Power consumption for server GPUs only goes up
A roadmap featuring Giga Computing (Gigabyte) data-center CPU/GPU architectures has been leaked by HXL.
Giga Computing is a separate entity that spun off from Gigabyte. The company specializes in the enterprise computing market and among the first to announce products based on the latest data-center architectures.
A new roadmap slide has emerged today, featuring products from all major companies who deliver cutting-edge enterprise equipment. The roadmap is divided between CPUs, GPUs and CPU+GPU (what is commonly known as APU). The roadmap shows products going as far as 2025.
CPU/GPU Server Roadmap 2022-2025, Source: Giga Computing/HXL
Giga Computing expects a transition from 270-280W CPUs from 2022 to reach as high as 500-600W by the end of 2025. These chips should include Intel’s upcoming Emerald and Granite Rapids CPU series as well as AMD’s EPYC Turing Zen5 series.
The company also expects PCIe-based GPU server solutions to exceed 500W by early 2025. In this case, the roadmap does not mention any particular product, but we do know that NVIDIA is now working on Blackwell server GPUs and AMD is more than likely to introduce CDNA4 architecture at some point in the future.
Interestingly, the roadmap does not list the AMD MI300 ‘exascale APU’, which is the next-gen AMD accelerator for data-centers already announced by the company in January this year. Similarly to the NVIDIA Grace/Hopper superchip, it would combine CPU and GPUs on a single package through advanced packaging (or in this case, stacking) technologies. According to Giga Computing, NVIDIA Grace Hopper CPU Superchip could reach 1000W.
Source: @9550pro