AMD Instinct MI300X, a powerhouse for AI large language models
AMD discloses how much power the MI300X requires.
The OAM-based (OCP Accelerator Module) design of the MI300X graphics processor is listed with 750W on AMD slides. This was not mentioned by early reports covering the introduction day earlier this week. This board power requirements was spotted in the footnotes after the event.
AMD developed a special version of its MI300 series accelerator exclusively with CDNA3 tiles. This accelerator will be offered alongside MI300A, the so-called data center exascale APU, which features 24 Zen4 cores combined with GPU tiles.
The MI300X has increased GPU core tiles and has more memory (192GB) thanks to the recently introduced 24GB HBM3 stacks. Raw GPU power combined with lots of memory have contributed in to higher power now reaching 750W for one OAM modules. This is more than the last-gen (CDNA2) MI250X which only went up to 560W.
Over the past few years, the power requirements for GPU data-center products have increased. NVIDIA H100 SXM GPU TDP goes up to 700W and 350W for the PCIe air-cooled version. The PCIe version of Intel Max 1550 Data Center GPU based on Ponte Vecchio silicon also requires at least 600W of power.
The naming of power requirements and designs varies depending on the manufacturer. Power characteristics like TDP, TBP or Power Envelope may not always mean the same thing. But the important thing is that the GPU rivalry in the data-center space has only just started, so who knows if we are going to see an even higher power specifications listed on official slides soon.
Source: AMD via Hoang Anh Phu