While NVIDIA's AI GPU dominance remains unparalleled, with the company holding the third-largest market capitalization globally and its CEO, Jensen Huang, boldly claiming that even if their competitors were to offer their products for free, they still wouldn't be able to beat NVIDIA, the reality is that AMD's MI300 series has made a notable impact this generation, attracting a significant number of users.
According to an independent survey conducted by TensorWave among 82 senior engineers and AI experts, approximately 50% of respondents expressed confidence in AMD MI300X GPUs. This confidence stems from the fact that, compared to NVIDIA's H100 series, AMD MI300X not only offers better value for money but also boasts ample supply, thus avoiding price hikes due to shortages.
In fact, TensorWave itself has procured AMD MI300X GPUs, further underscoring the growing appeal of AMD's offerings in the AI GPU market.
The AMD Instinct series has seen several iterations, but it hasn't quite matched up to its competitors in terms of performance or market acceptance.
With the latest MI300X, there's been a significant performance leap. Compared to the H100, it boasts 2.4 times more memory capacity, 1.6 times higher bandwidth, and 1.3 times higher FP8 and FP16 floating-point performance. In the FlashAttention 2 and Llama2 (with 700 billion parameters) 1v1 comparison tests, it's leading by 20%. In Bloom (with 1,760 billion parameters) and Llama2 (with 700 billion parameters) 8v8 tests, it's ahead by 60% and 40%, respectively.
AMD has also introduced the MI300A, which integrates CPU and GPU for the first time in data centers, although it's not widely deployed yet.
Su Zifeng of AMD previously predicted that AMD's AI chip revenue in 2024 would reach $3.5 billion, significantly higher than the earlier estimate of $2 billion.
However, AMD still has a significant gap to close in terms of development support and ecosystem building and will need to continue striving to catch up.
