Ampere Altra Max M128-30 128 Core Arm Server Update

Ampere Altra Q80 33 CPU In Wiwynn Socket
Ampere Altra Q80 33 CPU In Wiwynn Socket

Ampere has a new member of the family, the Ampere Altra MAX that is designed to offer even more cores in a single socket than its 80-core predecessor. These new chips will come out in 2021 and be socket compatible with the Ampere Altra we saw in the Wiwynn Mt. Jade server a few months ago, just with more performance due to a higher core count. Now, Ampere is giving an update on the progress of its new solution.

Ampere Altra Max M128-30 Update

Just by way of the roadmap, the 7nm part was defined as a socket-compatible DDR4/ PCIe Gen4 upgrade to the Amper Altra from 2020. This new part effectively adds more cores to the same sockets. For a cloud provider, the basic idea is that one can get fewer cores/ threads per socket with Intel, or fewer cores and the same number of threads per socket as AMD with Ampere. The Ampere Altra Max trick is that this is with all cores since it does not use SMT making the Max a 128C/ 128T part.

Ampere Altra Max Planned Sample 2020 And Shipping In 2021
Ampere Altra Max Planned Sample 2020 And Shipping In 2021

Ampere posted an update on the progress of its solution this week. Here is the company’s new performance chart.

Ampere Altra Max March 2021 Update
Ampere Altra Max March 2021 Update

What Ampere is showing is its scaling from 80 cores to 128 cores. Frankly, this is a good way to show its progress. Moving from 80 to 128 cores is a 60% increase in core count, and the -30 is usually the clock speed indicator for Ampere which would make us think these are the same clock speed. While Ampere is not getting a full 60% performance scaling, it is getting 51-57% uplift which is fairly good.

Final Words

Overall, it will awesome to see these chips come out over the next few months. The fact that Ampere is already testing them means that it is at least able to produce such large parts. Perhaps the biggest challenge now is building momentum for the platforms and products that will take the chip to market. We covered that the Wiwynn solution is an important step since that allows cloud providers, like Oracle or that class of customers, to build an Arm solution without having to get into chip design, and using a cloud-scale hardware provider. Gigabyte also has solutions in the market. Still, AMD Milan sounds like it has a lot of momentum and Intel will have a modern platform with the Ice Lake soon. It is not lost on us that Ampere is releasing this Altra Max data just before the “Milan Milano” March 15, 2021 AMD EPYC 7003 launch that we will cover on STH.


  1. Fairly good scaling? That’s extremely impressive in my book, especially if the additional cores add any memory latency.

    I try not to be a fanboy for any one architecture, but I want to see arm do well simply from a competitive standpoint. I don’t own any Apple products, so I’m reserving judgement when it comes to their processors, but I’ve thought the move to arm was a brilliant decision on their part from a business standpoint.

  2. Price, TDP, Peak Clock Speed, Sustained Clock Speed? This will be very interesting to Cloud Vendors as everything are priced per vCPU, which is one thread on x86 and one actual core on ( most ) ARM.

  3. Ampere is very promising but it’s hard to justify CapEx on a startup’s CPU because if they go bust you’re out of luck firmware / microcode / support wise.

    I’m sure they’ll be acquired at some point, though.

  4. Patrick sez “I’m sure they’ll be acquired at some point, though.” Ampere has several investors. Lenovo, NVidia and ARM. Now that NVidia has bought ARM, NVidia is now the lead investor in Ampere. So as far as being acquired, no one really needs to. Ampere has everything they need now to operate.

  5. spuwho: “now that nvidia has bought arm” — IIRC this is still not closed due to regulators and as nxp/qcom case shows they do have major voice there.


Please enter your comment!
Please enter your name here

This site uses Akismet to reduce spam. Learn how your comment data is processed.