Even if AMD isn’t as popular as NVIDIA in this sector, the MI300X still has its answer. If we look at the figures in relation to the chameleon’s H100, we find :
More VRAM
Greater memory bandwidth
Superior FP8 and FP16 performance
In 1 vs 1, the card is 20% faster than models such as Llama 2 70B or FlashAttention 2. In 8 vs 8, in a server configuration, AMD’s advantage is 40% over Llama 2 and 60% over FlashAttention 2.
NVIDIA’s difficulties in satisfying demand for its cards are turning in favor of the reds. Indeed, companies are keen to acquire hardware quickly to train their AI models, and don’t necessarily have time to waste. These include companies such as Microsoft, Meta and others. Estimates put AMD’s secure market share at 7%. Good news for the reds!
