Semi-OT: Oracle owns almost 1/3 of Ampere + Ampere working on 512 Core HBM AI CPU

Could conceivably take control of Ampere in 2027, from what I read here. And Ampere may be looking into GPUs. What’s interesting about that to me is, it elevates Oracle a bit more into the realm of the Google/MS/AWS cloud providers.

I haven’t been tracking Ampere closely at all. Apparently they have announced a planned 512 core Arm server CPU that will be paired with HBM and presumably meant to be usable for AI without a GPU… New ways to stake out parts of the market. I’m wondering whether Oracle would go all-in on developing their own AI-capable CPUs for the near-hyperscale market, would they reach/have they reached the scale required to make the economics work… lots of questions.

Ampere’s focus is on core-dense designs, but its particular focus on providing chips for air-cooled environments is another key value prop — Ampere says that despite 1mW racks being on the horizon amid a continued industry push for higher-performance silicon at the expense of voracious power consumption, 77% of currently-deployed racks are still under 20kW, and half fall under 10kW. Naturally, air cooling is much more forgiving on the pricing front than liquid cooling, so Ampere has focused heavily on power efficiency, naturally reducing the cooling burden.

The 512-core AmpereOne Aurora is currently in development and leverages the company’s IP, which it developed with its AmpereOne chips. The company isn’t sharing many details on this chip yet, but key highlights include Ampere’s custom Arm-powered AmpereOne cores, scalable AmpereOne mesh, and a die-to-die interconnect that facilitates communication across chiplets, with the latter comprised of a custom SERDES and protocol.

Perhaps most importantly in the AI-obsessed data center, the company also says its AmpereAI IP offers dedicated AI acceleration, and its addition of support for high bandwidth memory (HBM) will be instrumental in feeding enough memory bandwidth to the new engine. However, the chip is also geared for standard general-purpose workloads, too.

2 Likes