There’s been chatter over the weekend on just how big and powerful Nvidia’s next-generation HPC-optimised Hopper architecture will be.
As a refresher, Hopper is the codename ascribed to the underlying technology, and will likely interleave with consumer-facing Lovelace for upcoming cards in various segments.
Current-generation Ampere’s best chip, known as GA100, tops out at around 826mm², though no production GPU uses its firepower in full. Built on a 7nm process from TSMC, as opposed to 8nm Samsung for consumer GeForce GA102, this beastie is best exemplified by the A100 datacentre processor.
Top-bin Hopper, meanwhile, will be known as GH100, built on TSMC’s 5nm process, and be closer to 1,000mm², if conjecture is to be believed.
We feel it is unlikely to be quite this large as the reticle limit – the largest you can make a single die with traditional tools – is reckoned to be 858mm². It is conceivable that Nvidia will not use industry-standard i193 and EUV lithography steppers, but not doing so would incur significant extra expense.
Nevertheless, GH100 promises to be wonderfully complex and transistor-heavy; today’s GA100 ships with 54.2bn transistors so GH100 ought to be in the region of 75bn.
Fanning the flames of performance, there may be a derivative of the Hopper architecture employing two dies in a multi-chip module arrangement. Known by a different codename to GH100 (GH100X, perhaps?) 3Dcenter.org reckons, if true, it will pull around 1,000W alone.
Though we’re 95 per cent sure it won’t happen, wouldn’t it be something if Nvidia released a GeForce MCM based on the next Hopper/Lovelace architecture.
