I started my tech career in 1990 selling NCR’s UNIX and X86-based Intel servers to large financial institutions. Back then, mini-computers were all the rage and Intel-based servers were not even considered industrial strength enough for the datacenter but good enough for print servers. My how things have changed- X86-based servers now dominate the datacenter and mini-computers are nearly dead.
Eight failed attempts at Arm servers
Five years ago, IT did not consider Arm-based servers industrial-strength either, what looked like a life sentence of edge compute, and who could have blamed folks then? By my count, there were eight unsuccessful attempts at general purpose Arm-based datacenter processors: Marvell V1, Marvell V2, Calxeda, Samsung, AMD, Qualcomm, Broadcom, and APM. Literally, companies invested billions in cash with zero payback. I know the reason for the demise of each of these chips whether it be lack of a software ecosystem, not enough performance, lack of a systems ecosystem, bad timing, etc.
AWS provides Arm datacenter server credibility
But then in 2019 things changed. AWS introduced Graviton 1 based on its Nitro edge compute and then what seemed like overnight, the peanut gallery said that general-purpose, Arm-based servers had arrived. I was part of that peanut gallery and through a combination of Arm and AWS investments, a market was created. That market may have been for AWS users, but how about an Arm-based server cloud instance for everyone else? Enter Ampere Computing.
Ampere arrives on the scene
In just a few years, Ampere has racked up an impressive array of design wins from the world-class cloud providers and a particularly important on-prem, as a service cloud vendor, HPE. Today, end customers can buy Ampere-based compute instances from sixteen different companies from around the world.
I have never in my career seen this quick of an uptake of a new server chip vendor and I ran AMD’s corporate marketing group during Opteron. While I would like to spend more time on where Ampere is going, I will spend a little time talking about how Ampere had this level of success so quickly with cloud providers.
CSPs are different
First, off, it is important to understand that public cloud service software demands are different from legacy on-prem software like SAP, VMware, and Windows Server. Early on, Ampere optimized its processor and platforms for cloud workloads that favored high single threaded integer performance with consistent delivered and scalable performance at the lowest power draw. Ampere managed to cram 3,328 of these cores into a standard, 12kW rack, 3-4x AMD and Intel. Ampere’s super-dense design drove high performance and performance per watt for web services like NGINX, MySQL databases, in-memory caching like Redis and H.264 media transcoding.
More than IP
Some are confused and think that a vendor like Ampere just licenses technology from Arm, easily integrates it together, and goes to market with chips like Ampere Altra and Ampere Altra Max. Arm deserves a lot of credit for its investments and IPs, but there’s a tremendous amount of work to be done to then transform that IP into a performant, low power and reliable server solution. For Altra and Altra Max, Ampere does license the Arm N1 core and other Arm IPs like the MMU. Ampere also licenses key IPs from Cadence and Synopsis. Ampere creates a lot of its own IPs as well for its current product line related to optimizing the mesh, power delivery, and scaling I/O to larger core/socket counts, and integrates it with the Arm, Cadence, and Synopsis IPs to create the SoC.
You would think we would be done at this point, but you would be wrong. Ampere must create hardware and software platforms to make the SOC useful. The scalable hardware platforms need to conform to standards and include BMCs and all the peripherals like memory and storage. Ampere then needs to create UEFI compliant firmware and along with Arm and other IP vendors, integrate hardware-enabling software. At this point, Ampere integrates the SOC, hardware, and software platforms, validates the combination, and then you have an Ampere-enabled platform ready for an ODM or OEM. And you thought the work ended at the IP. 😊 For what it is worth, I know I am grossly simplifying the effort required above. It is a lot more complex than this but the point I am trying to make is that the work does not end at creating the IP.
Ampere now has an architectural license, like Apple
But now I would like to talk out of the other side of my mouth and talk about how excited I am about Ampere’s next generation SoC, AmpereOne, that contains more home-grown IP from Ampere. Like Apple, Ampere now has an architectural license from Arm that gives it the right to create its own, custom CPU core from scratch. And what have we seen Apple do with its architectural license? It has created SOCs that have twice the CPU performance of its competitors in the smartphone space and twice the performance per watt in the notebook space.
While Ampere is already sampling AmpereOne with its custom cores and IPs, it is not providing many details beyond that it will be fabbed in TSMC 5nm, will support DDR5 and PCIe Gen5 and slot into its current socket. What we do not know yet are performance, power, or in-market dates, obviously three vital variables to assess the chip. Given what Arm is projecting for its new Neoverse N2 processor, it would make sense to me that AmpereOne would at least provide 25% better power or performance or else why put all that investment into a custom core? Seeing what Alibaba did with the Yitian 710 and the prowess of Ampere’s engineering team, 25% does not seem like a stretch to me.
Like Apple, I would expect as Ampere gets more years of experience with its custom core it will look to license even less and create its own IPs around memory and IO where it could strive to be first, for example, with a memory type. I could see Ampere creating its own mesh to connect all its high-speed to get a competitive advantage in the marketplace. Just like Apple.
Arm-based cloud datacenter servers were over a decade in the making. Like X86 in the late 80’s and early 90’s Arm used to be considered unworthy in the datacenter. Now times have changed, and we have companies like Ampere Computing offering its cloud optimized instances via sixteen major public cloud providers and soon to be HPE for on-prem cloud. Like Apple, Ampere has taken an Arm architectural license for its AmpereOne SoC that could outperform anything in the Arm SoC market on a performance per watt basis. I am looking forward to seeing how AmpereOne performs, and we should know soon as it is sampling since May. It is an exciting time to be a compute consumer with the increased competitiveness brought companies like Ampere.
Note: Moor Insights & Strategy writers and editors may have contributed to this article.