Source:
The Register Multi-generation chip collab is more buzzwords than compute
Amazon Web Services on Tuesday tapped mobile chip titan and aspiring datacenter SoC maker Qualcomm to develop "customized" silicon for AI
inferenceand 1.6 Tbps optical networking gear to tie it all together, though it offered few details. Amazon's purchases and other commercial commitments under the agreement could total up to $60 billion in payments through September 2036. “Qualcomm is pleased to work with AWS on customized silicon and connectivity solutions, bringing decades of leadership in advanced processing and power-efficientcompute, to deliver breakthrough performance and enable the next generation of AI infrastructure,” Qualcomm CEO Cristiano Amon said in a canned statement that tells us absolutely nothing about the nature of the deal. While details remain vague, custom chip deals of this sort aren’t unusual for cloud providers like Amazon. The cloud giant has previously engaged Intel to develop customized versions of its Xeon processors. Similarly, AMD has built custom CPUs, like the Epyc 9V64H, for Microsoft. Qualcomm’s involvement could involve the development of fully custom silicon, chiplets, or customized versions of its emerging datacenter accelerator portfolio. Detailed during Qualcomm's annual investor day earlier this summer, the new rack-scale AI platform swaps pricey high-bandwidth memory (HBM) for what the chip biz calls high-bandwidth compute. Qualcomm boasts that the technology will enable it to achieve higher “effective” memory bandwidth — a key bottleneck for AI inference workloads — using cheaper LPDDR memory by shifting some of the compute to the memory’s base die. The technology is expected to launch next year as part of Qualcomm’s AI250-series of Dragonfly rack systems. It’s entirely possible that, just like Intel builds customized versions of its Xeon 6 processors for AWS, Qualcomm will bring a customized version of its AI250-series chips to the cloud provider as well. Qualcomm is also working on a new datacenter CPU called the C1000 that it claims will touch 5 GHz and scale to more than 250 cores when it arrives in the second half of 2028. We reached out to Amazon and Qualcomm for additional color on what the multi-generation tie will eventually entail, but hadn’t heard back at the time of publication. While the AI inference side of the equation is about as clear as mud, Qualcomm’s networking contributions are far more obvious. As part of the collab, Amazon will employ Qualcomm’s serializer-deserializers (SerDes) and optical digital signal processors (DSPs) — likely acquired with the purchase of Alphawave Semi late last year — to boost port speeds to 1.6 Tbps. This points to AWS expanding its supply chain to include Qualcomm as a source of optical pluggables, or potentially near- or co-packaged optics. High-speed optical interconnects have become a major supply chain hurdle for scaling AI infrastructure, particularly as compute scales from a rack with a few dozen accelerators to hundreds or thousands spanning multiple racks. The customer-supplier relationship also runs both ways. While AWS hoovers up its chip tech, Qually plans to expand its use of Amazon’s Bedrock platform for electronic design automation workloads — after all, this wouldn't be an AI deal if it didn't contribute to the circular economy somehow. ® Get AI news in your inbox
Daily digest of what matters in AI.