AMD will fully embrace CPU and GPU HBM?
AMD will fully embrace CPU and GPU HBM?
Recently, relevant information broke that AMD's next-generation Zen 4-core EPYC Genoa processor may be equipped with HBM content to form a formation with Intel's next-generation server CPU Xeon Sapphire Rapids. The recent Linux kernel patch also revealed that AMD’s next generation of cDNA 2-core-based Instinct MI200 GPUs will also use HBM2e with up to 128GB of video memory, which means that AMD is likely to embrace HBM in the server market.
HBM is no longer relevant to the consumer graphics market
HBM is a high-bandwidth memory, actually the first high-performance DRAM developed by AMD. In order to realize this vision, AMD introduced SK Hynix with 3D stacking process experience, and jointly developed HBM memory with the help of interconnection and packaging manufacturers.
AMD is also the first company to introduce it to the GPU market and apply it to its Fiji GPU. Samsung took the lead in mass production of HBM2, AMD was snatched by Nvidia, and Nvidia was the first to apply this new memory standard to its Tesla P100 accelerator card.
When the advantages and disadvantages of HBM are obvious, the initial bandwidth advantage is overtaken by GDDR6, and the design difficulty and cost are an insurmountable obstacle. Although these costs do not account for the bulk of high-end graphics cards, using HBM's low-end graphics cards is even more painful. But AMD did not abandon HBM2, but continued to introduce HBM2 on Vega graphics cards.
This may be the last time we have seen HBM on consumer GPUs. AMD has never used HBM in the subsequent RDNA architecture. Only those based on the CDNA architecture and used for throttle GPUs are still using HBM.
Why is the server market?
How did HBM take root in the server market? This is because one of the most suitable application scenarios for HBM is in power-constrained environments that require maximum bandwidth, which is very suitable for artificial intelligence computing in HPC clusters, or large and dense computing data centers.
This is why these companies with data center operations continue to use HBM, while NVIDIA is still using HBM2 and HBM2e in its powerful server GPU A100, and may even continue to use HBM2 and HBM2e in its next-generation Hopper architecture. Intel’s unreleased Xe-HP and Xe-HPC GPUs are also rumored to use HBM.
The consumer GPUs of the two manufacturers accidentally avoided HBM and chose GDDR6 and GDDR6X. It is conceivable that they do not want to take AMD's detours.
As for AMD's pioneering use of HBM on the CPU, in a patent issued by AMD last year, HBM appeared in the chip design. Intel’s competing Xeon Sapphire Rapids server CPU also officially announced the use of HBM, but mass production will not be until 2023. All these show how popular HBM is in the server market, and they have begun to transfer HBM to CPUs.
Next generation HBM
Although the standard-setting JEDEC has not yet released the HBM3 specifications, SK Hynix, which has been committed to the next generation of HBM, disclosed the latest information on HBM3 in June this year, which will usher in further performance improvements.
SK Hynix can achieve such a high performance improvement, most likely because it signed a patent license agreement with Xperi last year. These agreements include DBI Ultra 2.5D/3D interconnect technology, which can be used for the innovative development of 3DS, HBM2, HBM3 and subsequent DRAM products. Traditional copper pillar interconnects can only achieve 625 interconnects per square millimeter, while DBI Ultra can achieve 100,000 interconnects in the same area.
Since JEDEC announced the HBM2E standard in 2018, HBM has not been updated for nearly 3 years. Samsung even announced the development of HBM-PIM with an artificial intelligence engine in February this year. Whether HBM3 can continue to dominate the server field in the future, I believe that the proportion of HBM in server products planned by several major manufacturers has already given the answer.