ARM server CPUs have gotten a lot of attention over the last decade, as multiple companies attempted to bring competitive products to market without much success. In the past few years, the situation has changed somewhat. While ARM accounts for just a fraction of the overall server market, companies like Ampere and Marvell are starting to carve out a place for themselves.
It’s going to be a few years, however, before they’re much of a threat to x86. That’s the conclusion reached by TrendForce, which also reports that while both AMD and Intel launched new platforms earlier this year, these systems have reportedly only shipped in limited volume. TrendForce believes adoption will rise in Q3, with Ice Lake-SP accounting for 30 percent of the server market by Q4 2021.
Intel’s Eagle Stream platform for Sapphire Rapids CPUs will arrive in 2022, though the actual ramp timing is a little uncertain. Sapphire Rapids has bounced around a bit between being slated for 2021 or 2022 and is currently expected in Q2 2022 next year. The chip should enter qualification in Q4 and ship to customers in limited volume in Q1. Much depends on yields, OEM feedback, any bugs discovered throughout the process, and the ongoing semiconductor shortage. Ramping up a new part always incurs lower yields as the factory works out bugs in the manufacturing process. Intel and AMD may both be a little conservative with which SKUs they launch and when they launch them as a result. When demand is higher than supply, it can make more sense to build slightly older equipment at maximum yield than to build newer hardware at a yield penalty.
Sapphire Rapids is expected to include between 56 and 80 cores, with an option for on-die HBM. AMD’s Genoa may also include this feature — we’ve heard rumors to that effect, but nothing solid. We also don’t know yet if Genoa will include V-Cache. Milan-X is expected to get a V-Cache refresh but how that will play against Genoa’s release date in 2022 isn’t clear yet.
AMD’s Threadrippers, especially the 3990X, would probably benefit a great deal from V-Cache because that CPU is historically memory bandwidth bound. AMD might opt to offer HBM in lieu of V-Cache, or it could theoretically offer both. This type of configuration would give AMD an unprecedented amount of cache and memory bandwidth per socket — you have to look to chips in the POWER family to find CPUs that offer anything equivalent, and CPUs like POWER9 use a very different architecture from anything AMD and Nvidia currently field.
Data by Mercury Research, Graph by Next Platform. Intel continues to dominate the market, but AMD’s market share in servers is rising despite ongoing silicon shortages. The company has given up a bit of share in desktop and laptop to gain more in server.
TrendForce notes that AMD has seen robust adoption across the market from Google Cloud Platform, Azure, and Tencent, and that it expects the company to take 15 percent of the server market in 2022. AMD’s exact market share varies a bit by how you count, but the company has ~10 percent of the market this year. AMD and Intel collectively account for 97 percent of the current server market. This is expected to continue through at least 2023 because platforms from x86 vendors will include features like PCIe 5.0 support. PCIe 5.0 will further improve SSD throughput and storage server vendors are apparently looking forward to the new interface for demanding workloads.
But a feature like PCIe 5.0 isn’t the kind of thing long-term x86 dominance can rest on. ARM vendors will be adding interfaces, too, and the ARM server ecosystem will continue to mature. Adding features like HBM, V-Cache, and Intel technologies like Foveros are all part of an effort by both AMD and Intel to offer integration and packaging technologies that their rivals don’t match.
It’s going to be very interesting to see how AMD leverages its Xilinx purchase going forward. While Intel talked up the idea of an FPGA on a Xeon chip when it bought Altera, such chips have never become a major part of its product lineup. Integrated or improved FPGA functionality represents a potential way for AMD to encourage vendors to prefer x86 over potential ARM solutions.
Ultimately, we agree with TrendForce that ARM doesn’t pose the largest risk to Intel or AMD over the next few years. Just as it will take Apple and other ARM vendors several years to ramp CPUs that compete with desktop and laptop x86 chips across the entirety of the consumer product stack, it will take ARM vendors several years to build solutions that cater to every server market and requirement.
Now Read:
- Intel’s Sapphire Rapids Slips a Bit, but Ramp Still Planned for Q2 2022
- Ampere Will Launch 128-Core ARM CPUs by the End of 2021
- AMD’s Milan Brings Zen 3 to Epyc, With Mostly Positive Results
No comments:
Post a Comment