The startup announced Tencent is a customer for its data center processors, provided some performance details for its next IC, and offered a glimpse of the chip after that.
Ampere Computing is on a roll. The company’s 80-core Altra processor (released in 2020) is racking up customers, several of which Ampere has finally made public. Its 128-core Altra Max is sampling now, and it’s just about as fast as Ampere suggested it would be. In an interview with EE Times, CEO Renee James called it “just a screamer.” The Altra Max’s successor (still unnamed) is on schedule to sample in 2022.
Ampere was founded in 2018 by James, a former Intel executive, with the ambition of attacking Intel’s strength: processors for data centers. In the short time since Ampere was established, a resurgent AMD has begun to gain traction in that same market, some of the biggest data center operators began designing their own chips, and erstwhile Ampere partner Nvidia made clear its intention to also contend for the data center. Not to mention the scores of other startups with artificial intelligence (AI) accelerators designed for data centers.
With competition that formidable, James and Chief Product Officer Jeff Wittich were pleased to reel off a list of customers that were finally willing to go on record as such. Tencent was the biggest new fish. Others include ByteDance, Equinix, CloudFlare and uCloud. (Microsoft and Oracle we already knew about.) Among OEMs and ODMs, Ampere had previously announced working with GigaByte. Ampere just revealed it is also working with Foxconn and the Inspur Group.
All are customers for the 80-core Altra. There’s no telling what any customer might choose to buy in the future, but Ampere’s explicit value proposition is delivering processors optimized for data center workloads, with each successive product generation delivering performance improvements that scale almost linearly with the number of processor cores, at a cadence of one new product generation per year. The strategy is built to give customers a clear roadmap for improving data center performance annually. Given that value proposition, if a customer is using Ampere products today, it’s easy to infer they will keep doing so.
For decades, the most successful and enduring IC companies have conformed to the schedule set in Moore’s Law: improvements every two years or so. Those improvements have not always defined a straight line on a graph, either. Ampere, on the other hand, has been promising bigger performance improvements, scaling predictably, at a more rapid pace. The company’s most recent announcements combine to demonstrate that.
Ampere also released data demonstrating the performance of its processors scales almost linearly with the number of processor cores, while the performance of competing products from AMD and Intel tails off — see “The Ampere Architectural Advantage” (below). The stats for the 80-core Altra have been published before; this is the first time Ampere has provided any stats for the 128-core Altra Max.
Wittich said among the key reasons performance of x86-based processors doesn’t scale with the number of cores is reliance within the x86 architecture on hyperthreading. Hyperthreading requires sharing resources; if you want to add more cores “you have to throttle back on the frequency.”
Ampere’s 80-core Altra and 128-core Altra Max are both based on Arm designs (they are not the same cores, however), and both are produced at the 7nm node. The unnamed next-gen product is also Arm-based, but built by Ampere itself; it will be implemented at 5nm.
According to Wittich, doing its own designs allows Ampere to adjust the core and feature set on a yearly cadence. “Arm just doesn’t go that fast. We knew our customers would want features and performance specifically for cloud on a faster cadence,” he said.
Part of Ampere’s story has always been about power efficiency, but it seems the company is highlighting it more than before. If so, it’s because power consumption has become doubly important for its data center customers.
Controlling power consumption has always been critical from an operational standpoint, but most operators have now made public commitments to reduce power consumption, usually in the form of a promise to become carbon neutral. “When those goals went public, they became responsible for them. Now they have compliance reports to file,” Wittich said.
James added, “Equinix has customers coming in and asking about their carbon footprint. They’re now being made accountable for this.”
The upshot of all of it – the performance improvements, the improved power efficiency, the yearly technology updates — is that data center customers will be able to continuously and predictably increase performance simply by adding cores to the same racks.
“You don’t build a semiconductor company to do worse than you did before,” James noted.
This article was originally published on EE Times.
Brian Santo is Editor-in-Chief of EE Times. He has been writing about technology for over 30 years, for a number of publications including Electronic News, IEEE Spectrum, and CED; this is his second stint with EE Times (the first was 1989-1997). A former holder of a Radio Telephone Third Class Operator license, he once worked as an engineer at WWWG-AM. He is based in Portland, OR.