It is reported that AMD's AI specialized chip 'MI300X' demonstrated up to 2.1 times the performance performance in the same environment as NVIDIA chip 'H100'



AMD's AI-specialized chip '

MI300X ' announced on December 6, 2023 is touted for its superior performance compared to NVIDIA's AI-specialized chip ' H100 .' On the other hand, NVIDIA claimed that the test environment when comparing MI300X and H100 was not the same. Therefore, AMD verified the points raised by NVIDIA.

AMD strikes back at Nvidia with new MI300X benchmarks — MI300X shows 30% higher performance than H100, even with an optimized software stack | Tom's Hardware
https://www.tomshardware.com/pc-components/gpus/amd-strikes-back-at-nvidia-with-new-mi300x-benchmarks-mi300x-shows-30-higher-performance-than-h100-even- with-an-optimized-software-stack



The AI-specialized chip ``MI300X'' announced by AMD on December 6, 2023 is appealing for the fact that the GPU unit of MI300X outperforms the H100 in many applications.

Can AMD's AI specialized chip 'MI300X' surpass NVIDIA's chip 'H100' which is also used for ChatGPT? -GIGAZINE



On the other hand, NVIDIA claims that ``When comparing the performance of MI300X and H100, we did not use the software `` TensorRT-LLM '' optimized for H100.'' With the right software, the H100's performance can exceed that of the MI300X.

According to NVIDIA, AI-specific chips such as the H100 are designed to operate optimally with NVIDIA's proprietary 'TensorRT-LLM', and that ' vLLM ', which is widely used in open source, may degrade performance. About.

Therefore, NVIDIA conducted a performance comparison between MI300X and H100 when using TensorRT-LLM. After comparing the number of queries that can be processed per second, it was reported that the H100 overwhelmingly outperformed the MI300X.



In response to this, AMD also conducted a test to compare the performance of MI300X and H100. According to AMD, when both use vLLM and compare at

FP8 , MI300X shows approximately 2.1 times the score of H100, and even when H100 uses TensorRT-LLM and MI300X uses vLLM, It is said that MI300X has a score of about 1.3 times that of H100. Furthermore, when comparing H100 with 'FP8 using TensorRT-LLM' and MI300X with 'FP16 using vLLM', it became clear that the latency of MI300X was approximately 0.1 seconds lower.



Tom's Hardware, an overseas media outlet, says, ``It is up to NVIDIA to decide how to respond to this AMD performance test.''

Related Posts:

in Hardware, Posted by log1r_ut