diff --git a/README.md b/README.md
index 814f927b..6b7ce909 100644
--- a/README.md
+++ b/README.md
@@ -63,6 +63,7 @@ CPU backend:
       --asm=ASM                 ASM optimizations, possible values: auto, none, intel, ryzen, bulldozer
       --randomx-init=N          threads count to initialize RandomX dataset
       --randomx-no-numa         disable NUMA support for RandomX
+      --randomx-mode=MODE       RandomX mode: auto, fast, light
 
 API:
       --api-worker-id=ID        custom worker-id for API
diff --git a/doc/build/CMAKE_OPTIONS.md b/doc/build/CMAKE_OPTIONS.md
index 33a62bd2..654299de 100644
--- a/doc/build/CMAKE_OPTIONS.md
+++ b/doc/build/CMAKE_OPTIONS.md
@@ -1,5 +1,5 @@
 # CMake options
-This document contains list of useful cmake options.
+**Recent version of this document: https://xmrig.com/docs/miner/cmake-options**
 
 ## Algorithms
 
diff --git a/src/App.cpp b/src/App.cpp
index 04b05451..1908482c 100644
--- a/src/App.cpp
+++ b/src/App.cpp
@@ -33,6 +33,7 @@
 #include "base/io/Console.h"
 #include "base/io/log/Log.h"
 #include "base/kernel/Signals.h"
+#include "base/kernel/Platform.h"
 #include "core/config/Config.h"
 #include "core/Controller.h"
 #include "core/Miner.h"
@@ -89,6 +90,8 @@ int xmrig::App::exec()
 
     m_controller->start();
 
+    Platform::setThreadPriority(5);
+
     rc = uv_run(uv_default_loop(), UV_RUN_DEFAULT);
     uv_loop_close(uv_default_loop());
 
diff --git a/src/Summary.cpp b/src/Summary.cpp
index 2055e972..757297c5 100644
--- a/src/Summary.cpp
+++ b/src/Summary.cpp
@@ -102,6 +102,23 @@ static void print_cpu(Config *)
 }
 
 
+static void print_memory()
+{
+    constexpr size_t oneGiB = 1024U * 1024U * 1024U;
+    const auto freeMem      = static_cast<double>(uv_get_free_memory());
+    const auto totalMem     = static_cast<double>(uv_get_total_memory());
+
+    const double percent = freeMem > 0 ? ((totalMem - freeMem) / totalMem * 100.0) : 100.0;
+
+    Log::print(GREEN_BOLD(" * ") WHITE_BOLD("%-13s") CYAN_BOLD("%.1f/%.1f GB") BLACK_BOLD(" (%.0f%%)"),
+               "MEMORY",
+               (totalMem - freeMem) / oneGiB,
+               totalMem / oneGiB,
+               percent
+               );
+}
+
+
 static void print_threads(Config *config)
 {
     Log::print(GREEN_BOLD(" * ") WHITE_BOLD("%-13s") WHITE_BOLD("%s%d%%"),
@@ -144,6 +161,7 @@ void xmrig::Summary::print(Controller *controller)
     controller->config()->printVersions();
     print_memory(controller->config());
     print_cpu(controller->config());
+    print_memory();
     print_threads(controller->config());
     controller->config()->pools().print();
 
diff --git a/src/backend/common/Workers.cpp b/src/backend/common/Workers.cpp
index 8e195b66..1efa6cbe 100644
--- a/src/backend/common/Workers.cpp
+++ b/src/backend/common/Workers.cpp
@@ -109,6 +109,11 @@ void xmrig::Workers<T>::start(const std::vector<T> &data)
 
     for (Thread<T> *worker : m_workers) {
         worker->start(Workers<T>::onReady);
+
+        // This sleep is important for optimal caching!
+        // Threads must allocate scratchpads in order so that adjacent cores will use adjacent scratchpads
+        // Sub-optimal caching can result in up to 0.5% hashrate penalty
+        std::this_thread::sleep_for(std::chrono::milliseconds(20));
     }
 }
 
@@ -163,7 +168,7 @@ void xmrig::Workers<T>::onReady(void *arg)
     assert(worker != nullptr);
 
     if (!worker || !worker->selfTest()) {
-        LOG_ERR("%s " RED("thread ") RED_BOLD("#%zu") RED(" self-test failed"), T::tag(), worker->id());
+        LOG_ERR("%s " RED("thread ") RED_BOLD("#%zu") RED(" self-test failed"), T::tag(), worker ? worker->id() : 0);
 
         handle->backend()->start(worker, false);
         delete worker;
diff --git a/src/backend/common/interfaces/IRxStorage.h b/src/backend/common/interfaces/IRxStorage.h
index 49273ee3..9e0407b0 100644
--- a/src/backend/common/interfaces/IRxStorage.h
+++ b/src/backend/common/interfaces/IRxStorage.h
@@ -24,6 +24,9 @@
 #define XMRIG_IRXSTORAGE_H
 
 
+#include "crypto/rx/RxConfig.h"
+
+
 #include <cstdint>
 #include <utility>
 
@@ -41,9 +44,9 @@ class IRxStorage
 public:
     virtual ~IRxStorage() = default;
 
-    virtual RxDataset *dataset(const Job &job, uint32_t nodeId) const       = 0;
-    virtual std::pair<uint32_t, uint32_t> hugePages() const                 = 0;
-    virtual void init(const RxSeed &seed, uint32_t threads, bool hugePages) = 0;
+    virtual RxDataset *dataset(const Job &job, uint32_t nodeId) const                               = 0;
+    virtual std::pair<uint32_t, uint32_t> hugePages() const                                         = 0;
+    virtual void init(const RxSeed &seed, uint32_t threads, bool hugePages, RxConfig::Mode mode)    = 0;
 };
 
 
diff --git a/src/backend/cpu/CpuWorker.cpp b/src/backend/cpu/CpuWorker.cpp
index f64882ba..7584b686 100644
--- a/src/backend/cpu/CpuWorker.cpp
+++ b/src/backend/cpu/CpuWorker.cpp
@@ -185,8 +185,20 @@ void xmrig::CpuWorker<N>::start()
             consumeJob();
         }
 
+        uint64_t storeStatsMask = 7;
+
+#       ifdef XMRIG_ALGO_RANDOMX
+        bool first = true;
+        uint64_t tempHash[8] = {};
+
+        // RandomX is faster, we don't need to store stats so often
+        if (m_job.currentJob().algorithm().family() == Algorithm::RANDOM_X) {
+            storeStatsMask = 63;
+        }
+#       endif
+
         while (!Nonce::isOutdated(Nonce::CPU, m_job.sequence())) {
-            if ((m_count & 0x7) == 0) {
+            if ((m_count & storeStatsMask) == 0) {
                 storeStats();
             }
 
@@ -196,26 +208,34 @@ void xmrig::CpuWorker<N>::start()
                 break;
             }
 
+            uint32_t current_job_nonces[N];
+            for (size_t i = 0; i < N; ++i) {
+                current_job_nonces[i] = *m_job.nonce(i);
+            }
+
 #           ifdef XMRIG_ALGO_RANDOMX
             if (job.algorithm().family() == Algorithm::RANDOM_X) {
-                randomx_calculate_hash(m_vm->get(), m_job.blob(), job.size(), m_hash);
+                if (first) {
+                    first = false;
+                    randomx_calculate_hash_first(m_vm->get(), tempHash, m_job.blob(), job.size());
+                }
+                m_job.nextRound(kReserveCount, 1);
+                randomx_calculate_hash_next(m_vm->get(), tempHash, m_job.blob(), job.size(), m_hash);
             }
             else
 #           endif
             {
                 fn(job.algorithm())(m_job.blob(), job.size(), m_hash, m_ctx, job.height());
+                m_job.nextRound(kReserveCount, 1);
             }
 
             for (size_t i = 0; i < N; ++i) {
                 if (*reinterpret_cast<uint64_t*>(m_hash + (i * 32) + 24) < job.target()) {
-                    JobResults::submit(job, *m_job.nonce(i), m_hash + (i * 32));
+                    JobResults::submit(job, current_job_nonces[i], m_hash + (i * 32));
                 }
             }
 
-            m_job.nextRound(kReserveCount, 1);
             m_count += N;
-
-            std::this_thread::yield();
         }
 
         consumeJob();
diff --git a/src/backend/cuda/CudaBackend.cpp b/src/backend/cuda/CudaBackend.cpp
index 8ba7ed5c..5bb1cbdd 100644
--- a/src/backend/cuda/CudaBackend.cpp
+++ b/src/backend/cuda/CudaBackend.cpp
@@ -64,7 +64,7 @@ namespace xmrig {
 extern template class Threads<CudaThreads>;
 
 
-constexpr const size_t oneMiB   = 1024u * 1024u;
+constexpr const size_t oneMiB   = 1024U * 1024U;
 static const char *kLabel       = "CUDA";
 static const char *tag          = GREEN_BG_BOLD(WHITE_BOLD_S " nv  ");
 static const String kType       = "cuda";
@@ -249,7 +249,7 @@ public:
 
             std::string fans;
             if (!health.fanSpeed.empty()) {
-                for (uint32_t i = 0; i < health.fanSpeed.size(); ++i) {
+                for (size_t i = 0; i < health.fanSpeed.size(); ++i) {
                     fans += " fan" + std::to_string(i) + ":" CYAN_BOLD_S + std::to_string(health.fanSpeed[i]) + "%" CLEAR;
                 }
             }
diff --git a/src/backend/opencl/OclBackend.cpp b/src/backend/opencl/OclBackend.cpp
index bbf60443..af97b71e 100644
--- a/src/backend/opencl/OclBackend.cpp
+++ b/src/backend/opencl/OclBackend.cpp
@@ -58,7 +58,7 @@ namespace xmrig {
 extern template class Threads<OclThreads>;
 
 
-constexpr const size_t oneMiB   = 1024u * 1024u;
+constexpr const size_t oneMiB   = 1024U * 1024U;
 static const char *tag          = MAGENTA_BG_BOLD(WHITE_BOLD_S " ocl ");
 static const String kType       = "opencl";
 static std::mutex mutex;
diff --git a/src/base/api/Api.cpp b/src/base/api/Api.cpp
index 22f00b77..c66241c9 100644
--- a/src/base/api/Api.cpp
+++ b/src/base/api/Api.cpp
@@ -49,12 +49,48 @@
 #endif
 
 
+#include <thread>
+
+
+namespace xmrig {
+
+
+static rapidjson::Value getResources(rapidjson::Document &doc)
+{
+    using namespace rapidjson;
+    auto &allocator = doc.GetAllocator();
+
+    size_t rss = 0;
+    uv_resident_set_memory(&rss);
+
+    Value out(kObjectType);
+    Value memory(kObjectType);
+    Value load_average(kArrayType);
+
+    memory.AddMember("free",                uv_get_free_memory(), allocator);
+    memory.AddMember("total",               uv_get_total_memory(), allocator);
+    memory.AddMember("resident_set_memory", static_cast<uint64_t>(rss), allocator);
+
+    double loadavg[3] = { 0.0 };
+    uv_loadavg(loadavg);
+    load_average.PushBack(loadavg[0], allocator);
+    load_average.PushBack(loadavg[1], allocator);
+    load_average.PushBack(loadavg[2], allocator);
+
+    out.AddMember("memory",               memory, allocator);
+    out.AddMember("load_average",         load_average, allocator);
+    out.AddMember("hardware_concurrency", std::thread::hardware_concurrency(), allocator);
+
+    return out;
+}
+
+
+} // namespace xmrig
+
+
 xmrig::Api::Api(Base *base) :
     m_base(base),
-    m_id(),
-    m_workerId(),
-    m_timestamp(Chrono::currentMSecsSinceEpoch()),
-    m_httpd(nullptr)
+    m_timestamp(Chrono::currentMSecsSinceEpoch())
 {
     base->addListener(this);
 
@@ -117,10 +153,13 @@ void xmrig::Api::exec(IApiRequest &request)
         auto &allocator = request.doc().GetAllocator();
 
         request.accept();
-        request.reply().AddMember("id",         StringRef(m_id),       allocator);
-        request.reply().AddMember("worker_id",  StringRef(m_workerId), allocator);
-        request.reply().AddMember("uptime",     (Chrono::currentMSecsSinceEpoch() - m_timestamp) / 1000, allocator);
-        request.reply().AddMember("restricted", request.isRestricted(), allocator);
+
+        auto &reply = request.reply();
+        reply.AddMember("id",         StringRef(m_id),       allocator);
+        reply.AddMember("worker_id",  StringRef(m_workerId), allocator);
+        reply.AddMember("uptime",     (Chrono::currentMSecsSinceEpoch() - m_timestamp) / 1000, allocator);
+        reply.AddMember("restricted", request.isRestricted(), allocator);
+        reply.AddMember("resources",  getResources(request.doc()), allocator);
 
         Value features(kArrayType);
 #       ifdef XMRIG_FEATURE_API
@@ -144,7 +183,7 @@ void xmrig::Api::exec(IApiRequest &request)
 #       ifdef XMRIG_FEATURE_OPENCL
         features.PushBack("opencl", allocator);
 #       endif
-        request.reply().AddMember("features", features, allocator);
+        reply.AddMember("features", features, allocator);
     }
 
     for (IApiListener *listener : m_listeners) {
@@ -179,13 +218,13 @@ void xmrig::Api::genId(const String &id)
         if (!interfaces[i].is_internal && interfaces[i].address.address4.sin_family == AF_INET) {
             uint8_t hash[200];
             const size_t addrSize = sizeof(interfaces[i].phys_addr);
-            const size_t inSize   = strlen(APP_KIND) + addrSize + sizeof(uint16_t);
-            const uint16_t port   = static_cast<uint16_t>(m_base->config()->http().port());
+            const size_t inSize   = (sizeof(APP_KIND) - 1) + addrSize + sizeof(uint16_t);
+            const auto port       = static_cast<uint16_t>(m_base->config()->http().port());
 
-            uint8_t *input = new uint8_t[inSize]();
+            auto*input = new uint8_t[inSize]();
             memcpy(input, &port, sizeof(uint16_t));
             memcpy(input + sizeof(uint16_t), interfaces[i].phys_addr, addrSize);
-            memcpy(input + sizeof(uint16_t) + addrSize, APP_KIND, strlen(APP_KIND));
+            memcpy(input + sizeof(uint16_t) + addrSize, APP_KIND, (sizeof(APP_KIND) - 1));
 
             keccak(input, inSize, hash);
             Buffer::toHex(hash, 8, m_id);
diff --git a/src/base/api/Api.h b/src/base/api/Api.h
index 0c1a728d..130ce78e 100644
--- a/src/base/api/Api.h
+++ b/src/base/api/Api.h
@@ -70,10 +70,10 @@ private:
     void genWorkerId(const String &id);
 
     Base *m_base;
-    char m_id[32];
-    char m_workerId[128];
+    char m_id[32]{};
+    char m_workerId[128]{};
     const uint64_t m_timestamp;
-    Httpd *m_httpd;
+    Httpd *m_httpd = nullptr;
     std::vector<IApiListener *> m_listeners;
 };
 
diff --git a/src/base/api/Httpd.cpp b/src/base/api/Httpd.cpp
index e61e66f1..27e7dc44 100644
--- a/src/base/api/Httpd.cpp
+++ b/src/base/api/Httpd.cpp
@@ -23,9 +23,9 @@
  */
 
 
+#include "base/api/Httpd.h"
 #include "3rdparty/http-parser/http_parser.h"
 #include "base/api/Api.h"
-#include "base/api/Httpd.h"
 #include "base/io/log/Log.h"
 #include "base/net/http/HttpApiResponse.h"
 #include "base/net/http/HttpData.h"
@@ -58,9 +58,7 @@ xmrig::Httpd::Httpd(Base *base) :
 }
 
 
-xmrig::Httpd::~Httpd()
-{
-}
+xmrig::Httpd::~Httpd() = default;
 
 
 bool xmrig::Httpd::start()
diff --git a/src/base/api/Httpd.h b/src/base/api/Httpd.h
index 220bb7f5..ebfbcbf7 100644
--- a/src/base/api/Httpd.h
+++ b/src/base/api/Httpd.h
@@ -26,11 +26,12 @@
 #define XMRIG_HTTPD_H
 
 
-#include <stdint.h>
+#include <cstdint>
 
 
 #include "base/kernel/interfaces/IBaseListener.h"
 #include "base/kernel/interfaces/IHttpListener.h"
+#include "base/tools/Object.h"
 
 
 namespace xmrig {
@@ -44,6 +45,8 @@ class TcpServer;
 class Httpd : public IBaseListener, public IHttpListener
 {
 public:
+    XMRIG_DISABLE_COPY_MOVE_DEFAULT(Httpd)
+
     Httpd(Base *base);
     ~Httpd() override;
 
diff --git a/src/base/api/requests/ApiRequest.cpp b/src/base/api/requests/ApiRequest.cpp
index da73adee..4f0bd044 100644
--- a/src/base/api/requests/ApiRequest.cpp
+++ b/src/base/api/requests/ApiRequest.cpp
@@ -33,6 +33,4 @@ xmrig::ApiRequest::ApiRequest(Source source, bool restricted) :
 }
 
 
-xmrig::ApiRequest::~ApiRequest()
-{
-}
+xmrig::ApiRequest::~ApiRequest() = default;
diff --git a/src/base/api/requests/ApiRequest.h b/src/base/api/requests/ApiRequest.h
index ad4b0c35..c87e822c 100644
--- a/src/base/api/requests/ApiRequest.h
+++ b/src/base/api/requests/ApiRequest.h
@@ -29,6 +29,7 @@
 
 #include "base/api/interfaces/IApiRequest.h"
 #include "base/tools/String.h"
+#include "base/tools/Object.h"
 
 
 namespace xmrig {
@@ -37,6 +38,8 @@ namespace xmrig {
 class ApiRequest : public IApiRequest
 {
 public:
+    XMRIG_DISABLE_COPY_MOVE_DEFAULT(ApiRequest)
+
     ApiRequest(Source source, bool restricted);
     ~ApiRequest() override;
 
@@ -63,8 +66,8 @@ protected:
     String m_rpcMethod;
 
 private:
-    bool m_restricted;
-    Source m_source;
+    const bool m_restricted;
+    const Source m_source;
 };
 
 
diff --git a/src/base/kernel/config/BaseConfig.cpp b/src/base/kernel/config/BaseConfig.cpp
index 3da06b57..76056449 100644
--- a/src/base/kernel/config/BaseConfig.cpp
+++ b/src/base/kernel/config/BaseConfig.cpp
@@ -63,11 +63,6 @@
 #include "version.h"
 
 
-xmrig::BaseConfig::BaseConfig()
-{
-}
-
-
 void xmrig::BaseConfig::printVersions()
 {
     char buf[256] = { 0 };
diff --git a/src/base/kernel/config/BaseConfig.h b/src/base/kernel/config/BaseConfig.h
index 2e10c3ff..c700d109 100644
--- a/src/base/kernel/config/BaseConfig.h
+++ b/src/base/kernel/config/BaseConfig.h
@@ -31,9 +31,6 @@
 #include "base/net/stratum/Pools.h"
 
 
-struct option;
-
-
 namespace xmrig {
 
 
@@ -43,7 +40,7 @@ class IJsonReader;
 class BaseConfig : public IConfig
 {
 public:
-    BaseConfig();
+    BaseConfig() = default;
 
     inline bool isAutoSave() const                 { return m_autoSave; }
     inline bool isBackground() const               { return m_background; }
diff --git a/src/base/kernel/interfaces/IConfig.h b/src/base/kernel/interfaces/IConfig.h
index 7167e0bb..5044545a 100644
--- a/src/base/kernel/interfaces/IConfig.h
+++ b/src/base/kernel/interfaces/IConfig.h
@@ -89,6 +89,7 @@ public:
         AssemblyKey          = 1015,
         RandomXInitKey       = 1022,
         RandomXNumaKey       = 1023,
+        RandomXModeKey       = 1029,
         CPUMaxThreadsKey     = 1026,
         MemoryPoolKey        = 1027,
 
@@ -125,6 +126,7 @@ public:
         TlsProtocolsKey      = 1114,
         AlgoExtKey           = 1115,
         ProxyPasswordKey     = 1116,
+        LoginFileKey         = 'L',
 
         // xmrig nvidia
         CudaMaxThreadsKey    = 1200,
diff --git a/src/base/kernel/interfaces/IStrategy.h b/src/base/kernel/interfaces/IStrategy.h
index f2e58408..a0a24987 100644
--- a/src/base/kernel/interfaces/IStrategy.h
+++ b/src/base/kernel/interfaces/IStrategy.h
@@ -26,7 +26,7 @@
 #define XMRIG_ISTRATEGY_H
 
 
-#include <stdint.h>
+#include <cstdint>
 
 
 namespace xmrig {
diff --git a/src/base/net/http/Http.cpp b/src/base/net/http/Http.cpp
index 3c275824..c58080e6 100644
--- a/src/base/net/http/Http.cpp
+++ b/src/base/net/http/Http.cpp
@@ -23,9 +23,9 @@
  */
 
 
+#include "base/net/http/Http.h"
 #include "3rdparty/rapidjson/document.h"
 #include "base/io/json/Json.h"
-#include "base/net/http/Http.h"
 
 
 namespace xmrig {
@@ -41,10 +41,7 @@ static const char *kToken      = "access-token";
 
 
 xmrig::Http::Http() :
-    m_enabled(false),
-    m_restricted(true),
-    m_host(kLocalhost),
-    m_port(0)
+    m_host(kLocalhost)
 {
 }
 
diff --git a/src/base/net/http/Http.h b/src/base/net/http/Http.h
index 21eb581a..0e8e4fbb 100644
--- a/src/base/net/http/Http.h
+++ b/src/base/net/http/Http.h
@@ -38,7 +38,7 @@ class Http
 public:
     Http();
 
-    inline bool isAuthRequired() const         { return m_restricted == false || !m_token.isNull(); }
+    inline bool isAuthRequired() const         { return !m_restricted || !m_token.isNull(); }
     inline bool isEnabled() const              { return m_enabled; }
     inline bool isRestricted() const           { return m_restricted; }
     inline const String &host() const          { return m_host; }
@@ -58,11 +58,11 @@ public:
     void setPort(int port);
 
 private:
-    bool m_enabled;
-    bool m_restricted;
+    bool m_enabled      = false;
+    bool m_restricted   = true;
     String m_host;
     String m_token;
-    uint16_t m_port;
+    uint16_t m_port     = 0;
 };
 
 
diff --git a/src/base/net/http/HttpClient.cpp b/src/base/net/http/HttpClient.cpp
index 2699e663..9db1e809 100644
--- a/src/base/net/http/HttpClient.cpp
+++ b/src/base/net/http/HttpClient.cpp
@@ -66,7 +66,7 @@ public:
     inline static void onWrite(uv_write_t *req, int) { delete reinterpret_cast<ClientWriteBaton *>(req->data); }
 
 
-    uv_buf_t bufs[2];
+    uv_buf_t bufs[2]{};
 
 private:
     std::string m_body;
diff --git a/src/base/net/http/HttpContext.cpp b/src/base/net/http/HttpContext.cpp
index 1130d12c..af576257 100644
--- a/src/base/net/http/HttpContext.cpp
+++ b/src/base/net/http/HttpContext.cpp
@@ -24,27 +24,30 @@
  */
 
 
+#include "base/net/http/HttpContext.h"
+#include "3rdparty/http-parser/http_parser.h"
+#include "base/kernel/interfaces/IHttpListener.h"
+#include "base/tools/Chrono.h"
+
+
 #include <algorithm>
 #include <uv.h>
 
 
-#include "3rdparty/http-parser/http_parser.h"
-#include "base/kernel/interfaces/IHttpListener.h"
-#include "base/net/http/HttpContext.h"
-
-
 namespace xmrig {
 
+
 static http_parser_settings http_settings;
 static std::map<uint64_t, HttpContext *> storage;
 static uint64_t SEQUENCE = 0;
 
+
 } // namespace xmrig
 
 
 xmrig::HttpContext::HttpContext(int parser_type, IHttpListener *listener) :
     HttpData(SEQUENCE++),
-    m_wasHeaderValue(false),
+    m_timestamp(Chrono::steadyMSecs()),
     m_listener(listener)
 {
     storage[id()] = this;
@@ -96,6 +99,12 @@ std::string xmrig::HttpContext::ip() const
 }
 
 
+uint64_t xmrig::HttpContext::elapsed() const
+{
+    return Chrono::steadyMSecs() - m_timestamp;
+}
+
+
 void xmrig::HttpContext::close(int status)
 {
     if (status < 0 && m_listener) {
diff --git a/src/base/net/http/HttpContext.h b/src/base/net/http/HttpContext.h
index ba4418b6..9c304139 100644
--- a/src/base/net/http/HttpContext.h
+++ b/src/base/net/http/HttpContext.h
@@ -59,6 +59,7 @@ public:
 
     size_t parse(const char *data, size_t size);
     std::string ip() const;
+    uint64_t elapsed() const;
     void close(int status = 0);
 
     static HttpContext *get(uint64_t id);
@@ -74,7 +75,8 @@ private:
 
     void setHeader();
 
-    bool m_wasHeaderValue;
+    bool m_wasHeaderValue           = false;
+    const uint64_t m_timestamp;
     http_parser *m_parser;
     IHttpListener *m_listener;
     std::string m_lastHeaderField;
diff --git a/src/base/net/http/HttpData.h b/src/base/net/http/HttpData.h
index ceb19b8e..1705e881 100644
--- a/src/base/net/http/HttpData.h
+++ b/src/base/net/http/HttpData.h
@@ -38,12 +38,12 @@ namespace xmrig {
 class HttpData
 {
 public:
-    inline HttpData(uint64_t id) : method(0), status(0), m_id(id) {}
+    inline HttpData(uint64_t id) : m_id(id) {}
 
     inline uint64_t id() const { return m_id; }
 
-    int method;
-    int status;
+    int method  = 0;
+    int status  = 0;
     std::map<const std::string, const std::string> headers;
     std::string body;
     std::string url;
diff --git a/src/base/net/http/HttpResponse.cpp b/src/base/net/http/HttpResponse.cpp
index 7a4af838..ebd0bcf9 100644
--- a/src/base/net/http/HttpResponse.cpp
+++ b/src/base/net/http/HttpResponse.cpp
@@ -24,20 +24,23 @@
  */
 
 
-#include <sstream>
-#include <string.h>
-#include <uv.h>
-
-
+#include "base/net/http/HttpResponse.h"
 #include "3rdparty/http-parser/http_parser.h"
 #include "base/io/log/Log.h"
 #include "base/net/http/HttpContext.h"
-#include "base/net/http/HttpResponse.h"
 #include "base/tools/Baton.h"
+#include "base/tools/Object.h"
+
+
+#include <cinttypes>
+#include <cstring>
+#include <sstream>
+#include <uv.h>
 
 
 namespace xmrig {
 
+
 static const char *kCRLF      = "\r\n";
 static const char *kUserAgent = "user-agent";
 
@@ -45,6 +48,8 @@ static const char *kUserAgent = "user-agent";
 class WriteBaton : public Baton<uv_write_t>
 {
 public:
+    XMRIG_DISABLE_COPY_MOVE_DEFAULT(WriteBaton)
+
     inline WriteBaton(const std::stringstream &ss, const char *data, size_t size, HttpContext *ctx) :
         m_ctx(ctx),
         m_header(ss.str())
@@ -79,7 +84,7 @@ public:
     inline static void onWrite(uv_write_t *req, int) { delete reinterpret_cast<WriteBaton *>(req->data); }
 
 
-    uv_buf_t bufs[2];
+    uv_buf_t bufs[2]{};
 
 private:
     HttpContext *m_ctx;
@@ -98,7 +103,7 @@ xmrig::HttpResponse::HttpResponse(uint64_t id, int statusCode) :
 
 bool xmrig::HttpResponse::isAlive() const
 {
-    HttpContext *ctx = HttpContext::get(m_id);
+    auto ctx = HttpContext::get(m_id);
 
     return ctx && uv_is_writable(ctx->stream());
 }
@@ -129,8 +134,8 @@ void xmrig::HttpResponse::end(const char *data, size_t size)
 
     ss << kCRLF;
 
-    HttpContext *ctx  = HttpContext::get(m_id);
-    WriteBaton *baton = new WriteBaton(ss, data, size, ctx);
+    auto ctx   = HttpContext::get(m_id);
+    auto baton = new WriteBaton(ss, data, size, ctx);
 
 #   ifndef APP_DEBUG
     if (statusCode() >= 400)
@@ -138,13 +143,14 @@ void xmrig::HttpResponse::end(const char *data, size_t size)
     {
         const bool err = statusCode() >= 400;
 
-        Log::print(err ? Log::ERR : Log::INFO, CYAN("%s ") CLEAR MAGENTA_BOLD("%s") WHITE_BOLD(" %s ") CSI "1;%dm%d " CLEAR WHITE_BOLD("%zu ") BLACK_BOLD("\"%s\""),
+        Log::print(err ? Log::ERR : Log::INFO, CYAN("%s ") CLEAR MAGENTA_BOLD("%s") WHITE_BOLD(" %s ") CSI "1;%dm%d " CLEAR WHITE_BOLD("%zu ") CYAN_BOLD("%" PRIu64 "ms ") BLACK_BOLD("\"%s\""),
                    ctx->ip().c_str(),
                    http_method_str(static_cast<http_method>(ctx->method)),
                    ctx->url.c_str(),
                    err ? 31 : 32,
                    statusCode(),
                    baton->size(),
+                   ctx->elapsed(),
                    ctx->headers.count(kUserAgent) ? ctx->headers.at(kUserAgent).c_str() : nullptr
                    );
     }
diff --git a/src/base/net/http/HttpServer.cpp b/src/base/net/http/HttpServer.cpp
index 60db31f6..5aae1b6f 100644
--- a/src/base/net/http/HttpServer.cpp
+++ b/src/base/net/http/HttpServer.cpp
@@ -49,7 +49,7 @@ xmrig::HttpServer::~HttpServer()
 
 void xmrig::HttpServer::onConnection(uv_stream_t *stream, uint16_t)
 {
-    HttpContext *ctx = new HttpContext(HTTP_REQUEST, m_listener);
+    auto ctx = new HttpContext(HTTP_REQUEST, m_listener);
     uv_accept(stream, ctx->stream());
 
     uv_read_start(ctx->stream(),
@@ -65,11 +65,11 @@ void xmrig::HttpServer::onConnection(uv_stream_t *stream, uint16_t)
         },
         [](uv_stream_t *tcp, ssize_t nread, const uv_buf_t *buf)
         {
-            HttpContext *ctx = static_cast<HttpContext*>(tcp->data);
+            auto ctx = static_cast<HttpContext*>(tcp->data);
 
             if (nread >= 0) {
-                const size_t size   = static_cast<size_t>(nread);
-                const size_t parsed = ctx->parse(buf->base, size);
+                const auto size   = static_cast<size_t>(nread);
+                const auto parsed = ctx->parse(buf->base, size);
 
                 if (parsed < size) {
                     ctx->close();
diff --git a/src/base/net/http/HttpServer.h b/src/base/net/http/HttpServer.h
index 45f3843d..74e846b9 100644
--- a/src/base/net/http/HttpServer.h
+++ b/src/base/net/http/HttpServer.h
@@ -28,11 +28,12 @@
 #define XMRIG_HTTPSERVER_H
 
 
-typedef struct http_parser http_parser;
-typedef struct http_parser_settings http_parser_settings;
+using http_parser           = struct http_parser;
+using http_parser_settings  = struct http_parser_settings;
 
 
 #include "base/kernel/interfaces/ITcpServerListener.h"
+#include "base/tools/Object.h"
 
 
 namespace xmrig {
@@ -44,6 +45,8 @@ class IHttpListener;
 class HttpServer : public ITcpServerListener
 {
 public:
+    XMRIG_DISABLE_COPY_MOVE_DEFAULT(HttpServer)
+
     HttpServer(IHttpListener *listener);
     ~HttpServer() override;
 
diff --git a/src/base/net/stratum/SelfSelectClient.h b/src/base/net/stratum/SelfSelectClient.h
index c48c5078..44b07212 100644
--- a/src/base/net/stratum/SelfSelectClient.h
+++ b/src/base/net/stratum/SelfSelectClient.h
@@ -54,7 +54,7 @@ protected:
     inline const char *mode() const override                                        { return m_client->mode(); }
     inline const char *tlsFingerprint() const override                              { return m_client->tlsFingerprint(); }
     inline const char *tlsVersion() const override                                  { return m_client->tlsVersion(); }
-    inline const Job &job() const override                                          { return m_client->job(); }
+    inline const Job &job() const override                                          { return m_job; }
     inline const Pool &pool() const override                                        { return m_client->pool(); }
     inline const String &ip() const override                                        { return m_client->ip(); }
     inline int id() const override                                                  { return m_client->id(); }
diff --git a/src/base/tools/Arguments.cpp b/src/base/tools/Arguments.cpp
index 3713dd11..933173aa 100644
--- a/src/base/tools/Arguments.cpp
+++ b/src/base/tools/Arguments.cpp
@@ -60,17 +60,17 @@ void xmrig::Arguments::add(const char *arg)
 
     const size_t size = strlen(arg);
     if (size > 4 && arg[0] == '-' && arg[1] == '-') {
-        const char *p = strstr(arg, "=");
+        const char *p = strchr(arg, '=');
 
         if (p) {
-            const size_t keySize = static_cast<size_t>(p - arg);
+            const auto keySize = static_cast<size_t>(p - arg);
 
-            m_data.push_back(String(arg, keySize));
-            m_data.push_back(arg + keySize + 1);
+            m_data.emplace_back(arg, keySize);
+            m_data.emplace_back(arg + keySize + 1);
 
             return;
         }
     }
 
-    m_data.push_back(arg);
+    m_data.emplace_back(arg);
 }
diff --git a/src/base/tools/Buffer.cpp b/src/base/tools/Buffer.cpp
index c12e601d..982f9d36 100644
--- a/src/base/tools/Buffer.cpp
+++ b/src/base/tools/Buffer.cpp
@@ -31,10 +31,12 @@ static inline uint8_t hf_hex2bin(uint8_t c, bool &err)
     if (c >= '0' && c <= '9') {
         return c - '0';
     }
-    else if (c >= 'a' && c <= 'f') {
+
+    if (c >= 'a' && c <= 'f') {
         return c - 'a' + 0xA;
     }
-    else if (c >= 'A' && c <= 'F') {
+
+    if (c >= 'A' && c <= 'F') {
         return c - 'A' + 0xA;
     }
 
@@ -77,7 +79,9 @@ xmrig::Buffer::Buffer(const char *data, size_t size)
 xmrig::Buffer::Buffer(size_t size) :
     m_size(size)
 {
-    m_data = new char[size]();
+    if (size > 0) {
+        m_data = new char[size]();
+    }
 }
 
 
@@ -105,6 +109,10 @@ void xmrig::Buffer::from(const char *data, size_t size)
 
 xmrig::Buffer xmrig::Buffer::allocUnsafe(size_t size)
 {
+    if (size == 0) {
+        return {};
+    }
+
     Buffer buf;
     buf.m_size = size;
     buf.m_data = new char[size];
@@ -169,6 +177,13 @@ xmrig::String xmrig::Buffer::toHex() const
 
 void xmrig::Buffer::copy(const char *data, size_t size)
 {
+    if (size == 0) {
+        m_data = nullptr;
+        m_size = 0;
+
+        return;
+    }
+
     m_data = new char[size];
     m_size = size;
 
diff --git a/src/config.json b/src/config.json
index fbc60474..fbbf8225 100644
--- a/src/config.json
+++ b/src/config.json
@@ -16,6 +16,7 @@
     "colors": true,
     "randomx": {
         "init": -1,
+        "mode": "auto",
         "numa": true
     },
     "cpu": {
diff --git a/src/core/config/ConfigTransform.cpp b/src/core/config/ConfigTransform.cpp
index d0264258..667d9117 100644
--- a/src/core/config/ConfigTransform.cpp
+++ b/src/core/config/ConfigTransform.cpp
@@ -160,6 +160,9 @@ void xmrig::ConfigTransform::transform(rapidjson::Document &doc, int key, const
 
     case IConfig::RandomXNumaKey: /* --randomx-no-numa */
         return set(doc, kRandomX, "numa", false);
+
+    case IConfig::RandomXModeKey: /* --randomx-mode */
+        return set(doc, kRandomX, "mode", arg);
 #   endif
 
 #   ifdef XMRIG_FEATURE_OPENCL
diff --git a/src/core/config/Config_platform.h b/src/core/config/Config_platform.h
index 80fcf87f..608ac48e 100644
--- a/src/core/config/Config_platform.h
+++ b/src/core/config/Config_platform.h
@@ -96,6 +96,7 @@ static const option options[] = {
 #   ifdef XMRIG_ALGO_RANDOMX
     { "randomx-init",          1, nullptr, IConfig::RandomXInitKey        },
     { "randomx-no-numa",       0, nullptr, IConfig::RandomXNumaKey        },
+    { "randomx-mode",          1, nullptr, IConfig::RandomXModeKey        },
 #   endif
 #   ifdef XMRIG_FEATURE_OPENCL
     { "opencl",                0, nullptr, IConfig::OclKey                },
diff --git a/src/core/config/usage.h b/src/core/config/usage.h
index 48b34b81..dcae30f4 100644
--- a/src/core/config/usage.h
+++ b/src/core/config/usage.h
@@ -86,6 +86,7 @@ static inline const std::string &usage()
 #   ifdef XMRIG_ALGO_RANDOMX
     u += "      --randomx-init=N          threads count to initialize RandomX dataset\n";
     u += "      --randomx-no-numa         disable NUMA support for RandomX\n";
+    u += "      --randomx-mode=MODE       RandomX mode: auto, fast, light\n";
 #   endif
 
 #   ifdef XMRIG_FEATURE_HTTP
diff --git a/src/crypto/randomx/aes_hash.cpp b/src/crypto/randomx/aes_hash.cpp
index fe149dfe..4a400d0a 100644
--- a/src/crypto/randomx/aes_hash.cpp
+++ b/src/crypto/randomx/aes_hash.cpp
@@ -212,3 +212,84 @@ void fillAes4Rx4(void *state, size_t outputSize, void *buffer) {
 
 template void fillAes4Rx4<true>(void *state, size_t outputSize, void *buffer);
 template void fillAes4Rx4<false>(void *state, size_t outputSize, void *buffer);
+
+template<bool softAes>
+void hashAndFillAes1Rx4(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state) {
+	uint8_t* scratchpadPtr = (uint8_t*)scratchpad;
+	const uint8_t* scratchpadEnd = scratchpadPtr + scratchpadSize;
+
+	// initial state
+	rx_vec_i128 hash_state0 = rx_set_int_vec_i128(AES_HASH_1R_STATE0);
+	rx_vec_i128 hash_state1 = rx_set_int_vec_i128(AES_HASH_1R_STATE1);
+	rx_vec_i128 hash_state2 = rx_set_int_vec_i128(AES_HASH_1R_STATE2);
+	rx_vec_i128 hash_state3 = rx_set_int_vec_i128(AES_HASH_1R_STATE3);
+
+	const rx_vec_i128 key0 = rx_set_int_vec_i128(AES_GEN_1R_KEY0);
+	const rx_vec_i128 key1 = rx_set_int_vec_i128(AES_GEN_1R_KEY1);
+	const rx_vec_i128 key2 = rx_set_int_vec_i128(AES_GEN_1R_KEY2);
+	const rx_vec_i128 key3 = rx_set_int_vec_i128(AES_GEN_1R_KEY3);
+
+	rx_vec_i128 fill_state0 = rx_load_vec_i128((rx_vec_i128*)fill_state + 0);
+	rx_vec_i128 fill_state1 = rx_load_vec_i128((rx_vec_i128*)fill_state + 1);
+	rx_vec_i128 fill_state2 = rx_load_vec_i128((rx_vec_i128*)fill_state + 2);
+	rx_vec_i128 fill_state3 = rx_load_vec_i128((rx_vec_i128*)fill_state + 3);
+
+	constexpr int PREFETCH_DISTANCE = 4096;
+	const char* prefetchPtr = ((const char*)scratchpad) + PREFETCH_DISTANCE;
+	scratchpadEnd -= PREFETCH_DISTANCE;
+
+	for (int i = 0; i < 2; ++i) {
+		//process 64 bytes at a time in 4 lanes
+		while (scratchpadPtr < scratchpadEnd) {
+			hash_state0 = aesenc<softAes>(hash_state0, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 0));
+			hash_state1 = aesdec<softAes>(hash_state1, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 1));
+			hash_state2 = aesenc<softAes>(hash_state2, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 2));
+			hash_state3 = aesdec<softAes>(hash_state3, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 3));
+
+			fill_state0 = aesdec<softAes>(fill_state0, key0);
+			fill_state1 = aesenc<softAes>(fill_state1, key1);
+			fill_state2 = aesdec<softAes>(fill_state2, key2);
+			fill_state3 = aesenc<softAes>(fill_state3, key3);
+
+			rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 0, fill_state0);
+			rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 1, fill_state1);
+			rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 2, fill_state2);
+			rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 3, fill_state3);
+
+			rx_prefetch_t0(prefetchPtr);
+
+			scratchpadPtr += 64;
+			prefetchPtr += 64;
+		}
+		prefetchPtr = (const char*) scratchpad;
+		scratchpadEnd += PREFETCH_DISTANCE;
+	}
+
+	rx_store_vec_i128((rx_vec_i128*)fill_state + 0, fill_state0);
+	rx_store_vec_i128((rx_vec_i128*)fill_state + 1, fill_state1);
+	rx_store_vec_i128((rx_vec_i128*)fill_state + 2, fill_state2);
+	rx_store_vec_i128((rx_vec_i128*)fill_state + 3, fill_state3);
+
+	//two extra rounds to achieve full diffusion
+	rx_vec_i128 xkey0 = rx_set_int_vec_i128(AES_HASH_1R_XKEY0);
+	rx_vec_i128 xkey1 = rx_set_int_vec_i128(AES_HASH_1R_XKEY1);
+
+	hash_state0 = aesenc<softAes>(hash_state0, xkey0);
+	hash_state1 = aesdec<softAes>(hash_state1, xkey0);
+	hash_state2 = aesenc<softAes>(hash_state2, xkey0);
+	hash_state3 = aesdec<softAes>(hash_state3, xkey0);
+
+	hash_state0 = aesenc<softAes>(hash_state0, xkey1);
+	hash_state1 = aesdec<softAes>(hash_state1, xkey1);
+	hash_state2 = aesenc<softAes>(hash_state2, xkey1);
+	hash_state3 = aesdec<softAes>(hash_state3, xkey1);
+
+	//output hash
+	rx_store_vec_i128((rx_vec_i128*)hash + 0, hash_state0);
+	rx_store_vec_i128((rx_vec_i128*)hash + 1, hash_state1);
+	rx_store_vec_i128((rx_vec_i128*)hash + 2, hash_state2);
+	rx_store_vec_i128((rx_vec_i128*)hash + 3, hash_state3);
+}
+
+template void hashAndFillAes1Rx4<false>(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state);
+template void hashAndFillAes1Rx4<true>(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state);
diff --git a/src/crypto/randomx/aes_hash.hpp b/src/crypto/randomx/aes_hash.hpp
index b4d0e940..9f75f73a 100644
--- a/src/crypto/randomx/aes_hash.hpp
+++ b/src/crypto/randomx/aes_hash.hpp
@@ -38,3 +38,6 @@ void fillAes1Rx4(void *state, size_t outputSize, void *buffer);
 
 template<bool softAes>
 void fillAes4Rx4(void *state, size_t outputSize, void *buffer);
+
+template<bool softAes>
+void hashAndFillAes1Rx4(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state);
diff --git a/src/crypto/randomx/intrin_portable.h b/src/crypto/randomx/intrin_portable.h
index 346c433a..1dcd3ad3 100644
--- a/src/crypto/randomx/intrin_portable.h
+++ b/src/crypto/randomx/intrin_portable.h
@@ -102,6 +102,7 @@ typedef __m128d rx_vec_f128;
 #define rx_aligned_alloc(a, b) _mm_malloc(a,b)
 #define rx_aligned_free(a) _mm_free(a)
 #define rx_prefetch_nta(x) _mm_prefetch((const char *)(x), _MM_HINT_NTA)
+#define rx_prefetch_t0(x) _mm_prefetch((const char *)(x), _MM_HINT_T0)
 
 #define rx_load_vec_f128 _mm_load_pd
 #define rx_store_vec_f128 _mm_store_pd
@@ -201,6 +202,7 @@ typedef union{
 #define rx_aligned_alloc(a, b) malloc(a)
 #define rx_aligned_free(a) free(a)
 #define rx_prefetch_nta(x)
+#define rx_prefetch_t0(x)
 
 /* Splat 64-bit long long to 2 64-bit long longs */
 FORCE_INLINE __m128i vec_splat2sd (int64_t scalar)
@@ -399,6 +401,10 @@ inline void rx_prefetch_nta(void* ptr) {
 	asm volatile ("prfm pldl1strm, [%0]\n" : : "r" (ptr));
 }
 
+inline void rx_prefetch_t0(const void* ptr) {
+	asm volatile ("prfm pldl1strm, [%0]\n" : : "r" (ptr));
+}
+
 FORCE_INLINE rx_vec_f128 rx_load_vec_f128(const double* pd) {
 	return vld1q_f64((const float64_t*)pd);
 }
@@ -532,6 +538,7 @@ typedef union {
 #define rx_aligned_alloc(a, b) malloc(a)
 #define rx_aligned_free(a) free(a)
 #define rx_prefetch_nta(x)
+#define rx_prefetch_t0(x)
 
 FORCE_INLINE rx_vec_f128 rx_load_vec_f128(const double* pd) {
 	rx_vec_f128 x;
diff --git a/src/crypto/randomx/jit_compiler_x86.cpp b/src/crypto/randomx/jit_compiler_x86.cpp
index 2a342535..bfde7d00 100644
--- a/src/crypto/randomx/jit_compiler_x86.cpp
+++ b/src/crypto/randomx/jit_compiler_x86.cpp
@@ -29,6 +29,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include <stdexcept>
 #include <cstring>
 #include <climits>
+#include <atomic>
 #include "crypto/randomx/jit_compiler_x86.hpp"
 #include "crypto/randomx/jit_compiler_x86_static.hpp"
 #include "crypto/randomx/superscalar.hpp"
@@ -36,6 +37,12 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "crypto/randomx/reciprocal.h"
 #include "crypto/randomx/virtual_memory.hpp"
 
+#ifdef _MSC_VER
+#   include <intrin.h>
+#else
+#   include <cpuid.h>
+#endif
+
 namespace randomx {
 	/*
 
@@ -108,7 +115,7 @@ namespace randomx {
 	const int32_t codeSshPrefetchSize = codeShhEnd - codeShhPrefetch;
 	const int32_t codeSshInitSize = codeProgramEnd - codeShhInit;
 
-	const int32_t epilogueOffset = CodeSize - epilogueSize;
+	const int32_t epilogueOffset = (CodeSize - epilogueSize) & ~63;
 	constexpr int32_t superScalarHashOffset = 32768;
 
 	static const uint8_t REX_ADD_RR[] = { 0x4d, 0x03 };
@@ -183,6 +190,7 @@ namespace randomx {
 	static const uint8_t REX_ADD_I[] = { 0x49, 0x81 };
 	static const uint8_t REX_TEST[] = { 0x49, 0xF7 };
 	static const uint8_t JZ[] = { 0x0f, 0x84 };
+	static const uint8_t JZ_SHORT = 0x74;
 	static const uint8_t RET = 0xc3;
 	static const uint8_t LEA_32[] = { 0x41, 0x8d };
 	static const uint8_t MOVNTI[] = { 0x4c, 0x0f, 0xc3 };
@@ -197,20 +205,100 @@ namespace randomx {
 	static const uint8_t NOP7[] = { 0x0F, 0x1F, 0x80, 0x00, 0x00, 0x00, 0x00 };
 	static const uint8_t NOP8[] = { 0x0F, 0x1F, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00 };
 
-//	static const uint8_t* NOPX[] = { NOP1, NOP2, NOP3, NOP4, NOP5, NOP6, NOP7, NOP8 };
+	static const uint8_t* NOPX[] = { NOP1, NOP2, NOP3, NOP4, NOP5, NOP6, NOP7, NOP8 };
+
+	static const uint8_t JMP_ALIGN_PREFIX[14][16] = {
+		{},
+		{0x2E},
+		{0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x90, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x66, 0x90, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x66, 0x66, 0x90, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x0F, 0x1F, 0x40, 0x00, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x0F, 0x1F, 0x44, 0x00, 0x00, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+	};
+
+	bool JitCompilerX86::BranchesWithin32B = false;
 
 	size_t JitCompilerX86::getCodeSize() {
 		return codePos < prologueSize ? 0 : codePos - prologueSize;
 	}
 
+    static inline void cpuid(uint32_t level, int32_t output[4])
+    {
+        memset(output, 0, sizeof(int32_t) * 4);
+
+#   ifdef _MSC_VER
+        __cpuid(output, static_cast<int>(level));
+#   else
+        __cpuid_count(level, 0, output[0], output[1], output[2], output[3]);
+#   endif
+    }
+
+    // CPU-specific tweaks
+	void JitCompilerX86::applyTweaks() {
+		int32_t info[4];
+		cpuid(0, info);
+
+		int32_t manufacturer[4];
+		manufacturer[0] = info[1];
+		manufacturer[1] = info[3];
+		manufacturer[2] = info[2];
+		manufacturer[3] = 0;
+
+		if (strcmp((const char*)manufacturer, "GenuineIntel") == 0) {
+			struct
+			{
+				unsigned int stepping : 4;
+				unsigned int model : 4;
+				unsigned int family : 4;
+				unsigned int processor_type : 2;
+				unsigned int reserved1 : 2;
+				unsigned int ext_model : 4;
+				unsigned int ext_family : 8;
+				unsigned int reserved2 : 4;
+			} processor_info;
+
+			cpuid(1, info);
+			memcpy(&processor_info, info, sizeof(processor_info));
+
+			// Intel JCC erratum mitigation
+			if (processor_info.family == 6) {
+				const uint32_t model = processor_info.model | (processor_info.ext_model << 4);
+				const uint32_t stepping = processor_info.stepping;
+
+				// Affected CPU models and stepping numbers are taken from https://www.intel.com/content/dam/support/us/en/documents/processors/mitigations-jump-conditional-code-erratum.pdf
+				BranchesWithin32B =
+					((model == 0x4E) && (stepping == 0x3)) ||
+					((model == 0x55) && (stepping == 0x4)) ||
+					((model == 0x5E) && (stepping == 0x3)) ||
+					((model == 0x8E) && (stepping >= 0x9) && (stepping <= 0xC)) ||
+					((model == 0x9E) && (stepping >= 0x9) && (stepping <= 0xD)) ||
+					((model == 0xA6) && (stepping == 0x0)) ||
+					((model == 0xAE) && (stepping == 0xA));
+			}
+		}
+	}
+
+	static std::atomic<size_t> codeOffset;
+
 	JitCompilerX86::JitCompilerX86() {
-		code = (uint8_t*)allocExecutableMemory(CodeSize);
+		applyTweaks();
+		allocatedCode = (uint8_t*)allocExecutableMemory(CodeSize * 2);
+		// Shift code base address to improve caching - all threads will use different L2/L3 cache sets
+		code = allocatedCode + (codeOffset.fetch_add(59 * 64) % CodeSize);
 		memcpy(code, codePrologue, prologueSize);
 		memcpy(code + epilogueOffset, codeEpilogue, epilogueSize);
 	}
 
 	JitCompilerX86::~JitCompilerX86() {
-		freePagedMemory(code, CodeSize);
+		freePagedMemory(allocatedCode, CodeSize);
 	}
 
 	void JitCompilerX86::generateProgram(Program& prog, ProgramConfiguration& pcfg) {
@@ -307,6 +395,22 @@ namespace randomx {
 		emit(RandomX_CurrentConfig.codePrefetchScratchpadTweaked, prefetchScratchpadSize, code, codePos);
 		memcpy(code + codePos, codeLoopStore, loopStoreSize);
 		codePos += loopStoreSize;
+
+		if (BranchesWithin32B) {
+			const uint32_t branch_begin = static_cast<uint32_t>(codePos);
+			const uint32_t branch_end = static_cast<uint32_t>(branch_begin + 9);
+
+			// If the jump crosses or touches 32-byte boundary, align it
+			if ((branch_begin ^ branch_end) >= 32) {
+				uint32_t alignment_size = 32 - (branch_begin & 31);
+				if (alignment_size > 8) {
+					emit(NOPX[alignment_size - 9], alignment_size - 8, code, codePos);
+					alignment_size = 8;
+				}
+				emit(NOPX[alignment_size - 1], alignment_size, code, codePos);
+			}
+		}
+
 		emit(SUB_EBX, code, codePos);
 		emit(JNZ, code, codePos);
 		emit32(prologueSize - codePos - 4, code, codePos);
@@ -408,12 +512,13 @@ namespace randomx {
 		}
 	}
 
-	void JitCompilerX86::genAddressReg(const Instruction& instr, uint8_t* code, int& codePos, bool rax) {
-		emit(LEA_32, code, codePos);
-		emitByte(0x80 + instr.src + (rax ? 0 : 8), code, codePos);
-		if (instr.src == RegisterNeedsSib) {
-			emitByte(0x24, code, codePos);
-		}
+	template<bool rax>
+	FORCE_INLINE void JitCompilerX86::genAddressReg(const Instruction& instr, uint8_t* code, int& codePos) {
+		const uint32_t src = *((uint32_t*)&instr) & 0xFF0000;
+
+		*(uint32_t*)(code + codePos) = (rax ? 0x24808d41 : 0x24888d41) + src;
+		codePos += (src == (RegisterNeedsSib << 16)) ? 4 : 3;
+
 		emit32(instr.getImm32(), code, codePos);
 		if (rax)
 			emitByte(AND_EAX_I, code, codePos);
@@ -422,12 +527,14 @@ namespace randomx {
 		emit32(instr.getModMem() ? ScratchpadL1Mask : ScratchpadL2Mask, code, codePos);
 	}
 
-	void JitCompilerX86::genAddressRegDst(const Instruction& instr, uint8_t* code, int& codePos) {
-		emit(LEA_32, code, codePos);
-		emitByte(0x80 + instr.dst, code, codePos);
-		if (instr.dst == RegisterNeedsSib) {
-			emitByte(0x24, code, codePos);
-		}
+	template void JitCompilerX86::genAddressReg<false>(const Instruction& instr, uint8_t* code, int& codePos);
+	template void JitCompilerX86::genAddressReg<true>(const Instruction& instr, uint8_t* code, int& codePos);
+
+	FORCE_INLINE void JitCompilerX86::genAddressRegDst(const Instruction& instr, uint8_t* code, int& codePos) {
+		const uint32_t dst = static_cast<uint32_t>(instr.dst) << 16;
+		*(uint32_t*)(code + codePos) = 0x24808d41 + dst;
+		codePos += (dst == (RegisterNeedsSib << 16)) ? 4 : 3;
+
 		emit32(instr.getImm32(), code, codePos);
 		emitByte(AND_EAX_I, code, codePos);
 		if (instr.getModCond() < StoreL3Condition) {
@@ -438,7 +545,7 @@ namespace randomx {
 		}
 	}
 
-	void JitCompilerX86::genAddressImm(const Instruction& instr, uint8_t* code, int& codePos) {
+	FORCE_INLINE void JitCompilerX86::genAddressImm(const Instruction& instr, uint8_t* code, int& codePos) {
 		emit32(instr.getImm32() & ScratchpadL3Mask, code, codePos);
 	}
 
@@ -483,7 +590,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos);
+			genAddressReg<true>(instr, p, pos);
 			emit32(template_IADD_M[instr.dst], p, pos);
 		}
 		else {
@@ -523,7 +630,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos);
+			genAddressReg<true>(instr, p, pos);
 			emit(REX_SUB_RM, p, pos);
 			emitByte(0x04 + 8 * instr.dst, p, pos);
 			emitByte(0x06, p, pos);
@@ -561,7 +668,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos);
+			genAddressReg<true>(instr, p, pos);
 			emit(REX_IMUL_RM, p, pos);
 			emitByte(0x04 + 8 * instr.dst, p, pos);
 			emitByte(0x06, p, pos);
@@ -596,7 +703,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos, false);
+			genAddressReg<false>(instr, p, pos);
 			emit(REX_MOV_RR64, p, pos);
 			emitByte(0xc0 + instr.dst, p, pos);
 			emit(REX_MUL_MEM, p, pos);
@@ -635,7 +742,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos, false);
+			genAddressReg<false>(instr, p, pos);
 			emit(REX_MOV_RR64, p, pos);
 			emitByte(0xc0 + instr.dst, p, pos);
 			emit(REX_IMUL_MEM, p, pos);
@@ -704,7 +811,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos);
+			genAddressReg<true>(instr, p, pos);
 			emit(REX_XOR_RM, p, pos);
 			emitByte(0x04 + 8 * instr.dst, p, pos);
 			emitByte(0x06, p, pos);
@@ -801,7 +908,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		const uint32_t dst = instr.dst % RegisterCountFlt;
-		genAddressReg(instr, p, pos);
+		genAddressReg<true>(instr, p, pos);
 		emit(REX_CVTDQ2PD_XMM12, p, pos);
 		emit(REX_ADDPD, p, pos);
 		emitByte(0xc4 + 8 * dst, p, pos);
@@ -826,7 +933,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		const uint32_t dst = instr.dst % RegisterCountFlt;
-		genAddressReg(instr, p, pos);
+		genAddressReg<true>(instr, p, pos);
 		emit(REX_CVTDQ2PD_XMM12, p, pos);
 		emit(REX_SUBPD, p, pos);
 		emitByte(0xc4 + 8 * dst, p, pos);
@@ -862,7 +969,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		const uint32_t dst = instr.dst % RegisterCountFlt;
-		genAddressReg(instr, p, pos);
+		genAddressReg<true>(instr, p, pos);
 		emit(REX_CVTDQ2PD_XMM12, p, pos);
 		emit(REX_ANDPS_XMM12, p, pos);
 		emit(REX_DIVPD, p, pos);
@@ -902,19 +1009,39 @@ namespace randomx {
 		uint8_t* const p = code;
 		int pos = codePos;
 		
-		int reg = instr.dst;
+		const int reg = instr.dst;
+		int32_t jmp_offset = registerUsage[reg] - (pos + 16);
+
+		if (BranchesWithin32B) {
+			const uint32_t branch_begin = static_cast<uint32_t>(pos + 7);
+			const uint32_t branch_end = static_cast<uint32_t>(branch_begin + ((jmp_offset >= -128) ? 9 : 13));
+
+			// If the jump crosses or touches 32-byte boundary, align it
+			if ((branch_begin ^ branch_end) >= 32) {
+				const uint32_t alignment_size = 32 - (branch_begin & 31);
+				jmp_offset -= alignment_size;
+				emit(JMP_ALIGN_PREFIX[alignment_size], alignment_size, p, pos);
+			}
+		}
+
 		emit(REX_ADD_I, p, pos);
 		emitByte(0xc0 + reg, p, pos);
-		int shift = instr.getModCond() + RandomX_CurrentConfig.JumpOffset;
-		uint32_t imm = instr.getImm32() | (1UL << shift);
-		if (RandomX_CurrentConfig.JumpOffset > 0 || shift > 0)
-			imm &= ~(1UL << (shift - 1));
+		const int shift = instr.getModCond() + RandomX_CurrentConfig.JumpOffset;
+		const uint32_t imm = (instr.getImm32() | (1UL << shift)) & ~(1UL << (shift - 1));
 		emit32(imm, p, pos);
 		emit(REX_TEST, p, pos);
 		emitByte(0xc0 + reg, p, pos);
 		emit32(RandomX_CurrentConfig.ConditionMask_Calculated << shift, p, pos);
-		emit(JZ, p, pos);
-		emit32(registerUsage[reg] - (pos + 4), p, pos);
+
+		if (jmp_offset >= -128) {
+			emitByte(JZ_SHORT, p, pos);
+			emitByte(jmp_offset, p, pos);
+		}
+		else {
+			emit(JZ, p, pos);
+			emit32(jmp_offset - 4, p, pos);
+		}
+
 		//mark all registers as used
 		uint64_t* r = (uint64_t*) registerUsage;
 		uint64_t k = pos;
diff --git a/src/crypto/randomx/jit_compiler_x86.hpp b/src/crypto/randomx/jit_compiler_x86.hpp
index 30b16f58..f1864018 100644
--- a/src/crypto/randomx/jit_compiler_x86.hpp
+++ b/src/crypto/randomx/jit_compiler_x86.hpp
@@ -67,12 +67,17 @@ namespace randomx {
 
 		static InstructionGeneratorX86 engine[256];
 		int registerUsage[RegistersCount];
+		uint8_t* allocatedCode;
 		uint8_t* code;
 		int32_t codePos;
 
+		static bool BranchesWithin32B;
+
+		static void applyTweaks();
 		void generateProgramPrologue(Program&, ProgramConfiguration&);
 		void generateProgramEpilogue(Program&, ProgramConfiguration&);
-		static void genAddressReg(const Instruction&, uint8_t* code, int& codePos, bool rax = true);
+		template<bool rax>
+		static void genAddressReg(const Instruction&, uint8_t* code, int& codePos);
 		static void genAddressRegDst(const Instruction&, uint8_t* code, int& codePos);
 		static void genAddressImm(const Instruction&, uint8_t* code, int& codePos);
 		static void genSIB(int scale, int index, int base, uint8_t* code, int& codePos);
diff --git a/src/crypto/randomx/randomx.cpp b/src/crypto/randomx/randomx.cpp
index 4f56dc09..f9ce93f8 100644
--- a/src/crypto/randomx/randomx.cpp
+++ b/src/crypto/randomx/randomx.cpp
@@ -473,4 +473,22 @@ extern "C" {
 		machine->getFinalResult(output, RANDOMX_HASH_SIZE);
 	}
 
+	void randomx_calculate_hash_first(randomx_vm* machine, uint64_t (&tempHash)[8], const void* input, size_t inputSize) {
+		rx_blake2b(tempHash, sizeof(tempHash), input, inputSize, nullptr, 0);
+		machine->initScratchpad(tempHash);
+	}
+
+	void randomx_calculate_hash_next(randomx_vm* machine, uint64_t (&tempHash)[8], const void* nextInput, size_t nextInputSize, void* output) {
+		machine->resetRoundingMode();
+		for (uint32_t chain = 0; chain < RandomX_CurrentConfig.ProgramCount - 1; ++chain) {
+			machine->run(&tempHash);
+			rx_blake2b(tempHash, sizeof(tempHash), machine->getRegisterFile(), sizeof(randomx::RegisterFile), nullptr, 0);
+		}
+		machine->run(&tempHash);
+
+		// Finish current hash and fill the scratchpad for the next hash at the same time
+		rx_blake2b(tempHash, sizeof(tempHash), nextInput, nextInputSize, nullptr, 0);
+		machine->hashAndFill(output, RANDOMX_HASH_SIZE, tempHash);
+	}
+
 }
diff --git a/src/crypto/randomx/randomx.h b/src/crypto/randomx/randomx.h
index 6e074dba..ea3bb099 100644
--- a/src/crypto/randomx/randomx.h
+++ b/src/crypto/randomx/randomx.h
@@ -338,6 +338,9 @@ RANDOMX_EXPORT void randomx_destroy_vm(randomx_vm *machine);
 */
 RANDOMX_EXPORT void randomx_calculate_hash(randomx_vm *machine, const void *input, size_t inputSize, void *output);
 
+RANDOMX_EXPORT void randomx_calculate_hash_first(randomx_vm* machine, uint64_t (&tempHash)[8], const void* input, size_t inputSize);
+RANDOMX_EXPORT void randomx_calculate_hash_next(randomx_vm* machine, uint64_t (&tempHash)[8], const void* nextInput, size_t nextInputSize, void* output);
+
 #if defined(__cplusplus)
 }
 #endif
diff --git a/src/crypto/randomx/virtual_machine.cpp b/src/crypto/randomx/virtual_machine.cpp
index 2913c7e5..ecd187e2 100644
--- a/src/crypto/randomx/virtual_machine.cpp
+++ b/src/crypto/randomx/virtual_machine.cpp
@@ -114,6 +114,12 @@ namespace randomx {
         rx_blake2b(out, outSize, &reg, sizeof(RegisterFile), nullptr, 0);
 	}
 
+	template<bool softAes>
+	void VmBase<softAes>::hashAndFill(void* out, size_t outSize, uint64_t (&fill_state)[8]) {
+		hashAndFillAes1Rx4<softAes>(scratchpad, ScratchpadSize, &reg.a, fill_state);
+        rx_blake2b(out, outSize, &reg, sizeof(RegisterFile), nullptr, 0);
+	}
+
 	template<bool softAes>
 	void VmBase<softAes>::initScratchpad(void* seed) {
 		fillAes1Rx4<softAes>(seed, ScratchpadSize, scratchpad);
diff --git a/src/crypto/randomx/virtual_machine.hpp b/src/crypto/randomx/virtual_machine.hpp
index c85af009..d3718d04 100644
--- a/src/crypto/randomx/virtual_machine.hpp
+++ b/src/crypto/randomx/virtual_machine.hpp
@@ -39,6 +39,7 @@ public:
 	virtual ~randomx_vm() = 0;
 	virtual void setScratchpad(uint8_t *scratchpad) = 0;
 	virtual void getFinalResult(void* out, size_t outSize) = 0;
+	virtual void hashAndFill(void* out, size_t outSize, uint64_t (&fill_state)[8]) = 0;
 	virtual void setDataset(randomx_dataset* dataset) { }
 	virtual void setCache(randomx_cache* cache) { }
 	virtual void initScratchpad(void* seed) = 0;
@@ -82,6 +83,7 @@ namespace randomx {
 		void setScratchpad(uint8_t *scratchpad) override;
 		void initScratchpad(void* seed) override;
 		void getFinalResult(void* out, size_t outSize) override;
+		void hashAndFill(void* out, size_t outSize, uint64_t (&fill_state)[8]) override;
 
 	protected:
 		void generateProgram(void* seed);
diff --git a/src/crypto/rx/Rx.cpp b/src/crypto/rx/Rx.cpp
index 115c0f62..b08f844e 100644
--- a/src/crypto/rx/Rx.cpp
+++ b/src/crypto/rx/Rx.cpp
@@ -70,7 +70,7 @@ bool xmrig::Rx::init(const Job &job, const RxConfig &config, bool hugePages)
         return true;
     }
 
-    d_ptr->queue.enqueue(job, config.nodeset(), config.threads(), hugePages);
+    d_ptr->queue.enqueue(job, config.nodeset(), config.threads(), hugePages, config.mode());
 
     return false;
 }
diff --git a/src/crypto/rx/RxBasicStorage.cpp b/src/crypto/rx/RxBasicStorage.cpp
index dcabad5b..bc1ceb2d 100644
--- a/src/crypto/rx/RxBasicStorage.cpp
+++ b/src/crypto/rx/RxBasicStorage.cpp
@@ -69,11 +69,11 @@ public:
     }
 
 
-    inline void createDataset(bool hugePages)
+    inline void createDataset(bool hugePages, RxConfig::Mode mode)
     {
         const uint64_t ts = Chrono::steadyMSecs();
 
-        m_dataset = new RxDataset(hugePages, true);
+        m_dataset = new RxDataset(hugePages, true, mode);
         printAllocStatus(ts);
     }
 
@@ -150,19 +150,19 @@ xmrig::RxDataset *xmrig::RxBasicStorage::dataset(const Job &job, uint32_t) const
 std::pair<uint32_t, uint32_t> xmrig::RxBasicStorage::hugePages() const
 {
     if (!d_ptr->dataset()) {
-        return { 0u, 0u };
+        return { 0U, 0U };
     }
 
     return d_ptr->dataset()->hugePages();
 }
 
 
-void xmrig::RxBasicStorage::init(const RxSeed &seed, uint32_t threads, bool hugePages)
+void xmrig::RxBasicStorage::init(const RxSeed &seed, uint32_t threads, bool hugePages, RxConfig::Mode mode)
 {
     d_ptr->setSeed(seed);
 
     if (!d_ptr->dataset()) {
-        d_ptr->createDataset(hugePages);
+        d_ptr->createDataset(hugePages, mode);
     }
 
     d_ptr->initDataset(threads);
diff --git a/src/crypto/rx/RxBasicStorage.h b/src/crypto/rx/RxBasicStorage.h
index 63eba1d9..bd6575d2 100644
--- a/src/crypto/rx/RxBasicStorage.h
+++ b/src/crypto/rx/RxBasicStorage.h
@@ -50,7 +50,7 @@ public:
 protected:
     RxDataset *dataset(const Job &job, uint32_t nodeId) const override;
     std::pair<uint32_t, uint32_t> hugePages() const override;
-    void init(const RxSeed &seed, uint32_t threads, bool hugePages) override;
+    void init(const RxSeed &seed, uint32_t threads, bool hugePages, RxConfig::Mode mode) override;
 
 private:
     RxBasicStoragePrivate *d_ptr;
diff --git a/src/crypto/rx/RxConfig.cpp b/src/crypto/rx/RxConfig.cpp
index 07f45eac..ef823e60 100644
--- a/src/crypto/rx/RxConfig.cpp
+++ b/src/crypto/rx/RxConfig.cpp
@@ -25,9 +25,54 @@
 
 #include "crypto/rx/RxConfig.h"
 #include "backend/cpu/Cpu.h"
+#include "rapidjson/document.h"
+
+
+#include <array>
+#include <algorithm>
+
+
+#ifdef _MSC_VER
+#   define strcasecmp  _stricmp
+#endif
+
+
+namespace xmrig {
+
+
+static const std::array<const char *, RxConfig::ModeMax> modeNames = { "auto", "fast", "light" };
+
+
+} // namespace xmrig
+
+
+const char *xmrig::RxConfig::modeName() const
+{
+    return modeNames[m_mode];
+}
 
 
 uint32_t xmrig::RxConfig::threads() const
 {
     return m_threads < 1 ? static_cast<uint32_t>(Cpu::info()->threads()) : static_cast<uint32_t>(m_threads);
 }
+
+
+xmrig::RxConfig::Mode xmrig::RxConfig::readMode(const rapidjson::Value &value) const
+{
+    if (value.IsUint()) {
+        return static_cast<Mode>(std::min(value.GetUint(), ModeMax - 1));
+    }
+
+    if (value.IsString()) {
+        auto mode = value.GetString();
+
+        for (size_t i = 0; i < modeNames.size(); i++) {
+            if (strcasecmp(mode, modeNames[i]) == 0) {
+                return static_cast<Mode>(i);
+            }
+        }
+    }
+
+    return AutoMode;
+}
diff --git a/src/crypto/rx/RxConfig.h b/src/crypto/rx/RxConfig.h
index 52a832a2..13ebb9d1 100644
--- a/src/crypto/rx/RxConfig.h
+++ b/src/crypto/rx/RxConfig.h
@@ -38,6 +38,13 @@ namespace xmrig {
 class RxConfig
 {
 public:
+    enum Mode : uint32_t {
+        AutoMode,
+        FastMode,
+        LightMode,
+        ModeMax
+    };
+
     bool read(const rapidjson::Value &value);
     rapidjson::Value toJSON(rapidjson::Document &doc) const;
 
@@ -47,11 +54,17 @@ public:
     inline std::vector<uint32_t> nodeset() const { return std::vector<uint32_t>(); }
 #   endif
 
+    const char *modeName() const;
     uint32_t threads() const;
 
+    inline Mode mode() const    { return m_mode; }
+
 private:
+    Mode readMode(const rapidjson::Value &value) const;
+
     bool m_numa     = true;
     int m_threads   = -1;
+    Mode m_mode     = AutoMode;
 
 #   ifdef XMRIG_FEATURE_HWLOC
     std::vector<uint32_t> m_nodeset;
diff --git a/src/crypto/rx/RxConfig_basic.cpp b/src/crypto/rx/RxConfig_basic.cpp
index 26ef7a90..1311b220 100644
--- a/src/crypto/rx/RxConfig_basic.cpp
+++ b/src/crypto/rx/RxConfig_basic.cpp
@@ -31,6 +31,7 @@
 namespace xmrig {
 
 static const char *kInit = "init";
+static const char *kMode = "mode";
 
 }
 
@@ -42,6 +43,7 @@ rapidjson::Value xmrig::RxConfig::toJSON(rapidjson::Document &doc) const
 
     Value obj(kObjectType);
     obj.AddMember(StringRef(kInit), m_threads, allocator);
+    obj.AddMember(StringRef(kMode), StringRef(modeName()), allocator);
 
     return obj;
 }
@@ -51,6 +53,7 @@ bool xmrig::RxConfig::read(const rapidjson::Value &value)
 {
     if (value.IsObject()) {
         m_threads = Json::getInt(value, kInit, m_threads);
+        m_mode    = readMode(Json::getValue(value, kMode));
 
         return true;
     }
diff --git a/src/crypto/rx/RxConfig_hwloc.cpp b/src/crypto/rx/RxConfig_hwloc.cpp
index 66f086f2..4fc57fb6 100644
--- a/src/crypto/rx/RxConfig_hwloc.cpp
+++ b/src/crypto/rx/RxConfig_hwloc.cpp
@@ -33,6 +33,7 @@
 namespace xmrig {
 
 static const char *kInit = "init";
+static const char *kMode = "mode";
 static const char *kNUMA = "numa";
 
 }
@@ -46,6 +47,7 @@ rapidjson::Value xmrig::RxConfig::toJSON(rapidjson::Document &doc) const
     Value obj(kObjectType);
 
     obj.AddMember(StringRef(kInit), m_threads, allocator);
+    obj.AddMember(StringRef(kMode), StringRef(modeName()), allocator);
 
     if (!m_nodeset.empty()) {
         Value numa(kArrayType);
@@ -68,6 +70,13 @@ bool xmrig::RxConfig::read(const rapidjson::Value &value)
 {
     if (value.IsObject()) {
         m_threads = Json::getInt(value, kInit, m_threads);
+        m_mode    = readMode(Json::getValue(value, kMode));
+
+        if (m_mode == LightMode) {
+            m_numa = false;
+
+            return true;
+        }
 
         const auto &numa = Json::getValue(value, kNUMA);
         if (numa.IsArray()) {
diff --git a/src/crypto/rx/RxDataset.cpp b/src/crypto/rx/RxDataset.cpp
index 62887a01..ce41a58e 100644
--- a/src/crypto/rx/RxDataset.cpp
+++ b/src/crypto/rx/RxDataset.cpp
@@ -26,6 +26,8 @@
 
 
 #include "crypto/rx/RxDataset.h"
+#include "backend/common/Tags.h"
+#include "base/io/log/Log.h"
 #include "crypto/common/VirtualMemory.h"
 #include "crypto/randomx/randomx.h"
 #include "crypto/rx/RxAlgo.h"
@@ -33,12 +35,14 @@
 
 
 #include <thread>
+#include <uv.h>
 
 
 static_assert(RANDOMX_FLAG_LARGE_PAGES == 1, "RANDOMX_FLAG_LARGE_PAGES flag mismatch");
 
 
-xmrig::RxDataset::RxDataset(bool hugePages, bool cache)
+xmrig::RxDataset::RxDataset(bool hugePages, bool cache, RxConfig::Mode mode) :
+    m_mode(mode)
 {
     allocate(hugePages);
 
@@ -118,7 +122,7 @@ size_t xmrig::RxDataset::size(bool cache) const
 
 std::pair<uint32_t, uint32_t> xmrig::RxDataset::hugePages(bool cache) const
 {
-    constexpr size_t twoMiB     = 2u * 1024u * 1024u;
+    constexpr size_t twoMiB     = 2U * 1024U * 1024U;
     constexpr size_t cacheSize  = VirtualMemory::align(RxCache::maxSize(), twoMiB) / twoMiB;
     size_t total                = VirtualMemory::align(maxSize(), twoMiB) / twoMiB;
 
@@ -157,6 +161,18 @@ void xmrig::RxDataset::setRaw(const void *raw)
 
 void xmrig::RxDataset::allocate(bool hugePages)
 {
+    if (m_mode == RxConfig::LightMode) {
+        LOG_ERR(CLEAR "%s" RED_BOLD_S "fast RandomX mode disabled by config", rx_tag());
+
+        return;
+    }
+
+    if (m_mode == RxConfig::AutoMode && uv_get_total_memory() < (maxSize() + RxCache::maxSize())) {
+        LOG_ERR(CLEAR "%s" RED_BOLD_S "not enough memory for RandomX dataset", rx_tag());
+
+        return;
+    }
+
     if (hugePages) {
         m_flags   = RANDOMX_FLAG_LARGE_PAGES;
         m_dataset = randomx_alloc_dataset(static_cast<randomx_flags>(m_flags));
diff --git a/src/crypto/rx/RxDataset.h b/src/crypto/rx/RxDataset.h
index 9b4f4120..30455956 100644
--- a/src/crypto/rx/RxDataset.h
+++ b/src/crypto/rx/RxDataset.h
@@ -28,9 +28,10 @@
 #define XMRIG_RX_DATASET_H
 
 
+#include "base/tools/Object.h"
 #include "crypto/common/Algorithm.h"
 #include "crypto/randomx/configuration.h"
-#include "base/tools/Object.h"
+#include "crypto/rx/RxConfig.h"
 
 
 struct randomx_dataset;
@@ -49,7 +50,7 @@ class RxDataset
 public:
     XMRIG_DISABLE_COPY_MOVE_DEFAULT(RxDataset)
 
-    RxDataset(bool hugePages, bool cache);
+    RxDataset(bool hugePages, bool cache, RxConfig::Mode mode);
     RxDataset(RxCache *cache);
     ~RxDataset();
 
@@ -69,9 +70,10 @@ public:
 private:
     void allocate(bool hugePages);
 
-    int m_flags                = 0;
-    randomx_dataset *m_dataset = nullptr;
-    RxCache *m_cache           = nullptr;
+    const RxConfig::Mode m_mode = RxConfig::FastMode;
+    int m_flags                 = 0;
+    randomx_dataset *m_dataset  = nullptr;
+    RxCache *m_cache            = nullptr;
 };
 
 
diff --git a/src/crypto/rx/RxNUMAStorage.cpp b/src/crypto/rx/RxNUMAStorage.cpp
index 6d8ec167..646d3d95 100644
--- a/src/crypto/rx/RxNUMAStorage.cpp
+++ b/src/crypto/rx/RxNUMAStorage.cpp
@@ -198,7 +198,7 @@ private:
             return;
         }
 
-        auto dataset = new RxDataset(hugePages, false);
+        auto dataset = new RxDataset(hugePages, false, RxConfig::FastMode);
         if (!dataset->get()) {
             printSkipped(nodeId, "failed to allocate dataset");
 
@@ -339,14 +339,14 @@ xmrig::RxDataset *xmrig::RxNUMAStorage::dataset(const Job &job, uint32_t nodeId)
 std::pair<uint32_t, uint32_t> xmrig::RxNUMAStorage::hugePages() const
 {
     if (!d_ptr->isAllocated()) {
-        return { 0u, 0u };
+        return { 0U, 0U };
     }
 
     return d_ptr->hugePages();
 }
 
 
-void xmrig::RxNUMAStorage::init(const RxSeed &seed, uint32_t threads, bool hugePages)
+void xmrig::RxNUMAStorage::init(const RxSeed &seed, uint32_t threads, bool hugePages, RxConfig::Mode)
 {
     d_ptr->setSeed(seed);
 
diff --git a/src/crypto/rx/RxNUMAStorage.h b/src/crypto/rx/RxNUMAStorage.h
index 3afdd81d..bed0bc75 100644
--- a/src/crypto/rx/RxNUMAStorage.h
+++ b/src/crypto/rx/RxNUMAStorage.h
@@ -53,7 +53,7 @@ public:
 protected:
     RxDataset *dataset(const Job &job, uint32_t nodeId) const override;
     std::pair<uint32_t, uint32_t> hugePages() const override;
-    void init(const RxSeed &seed, uint32_t threads, bool hugePages) override;
+    void init(const RxSeed &seed, uint32_t threads, bool hugePages, RxConfig::Mode mode) override;
 
 private:
     RxNUMAStoragePrivate *d_ptr;
diff --git a/src/crypto/rx/RxQueue.cpp b/src/crypto/rx/RxQueue.cpp
index 6614d407..b420d0c2 100644
--- a/src/crypto/rx/RxQueue.cpp
+++ b/src/crypto/rx/RxQueue.cpp
@@ -90,11 +90,11 @@ std::pair<uint32_t, uint32_t> xmrig::RxQueue::hugePages()
 {
     std::lock_guard<std::mutex> lock(m_mutex);
 
-    return m_storage && m_state == STATE_IDLE ? m_storage->hugePages() : std::pair<uint32_t, uint32_t>(0u, 0u);
+    return m_storage && m_state == STATE_IDLE ? m_storage->hugePages() : std::pair<uint32_t, uint32_t>(0U, 0U);
 }
 
 
-void xmrig::RxQueue::enqueue(const RxSeed &seed, const std::vector<uint32_t> &nodeset, uint32_t threads, bool hugePages)
+void xmrig::RxQueue::enqueue(const RxSeed &seed, const std::vector<uint32_t> &nodeset, uint32_t threads, bool hugePages, RxConfig::Mode mode)
 {
     std::unique_lock<std::mutex> lock(m_mutex);
 
@@ -114,7 +114,7 @@ void xmrig::RxQueue::enqueue(const RxSeed &seed, const std::vector<uint32_t> &no
         return;
     }
 
-    m_queue.emplace_back(seed, nodeset, threads, hugePages);
+    m_queue.emplace_back(seed, nodeset, threads, hugePages, mode);
     m_seed  = seed;
     m_state = STATE_PENDING;
 
@@ -156,7 +156,7 @@ void xmrig::RxQueue::backgroundInit()
                  Buffer::toHex(item.seed.data().data(), 8).data()
                  );
 
-        m_storage->init(item.seed, item.threads, item.hugePages);
+        m_storage->init(item.seed, item.threads, item.hugePages, item.mode);
 
         lock = std::unique_lock<std::mutex>(m_mutex);
 
diff --git a/src/crypto/rx/RxQueue.h b/src/crypto/rx/RxQueue.h
index 28407a87..2d77e2e3 100644
--- a/src/crypto/rx/RxQueue.h
+++ b/src/crypto/rx/RxQueue.h
@@ -29,6 +29,7 @@
 
 
 #include "base/tools/Object.h"
+#include "crypto/rx/RxConfig.h"
 #include "crypto/rx/RxSeed.h"
 
 
@@ -52,14 +53,16 @@ class RxDataset;
 class RxQueueItem
 {
 public:
-    RxQueueItem(const RxSeed &seed, const std::vector<uint32_t> &nodeset, uint32_t threads, bool hugePages) :
+    RxQueueItem(const RxSeed &seed, const std::vector<uint32_t> &nodeset, uint32_t threads, bool hugePages, RxConfig::Mode mode) :
         hugePages(hugePages),
+        mode(mode),
         seed(seed),
         nodeset(nodeset),
         threads(threads)
     {}
 
     const bool hugePages;
+    const RxConfig::Mode mode;
     const RxSeed seed;
     const std::vector<uint32_t> nodeset;
     const uint32_t threads;
@@ -77,7 +80,7 @@ public:
     bool isReady(const Job &job);
     RxDataset *dataset(const Job &job, uint32_t nodeId);
     std::pair<uint32_t, uint32_t> hugePages();
-    void enqueue(const RxSeed &seed, const std::vector<uint32_t> &nodeset, uint32_t threads, bool hugePages);
+    void enqueue(const RxSeed &seed, const std::vector<uint32_t> &nodeset, uint32_t threads, bool hugePages, RxConfig::Mode mode);
 
 private:
     enum State {
diff --git a/src/version.h b/src/version.h
index 3b976705..1cbae1b0 100644
--- a/src/version.h
+++ b/src/version.h
@@ -28,15 +28,15 @@
 #define APP_ID        "xmrig"
 #define APP_NAME      "XMRig"
 #define APP_DESC      "XMRig miner"
-#define APP_VERSION   "5.0.1"
+#define APP_VERSION   "5.1.0"
 #define APP_DOMAIN    "xmrig.com"
 #define APP_SITE      "www.xmrig.com"
 #define APP_COPYRIGHT "Copyright (C) 2016-2019 xmrig.com"
 #define APP_KIND      "miner"
 
 #define APP_VER_MAJOR  5
-#define APP_VER_MINOR  0
-#define APP_VER_PATCH  1
+#define APP_VER_MINOR  1
+#define APP_VER_PATCH  0
 
 #ifdef _MSC_VER
 #   if (_MSC_VER >= 1920)