Ampere Computing has delivered its annual state-of-the-company address, announcing a second AmpereOne processor model (Ampere AmpereOne II?) but withholding when it expects to sample. The company also announced a collaboration with Qualcomm to jointly tackle AI inference with Ampere processors and Qualcomm’s AI accelerator (NPU). Ampere peppered its address with product comparisons and industry commentary.
Selling the Next
- Availability—Ampere in its address claimed it’s shipping the first-gen AmpereOne but its press release states it’s still expected later this year. Online sellers show their AmpereOne systems as coming soon. Thus, talk of a next-generation processor is premature.
- Tick tock—Enabled by its 3 nm process, the new product raises the AmpereOne core count from 192 to 256. Keeping the microarchitecture the same, Ampere—helmed by Intel veteran Rene James—is following the model of reducing design changes when adopting a new process.
Riding the AI Hype
- AI hype—In the address, James stated, “We’re in an extreme hype cycle on AI.” We must allude to Ampere itself because she went on to say for the past six years the company has set out to build a microprocessor to do efficient inferencing.
- Ampere-versus-Ampere—For AI processing, Ampere compared the older, currently shipping Altra Max 128-core processor with an x86 processor attached to an Nvidia A10 card. Running Llama 3-8B, Ampere found the Altra system performed the same but needed only one-third the power. However, the A10 employs Nvidia’s Ampere architecture (no relation) from 2021, which delivers less than half the raw performance of the Nvidia L40S (Ada Lovelace) from 2023. For further context, the 8-billion-parameter Llama 3 is the large language model Qualcomm showed its Snapdragon 8 Gen 3 handset running. By contrast, the Nvidia Blackwell launch referenced it running a 1.8-trillion-parameter model.
- Qualcomm has been diligently promoting the Cloud AI 100 NPU despite disclosing little customer interest. Launched last year, the company’s newest add-in-board based on the AI 100, Ultra, raises performance and power over previous offerings, presumably by integrating additional AI 100 NPU chips. The Qualcomm-Ampere collab is a good fit because both want to show data-center customers an ecosystem and ride the AI hype wave. The two companies have also achieved similar data-center success. (Cue sad trombones.)
Delivering Efficiency but not Performance
- SpecInt—Comparing the 192-CPU, 3.2 GHz AmpereOne to the AMD Epyc 9754 (Bergamo), Ampere claims its chip is 15% more efficient on SpecInt. However, by the company’s estimates, Bergamo is 5.6% faster. Per physical core, the 128-core Bergamo is a whopping 58% faster.
- Realistic workloads—Ampere further compared performance and power for data-center workloads Nginx, Redis, Memcached, and MySQL. In all cases except Memcached, the AmpereOne processor outperformed Bergamo and required less power, giving it a 30% or better efficiency advantage. Again, however, Bergamo is faster per physical core.
- Xeon—Ampere also compared its chips with the Epyc 9654 (Genoa) but not with Xeon. James was nice not to embarrass her former employer.
Customers—Never Really Wanted to Be Your Friend, Anyway
In the address, Ampere decried build-your-own processors, calling them dangerous (presumably not electrically hazardous but dangerous philosophically or economically) and saying they reflect a misunderstanding of the computing ecosystem. Unfortunately for Ampere, times have changed. Amazon Graviton, Google Axion, Microsoft Cobalt, and Nvidia Grace emerged because those companies are big enough to create their own weather. They have prodigious cash, massive scale, expertise, and big software teams. Simultaneously, Arm Neoverse lowers the barriers to server-processor development and enables compatibility among chipmakers. Ampere, thus, struggles to gain traction with the biggest cloud service providers, leaving only antiestablishment ones like Oracle and Scaleway. Meanwhile, smaller providers and enterprises will favor incumbent Intel to minimize perceived risk.
Bottom Line
Ampere is in a tough position. Falling short of competing x86 processors’ per-core performance, AmpereOne’s efficiency is its key selling point. However, the data-center customers most concerned with opex—power efficiency—can roll their own chip. They’re also the biggest and the most engaged in AI, the big trend Ampere is trying to capitalize on. Developing a CPU instead of licensing one from Arm, Ampere is surely consuming cash.
Meanwhile, the Arm tide is rising. Not only is it gaining in the cloud, it’s showing promise on the desktop. Microsoft just launched new Surface PCs employing the Qualcomm Snapdragon X processor. Ampere should consider a niche beyond the data center: software-development and content-creation workstations. The latter can run the editing and other software ported to Arm for PC use. The former would help those targeting the new Windows machines, Arm-based cloud servers, and possibly also Apple systems. For example, Linux Torvalds has stated he’s building more Arm kernels now that he has a high-throughput Ampere Altra Max system. Between the lucrative data center and desktop markets, an opportunity must exist for a company seeking to capitalize on Arm’s ascendance.