AMD recently presented its latest technological advancement, the Instinct MI300X GPU, at the Data Center and AI Technology Premiere event on June 15th. In the keynote address, specific information regarding the power consumption of this new accelerator model was not disclosed. 

However, after thorough investigation, Hoang Anh Phu uncovered the relevant details in AMD's post-event footnotes. A comparative analysis demonstrates that the MI300X's total board power (TBP) amounts to 750W, surpassing its predecessor, the MI250X, which ranged from 500-560W. Earlier projections from Giga Computing's roadmap had already hinted at server-grade GPUs reaching the 700W threshold.

This new AMD offering, referred to as a "leadership generative AI accelerator," boasts 304 CDNA 3 compute units, a notable upgrade compared to the MI250X's 220 (CDNA 2) CUs. Additionally, engineers have introduced new 24GB HBM3 stacks, enabling the MI300X to support a maximum memory capacity of 192GB. In contrast, the MI250X, equipped with slower HBM2E stacks, is limited to 128GB memory capacity. Benchmark results that compare the MI300X to the H100 are highly anticipated, as they will provide valuable insights into its performance. The increased power consumption of the MI300X, now at 750W, is primarily attributed to the inclusion of additional GPU core tiles and the expanded memory capacity of 192GB, made possible by the introduction of 24GB HBM3 stacks. This surpasses the power consumption of the previous-generation (CDNA2) MI250X, which had a maximum power of 560W.

AMD Instinct MI300X Datacenter GPU gets 750W TBP