Amazon Updates Homegrown Chips, Even as It Grows Nvidia Ties (bloomberg.com) 3
Amazon's cloud-computing unit announced updated versions of its in-house computer chips while also forging closer ties with Nvidia -- dual efforts designed to ensure it can get enough supplies of crucial data-center processors. From a report: New homegrown Graviton4 chips will have as much as 30% better performance than their predecessors, Amazon Web Services said at its annual re:Invent conference in Las Vegas. Computers using the processors will start coming online in the coming months.
The company also unveiled Trainium2, an updated version of a processor designed for artificial intelligence systems. It will begin powering new services starting next year, Amazon said. That chip provides an alternative to so-called AI accelerators sold by Nvidia -- processors that have been vital to the build-out of artificial intelligence services. But Amazon also touted "an expansion of its partnership" with Nvidia, whose chief executive officer, Jensen Huang, joined AWS counterpart Adam Selipsky on stage. AWS will be the first big user of an updated version of that company's Grace Hopper Superchip, and it will be one of the data-center companies hosting Nvidia's DGX Cloud service.
The company also unveiled Trainium2, an updated version of a processor designed for artificial intelligence systems. It will begin powering new services starting next year, Amazon said. That chip provides an alternative to so-called AI accelerators sold by Nvidia -- processors that have been vital to the build-out of artificial intelligence services. But Amazon also touted "an expansion of its partnership" with Nvidia, whose chief executive officer, Jensen Huang, joined AWS counterpart Adam Selipsky on stage. AWS will be the first big user of an updated version of that company's Grace Hopper Superchip, and it will be one of the data-center companies hosting Nvidia's DGX Cloud service.
Why are these chips different? (Score:2)
Amazon, Microsoft, Google, Meta, Apple, Tesla, etc. all have developed their own AI chips. AMD and Intel have GPUs that can be used for AI. And there are many other smaller companies that also sell their own accelerators. That's been the case for many years. Not much has changed. All those chips have been available for many years, and yet Nvidia is dominant (at least for training).
Simple existence of these chips isn't that interesting. What are the market share and revenue numbers? Or at least, what
Re: (Score:2)
These are highly specialized chips. They are more like the proliferation of special purpose graphics accelerators with their own APIs back in Windows 95 days (does anyone remember minigl or glide?)
Why it is important?
The hardware is optimized for the models they run. For example, large "embedding" models require large lookup tables, hence local RAM speed is important. Others might require faster node-to-node communication. Google for example, has their own hardware optical routers for the TPU chips (real op
Re: (Score:2)