Recently we covered the Ampere Altra Launch which is a more modern chip design from the Arm-server CPU vendor. A few days later, I was able to stop by the Ampere headquarters and see the Ampere Altra Q80-30 in action. Since then, the world has changed, but Ampere has been working on new Arm server processors.
The 2020 Ampere Altra Generation
Our discussion started with the current generation. Here is the official SKU list where the Ampere Altra Q80-30 seen earlier this year is one of the top-bin SKUs:
The SKUs are differentiated by core counts, clock speeds, and TDP. Given the correlation between TDP and clock speeds, I asked why these are not a single part that can be tuned via a clock/ TDP slider. Ampere said this is possible, but it needed SKUs to match the current selling model that customers are accustomed to.
From what Ampere told STH on a pre-briefing call, the Altra chips are now available through systems for large customers. If you are buying thousands or tens of thousands of Ampere servers, you can now get them through OEMs. For smaller markets, Ampere is adding Phoenics Electronics, an Avnet Company as a distribution partner for selling smaller numbers of systems. Ampere still has a long way to go building out a robust distribution market.
While the Altra chips are starting to ship, the next update is coming.
Ampere Altra Max at 128 Cores in 2021
The Ampere Altra Q80-30 has 80 cores, but the company plans to begin sampling 128 core parts in Q4 2020 which would put volume sales in 2021. The new line the company is calling the Ampere Altra Max.
With the Ampere Altra Max TDPs rise to 250 watts matching the top-bin Intel Xeon Platinum 8380H 3rd Generation Intel Xeon Scalable Cooper Lake parts. Since Ice Lake Xeons will hit 270W an AMD EPYC 7002 generations from 2019 are hitting 280W, the 250W seems like a lower target than what we would expect from a chip looking to enter the market in 2021.
While the chip will have the same L1/ L2 caches as the original Altra, we asked about the L3 cache. The company said that it is still being evaluated. Although Ampere did not say this, our sense is that designing a monolithic 7nm die this large will mean that Ampere needs to take a look at both yields as well as thermals of the Altra Max and that is why the shared L3 cache sizes have not been fully set at this point.
I told the Ampere team that the current customer list is disappointing. As much as I love Packet and Cloudflare, and think they are good testimonials, they are also known for being reference customers for every Arm server CPU launch. Cloudflare, for example, gave a glowing testimonial for the Qualcomm Centriq along with Microsoft. Qualcomm decided to no longer market that server CPU line. As a quick note, Qualcomm and Microsoft are investors in Cloudflare. Our sense is that Ampere likely has some larger customers, but is simply unable to disclose who they are yet.
Looking ahead, the company has now taped out its 5nm generation of CPUs so it is moving ahead with an aggressive roadmap.
“250W seems like a lower target than what we would expect from a chip looking to enter the market in 2021”
Why? A quick calculation shows 128 cores at 250W would run at 2.8GHz, so it gives a 36% increase in compute over Q80-33 at the same power. Wouldn’t that be hard for EPYC 3 to match?