Ampere Will Launch 128-Core ARM CPUs by the End of 2021
Ampere has announced an upcoming 7nm variant of its existing Altra architecture that will scale the CPU up to 128 cores, with the new CPU shipping later in 2021. Additionally, the company will transition to its own custom CPU architecture rather than continuing to tap ARM’s Neoverse line of products.
ServeTheHome wrote an excellent review of Ampere’s current 80-core Q80-33 back in December, if you’re looking for more information on how the company’s current chips compare against what’s on tap from Intel and AMD. The Altra doesn’t support SMT, so the CPU offers up to 80 cores currently and will soon scale up to 128. Performance from dual Q80-33 CPUs (160C/160T) often scaled well against dual Epyc 7742 CPUs (128C/256T). For its 128-core Ampere Altra Max, Ampere is predicting overall performance uplift of 1.5x – 1.6x per core over AMD’s Zen 2-based Epyc hardware:
Ampere is predicting more-or-less linear scaling here; 128 CPU cores works out to a 1.6x increase over an 80-core CPU. No details about the media encoding workload were provided, but the 1.5x uplift suggests the workload is only somewhat dependent on per-core memory bandwidth. As our recent TRACBench benchmark debut illustrates, some media encoding workloads put more pressure on main memory than others. Ampere’s scales quite well to the 128-core mark.
After 2021, Ampere will transition to 5nm and a next-generation CPU design based on its own architecture rather than an ARM-licensed chip. We don’t have details on the new core yet, beyond that it will be ARM ISA-compliant, and that the CPU will offer larger amounts of memory and I/O bandwidth. According to Ampere, it can exceed the recent performance predictions from ARM’s Neoverse V1 and N2 platforms, offering even better results.
At Long Last, ARM Is Coming for the x86 Server Market
It’s been over 10 years since Calxeda announced a 480-core server it had in development, built from some 120 quad-core Cortex A9 CPU cores. We’ve seen plenty of companies announce their intent to compete with ARM-based servers or bring products to market, including AMD and Qualcomm (via Centriq), only to back away later.
But the market in 2021 is different than the market in 2011, or even in 2016. Nvidia has a server CPU, Grace, headed to market in 2023. Amazon has fielded its own ARM servers based on custom silicon for use in AWS cloud instances. The fastest computer on Earth, Fugaku, is based on a custom ARM processor built by Fujitsu, the A64FX. ARM is investing more in building ecosystem support for servers and data centers, as evidenced by the recent Neoverse announcements.
This slide, courtesy of Next Platform, shows the rise of Graviton2 support as a percentage of all EC2 instances (the pie chart to the right shows the percentage of Graviton2 instances added in 2020, not the total install base). Vendor A is obviously Intel, Vendor B is AMD. AMD has taken share from Intel, to be sure, and appears to have roughly doubled its market share, but Graviton has grown at a considerably faster rate. We may not know how much a Graviton2 costs Amazon, specifically, but it’s cheaper than an x86 core purchased from a different company.
Discussing the rise of ARM in servers naturally invites contemplation of what AMD and Intel are doing about the problem. The two companies are taking somewhat different approaches. For AMD, taking space in server is more about out-competing Intel than pivoting to specifically attack ARM right now. Zen 3 improved overall Epyc performance, though power consumption figures vary depending on workload. For Intel, the recent Ice Lake-SP launch was a chance to make up some ground against AMD and position itself more effectively in the future. Intel’s gross margins and overall data center earnings remain much larger than AMD’s, and it is arguably more exposed to the potential cost impact of effective ARM competition for this reason.
Neither AMD nor Intel will meet these challenges sitting still. AMD has been steadily improving the performance and performance per watt of Epyc with every generation, and Zen 4 should bring additional gains when it debuts on 5nm. Intel is working on hybrid CPUs with the potential to sharply improve power consumption at idle and in low-power states. Server customers tend to be conservative with their purchases, and hyperscalers like Amazon are only part of the server market. Writing x86 off from either vendor would be highly premature at this point in time. But the next few years should offer ringside seats to the kind of fight x86 hasn’t faced since Intel was the plucky upstart shoving into data centers dominated by big iron servers from the likes of IBM, HP, and Sun.
Amazon, Qualcomm, Ampere, Apple — each of these companies has one or more custom ARM SoCs, and each intends to offer its own challenge to the dominant x86 computing model over the next 2-4 years.
Feature image by Ampere.