diff --git a/src/backend/opencl/runners/OclBaseRunner.cpp b/src/backend/opencl/runners/OclBaseRunner.cpp
index 71b35edc4..d8497e94f 100644
--- a/src/backend/opencl/runners/OclBaseRunner.cpp
+++ b/src/backend/opencl/runners/OclBaseRunner.cpp
@@ -23,10 +23,11 @@
  */
 
 
+#include "backend/opencl/runners/OclBaseRunner.h"
 #include "backend/opencl/cl/OclSource.h"
 #include "backend/opencl/OclCache.h"
 #include "backend/opencl/OclLaunchData.h"
-#include "backend/opencl/runners/OclBaseRunner.h"
+#include "backend/opencl/runners/tools/OclSharedState.h"
 #include "backend/opencl/wrappers/OclError.h"
 #include "backend/opencl/wrappers/OclLib.h"
 #include "base/io/log/Log.h"
@@ -34,6 +35,9 @@
 #include "crypto/common/VirtualMemory.h"
 
 
+constexpr size_t oneGiB = 1024 * 1024 * 1024;
+
+
 xmrig::OclBaseRunner::OclBaseRunner(size_t id, const OclLaunchData &data) :
     m_algorithm(data.algorithm),
     m_ctx(data.ctx),
@@ -93,16 +97,17 @@ void xmrig::OclBaseRunner::build()
 
 void xmrig::OclBaseRunner::init()
 {
-    m_queue  = OclLib::createCommandQueue(m_ctx, data().device.id());
+    m_queue = OclLib::createCommandQueue(m_ctx, data().device.id());
 
-    constexpr size_t oneGiB = 1024 * 1024 * 1024;
-    size_t size             = bufferSize();
+    size_t size = align(bufferSize());
 
     if (size < oneGiB && data().device.vendorId() == OCL_VENDOR_AMD && data().device.freeMemSize() >= oneGiB) {
-        size = oneGiB;
+        m_buffer = OclSharedState::get(data().device.index()).createBuffer(m_ctx, size, m_offset);
+    }
+    else {
+        m_buffer = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, size);
     }
 
-    m_buffer = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, size);
     m_input  = createSubBuffer(CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY, Job::kMaxBlobSize);
     m_output = createSubBuffer(CL_MEM_READ_WRITE, sizeof(cl_uint) * 0x100);
 }
diff --git a/src/backend/opencl/runners/tools/OclSharedData.cpp b/src/backend/opencl/runners/tools/OclSharedData.cpp
index 51571c027..355ab7755 100644
--- a/src/backend/opencl/runners/tools/OclSharedData.cpp
+++ b/src/backend/opencl/runners/tools/OclSharedData.cpp
@@ -31,11 +31,30 @@
 #include "crypto/rx/RxDataset.h"
 
 
+#include <algorithm>
 #include <cinttypes>
 #include <stdexcept>
 #include <thread>
 
 
+constexpr size_t oneGiB = 1024 * 1024 * 1024;
+
+
+cl_mem xmrig::OclSharedData::createBuffer(cl_context context, size_t size, size_t &offset)
+{
+    std::lock_guard<std::mutex> lock(m_mutex);
+
+    offset += size * m_offset++;
+    size   = std::max(size * m_threads, oneGiB);
+
+    if (!m_buffer) {
+        m_buffer = OclLib::createBuffer(context, CL_MEM_READ_WRITE, size);
+    }
+
+    return OclLib::retain(m_buffer);
+}
+
+
 uint64_t xmrig::OclSharedData::adjustDelay(size_t id)
 {
     if (m_threads < 2) {
@@ -113,6 +132,8 @@ uint64_t xmrig::OclSharedData::resumeDelay(size_t id)
 
 void xmrig::OclSharedData::release()
 {
+    OclLib::release(m_buffer);
+
 #   ifdef XMRIG_ALGO_RANDOMX
     OclLib::release(m_dataset);
 #   endif
diff --git a/src/backend/opencl/runners/tools/OclSharedData.h b/src/backend/opencl/runners/tools/OclSharedData.h
index 75ce04830..caf7b4d66 100644
--- a/src/backend/opencl/runners/tools/OclSharedData.h
+++ b/src/backend/opencl/runners/tools/OclSharedData.h
@@ -45,13 +45,16 @@ class OclSharedData
 public:
     OclSharedData() = default;
 
+    cl_mem createBuffer(cl_context context, size_t size, size_t &offset);
     uint64_t adjustDelay(size_t id);
     uint64_t resumeDelay(size_t id);
     void release();
     void setResumeCounter(uint32_t value);
     void setRunTime(uint64_t time);
 
-    inline OclSharedData &operator++() { ++m_threads; return *this; }
+    inline size_t threads() const       { return m_threads; }
+
+    inline OclSharedData &operator++()  { ++m_threads; return *this; }
 
 #   ifdef XMRIG_ALGO_RANDOMX
     cl_mem dataset() const;
@@ -59,8 +62,10 @@ public:
 #   endif
 
 private:
+    cl_mem m_buffer           = nullptr;
     double m_averageRunTime   = 0.0;
     double m_threshold        = 0.95;
+    size_t m_offset           = 0;
     size_t m_threads          = 0;
     std::mutex m_mutex;
     uint32_t m_resumeCounter  = 0;