Ryzen 8700G Incorrect iGPU memory read speed in GPGPU test.

[Alexander_A]

In GPGPU test Ryzen 8700G shows:
Memory Read GPU 38280 Mb/s - CPU 63740 Mb/s
Memory Write GPU 36536 Mb/s - CPU 88447 Mb/s
Memory Copy GPU 75385 Mb/s - CPU 74036 Mb/s

So it can not be true because GPU and CPU have same memory controller. Interesting that hwinfo64 shows same values when GPGPU test running, but in LLM inference with gpu only i see full memory read speed (about 64GB/sec).

Same situation on Ryzen 9950X when you testing iGPU.

[Fiery]

11 hours ago, Alexander_A said:

In GPGPU test Ryzen 8700G shows:
Memory Read GPU 38280 Mb/s - CPU 63740 Mb/s
Memory Write GPU 36536 Mb/s - CPU 88447 Mb/s
Memory Copy GPU 75385 Mb/s - CPU 74036 Mb/s

So it can not be true because GPU and CPU have same memory controller. Interesting that hwinfo64 shows same values when GPGPU test running, but in LLM inference with gpu only i see full memory read speed (about 64GB/sec).

Same situation on Ryzen 9950X when you testing iGPU.

It can be true since the GPU column lists the memory bandwidth that the iGPU can actually utilize through OpenCL.  While the CPU column lists the full data path available for the CPU without the overhead of OpenCL and the overhead of the CPU-iGPU connection.

[Alexander_A]

As i wrote before LM Studio & llama.cpp(vulkan) shows(via hwinfo) full speed of memory read in inference with iGPU only - 64...69 Gb/sec. It can not be so bad that in OpenCL speed is 2 times slower. I will search some another tests to compare. 

[Alexander_A]

Okay, deepseek 😄 explained to me why OpenCL is slow when measuring memory read speed. But then this benchmark doesn't show the full capabilities of the iGPU. 

[Fiery]

On 6/18/2026 at 3:01 PM, Alexander_A said:

Okay, deepseek 😄 explained to me why OpenCL is slow when measuring memory read speed. But then this benchmark doesn't show the full capabilities of the iGPU. 

Our benchmark reflects the performance of GPUs using OpenCL.  When there's a bottleneck due to OpenCL limitations then AIDA64 will reflect that too.