NVIDIA slaps AMD in the face: With software support, H100's AI performance is 47% faster than MI300X!

According to news on December 14, AMD launched its most powerful AI chip Instinct MI300X earlier this month. The AI ​​performance of its 8-GPU server is 60% higher than that of Nvidia H100 8-GPU. In this regard, NVIDIA recently released a set of the latest performance comparison data between H100 and MI300X, showing how H100 can use the right software to provide faster AI performance than MI300X.

According to data previously released by AMD, the FP8/FP16 performance of MI300X has reached 1.3 times that of NVIDIA H100, and the speed of running Llama 2 70B and FlashAttention 2 models is 20% faster than H100. In the 8v8 server, when running the Llama 2 70B model, MI300X is 40% faster than H100; when running the Bloom 176B model, MI300X is 60% faster than H100.

However, it should be noted that when comparing MI300X with NVIDIA H100, AMD used the optimization library in the latest ROCm 6.0 suite (which can support the latest computing formats such as FP16, Bf16 and FP8, including Sparsity etc.) to get these numbers. In contrast, the NVIDIA H100 was not tested without the use of optimization software such as NVIDIA's TensorRT-LLM.

AMD's implicit statement on the NVIDIA H100 test shows that using vLLM v.02.2.2 inference software and the NVIDIA DGX H100 system, the Llama 2 70B query has an input sequence length of 2048 and an output sequence length of 128

The latest test results released by NVIDIA for DGX H100 (with 8 NVIDIA H100 Tensor Core GPUs, with 80 GB HBM3) show that the public NVIDIA TensorRT LLM software is used, of which v0.5.0 is used for Batch-1 Test, v0.6.1 for latency threshold measurements. Test workload details are the same as previously conducted AMD tests

According to the results, after using optimized software, the performance of the NVIDIA DGX H100 server has increased by more than 2 times, and is 47% faster than the MI300X 8-GPU server displayed by AMD

DGX H100 can handle a single inference task in 1.7 seconds. In order to optimize response time and data center throughput, cloud services set fixed response times for specific services. This allows them to combine multiple inference requests into larger "batches", thereby increasing the overall number of inferences per second on the server. Industry standard benchmarks such as MLPerf also use this fixed response time metric to measure performance

Slight tradeoffs in response time can create uncertainty in the number of inference requests the server can handle in real time. Using a fixed 2.5 second response time budget, the NVIDIA DGX H100 server can handle more than 5 Llama 2 70B inferences per second, while Batch-1 handles less than one per second.

Obviously, it is relatively fair for Nvidia to use these new benchmarks. After all, AMD also uses its optimized software to evaluate the performance of its GPUs, so why not do the same when testing the Nvidia H100?

You must know that NVIDIA's software stack revolves around the CUDA ecosystem, and after years of hard work and development, it has a very strong position in the artificial intelligence market, while AMD's ROCm 6.0 is new and has not yet been tested in real-world scenarios.

According to information previously disclosed by AMD, it has reached a large part of the deal with large companies such as Microsoft and Meta. These companies regard its MI300X GPU as a replacement for Nvidia's H100 solution.

AMD’s latest Instinct MI300X is expected to be shipped in large quantities in the first half of 2024. However, NVIDIA’s more powerful H200 GPU will also be shipped by then, and NVIDIA will also launch a new generation of Blackwell B100 in the second half of 2024. In addition, Intel will also launch its new generation AI chip Gaudi 3. Next, competition in the field of artificial intelligence seems to become more intense.

Editor: Xinzhixun-Rurounijian

The above is the detailed content of NVIDIA slaps AMD in the face: With software support, H100's AI performance is 47% faster than MI300X!. For more information, please follow other related articles on the PHP Chinese website!