Cloud Hosting

Ampere Computing today made public their roadmap update concerning current and future AArch64 server processors. AmpereOne availability remains tough but the company is hoping next year to introduce a 3nm CPU with up to 256 cores and supporting 12 channel DDR5 memory.

Ampere's first update to share was that they are collaborating with Qualcomm to pair Ampere CPUs with Qualcomm Cloud AI100 Ultra accelerators. Ampere Computing CPUs will be paired with Qualcomm AI accelerators in Supermicro servers as a new option for AI inference servers. (As a reminder, the Qualcomm AI accelerator hardware does have an open-source Linux kernel driver: QAIC). This announcement came as a bit of a surprise as Ampere Computing has been promoting "GPU-Free AI Inferencing" and a lot around just CPU-based inferencing... With the Qualcomm accelerators still technically GPU-Free. Initially these LLM AI servers will be using Ampere Altra processors while in the coming "months" will have AmpereOne processor options.

And then the most exciting roadmap update: Ampere Computing is planning for a 3nm AmpereOne CPU with up to 256 cores and 12 channels of DDR5 memory support in 2025. Ampere Computing says that the next iteration of AmpereOne CPUs is "ready at fab". When inquiring about the 12 channel DDR5 memory, it's anticipated to be at DDR5-5200 speeds or higher. Twelve channels of DDR5 memory support matches that of AMD EPYC Bergamo/Genoa as well as upcoming Intel Xeon 6 processors.

Being curious if the 256 core AmpereOne was simply a die shrink and core count increase, I brought up the matter of Ampere-1B... Several months back I spotted Ampere1B appearing in the LLVM Clang compiler and also the GCC compiler. When bringing up Ampere-1B in the context of the new AmpereOne CPUs in 2025, I was able to get it confirmed that indeed will be the "1B" variant. So with that some ISA additions with Ampere-1B having Armv8.7+ with FEAT WFxT, FEAT CSSC, FEAT PAN3 and FEAT AFP extensions plus 1A additions like the Memory Tagging Extension (MTE) and SM3/SM4 cryptography.

While this teasing over a 256-core AmpereOne processor and indicating AmpereOne is "generally available", that leads to the elephant in the room: where the hell is AmpereOne? It seems next to impossible to find. Ampere Computing has been talking about AmpereOne since 2022, last May it officially was "announced" with up to 192 cores and last September AmpereOne was announced for Oracle Cloud with limited availability in Q4'2023, but it seems next to impossible to actually find in the middle of 2024.

There have also been no independent reviews/benchmarks of AmpereOne and Ampere Computing is all the happy continuing to promote Ampere Altra on social media channels, new blog posts, and with these roadmap updates. Ampere Altra has stood up well but mind you it was announced in 2020 and has been benchmarked since late 2020... Not 1~2 years later like we are now in the AmpereOne cycle. We are now at the middle of 2024 with AmpereOne supposedly "generally available" but seems next to impossible to find and still awaiting any independent AmpereOne benchmarking analysis or even access to the Oracle Cloud AmpereOne instances... Everyone I know that is independently interested in AmpereOne and trying to acquire the hardware has yet to be successful in doing so. Simply put, it's been frustrating and rather ghostly all the meanwhile Intel Xeon 6 is imminent with Sierra Forest followed by Granite Rapids for higher core counts there with 144 cores per socket and even a 288 core variant of Xeon SRF. AMD EPYC 9754 Bergamo has been available since last year with 128 cores / 256 threads per socket, Zen 4C has boasted great power efficiency improvements, and there is also AMD EPYC Zen 5 on the horizon.

When inquiring about AmpereOne availability during the advanced briefing, I was told they've been busy ramping up with their large customers and then over the months ahead they will be expanding more to their medium and smaller customers. Today's press release goes on to note that "new AmpereOne OEM and ODM platforms would be shipping within a few months." The explanation also seems a bit odd considering the lack of GA access still to AmpereOne at any of the large cloud providers. We'll see when the availability story pans out and if/when we see any AmpereOne hardware for independent reviews/benchmarking. The battle is only becoming more tough if the broad availability on AmpereOne drags out past the launch of Intel Xeon 6 Sierra Forest / Granite Rapids and AMD EPYC Zen 5 hardware. AmpereOne when originally announced was before the 128-core Bergamo even shipped while there are rumors its successor could be pushing 192 cores per socket this year and further improvements to x86_64 power efficiency.

Last year's roadmap update also signaled the AmpereOne family is for 136 to 192 cores (or now, 256 cores in 2025) while Ampere Altra is for 128 cores and less. Sadly this year's update didn't shed any light on any AmpereOne expansion for 128 cores or less or any sort of refresh for those pursuing any smaller AArch64 core counts while wanting a modern design. So it's not clear if Ampere is just pursuing the larger (cloud) servers or if in time they will release any new products for 128 cores and less. There certainly is interest and a need for AArch64 workstation type systems while for now Ampere Altra is working out for the lack of any viable competition there for socketed AArch64 desktops/workstations. It will be interesting to see what Intel Sierra Forest offers at lower core counts for pure E core efficiency and similarly what AMD is already delivering with the EPYC 8004 Siena processors for great power efficiency and scalability.

Continue reading here:
Ampere Computing 2024 Roadmap Update: 256 Core 3nm CPU In 2025 - Phoronix