Now AMD is accusing Nvidia of incorrectly comparing the H100 and Instinct MI300X accelerators. New AMD data shows its product is faster

Companies are actively trying to prove their superiority

A couple of days ago, Nvidia published a material where it said that at its recent presentation, AMD incorrectly compared the Instinct MI300X accelerator with the Nvidia H100. And if you compare correctly, Nvidia's solution will be faster. Now AMD has responded to the competitor's attack, saying that the competitor is making the wrong comparisons. 

AMD notes that Nvidia used three tricks at once in its comparison:  

• For the H100 accelerator, the TensorRT-LLM library was used instead of vLLM in AMD tests 
• Comparison of performance in FP16 mode on AMD Instinct MI300X with FP8 mode on H100 
• AMD's published performance data has been inverted from relative latency to absolute throughput 

As a result, AMD has now retested the accelerators, supposedly comparing them as correctly as possible. The company made two comparisons at once: with the vLLM library for both accelerators and with different libraries.  

As you can see, the Instinct MI300X, according to AMD tests, is still significantly faster.