Somehow I really doubt that GB203 will be smaller than AD103 on the same process.
Well @xpea has implied it will be since they've been able to increase transistor density with AI based innovations. I mean 30% with B100 so I guess GB203 is probably less but still substantial (maybe 10-20%?), So could be 350mm2 or there abouts? Basically smaller than AD103 (376mm2).
Higher density on the same process could be just a result of using different process libraries and not necessarily related to any innovations. Blackwell is a new architecture so whatever they will save this way will likely be eaten up by additional features. This transition will probably be similar to Pascal->Turing, maybe not as extreme but comparable. Thus chips being smaller is something I find unlikely.
Nvidia isn't going to offer better value if consumers like yourself are just going to give in anyways. Raising prices by a foot, then conceding an inch or so is not at all a 'correction', either. It's just giving us crumbs as they gorge themselves at the banquet!
Proportionally, GB203 is supposed to be smaller compared to the top die than with Lovelace, but we dont know top die size. Either way, if it's really just 96SM's, that's not much more than AD103. I reckon it'll be pretty similar in size to AD103. Smaller isn't impossible.
wccftech.com
In a new blog post, NVIDIA has shared how Blackwell GPUs are going to add more performance to the research segment which includes Quantum Computing, Drug Discovery, Fusion Energy, Physics-based simulations, scientific computing, & more. When the architecture was originally announced at GTC 2024, the company showcased some big numbers but we have yet to get a proper look at the architecture itself. While we wait for that, the company has more figures for us to consume.
...
NVIDIA also sheds light on the double-precision of FP64 (Floating Point) capabilities of its Blackwell GPUs which are rated at 30% more TFLOPs than Hopper. A single Hopper H100 GPU offers around 34 TFLOPs of FP64 compute and a single Blackwell B100 GPU offers around 45 TFLOPs of compute performance. Blackwell mostly comes in the GB200 Superchip which includes two GPUs along with the Grace CPU so that's around 90 TFLOPs of FP64 compute capabilities. A single chip is behind the AMD MI300X and MI300A Instinct accelerators which offer 81.7 & 61.3 TFLOPs of FP64 capabilities on a single chip.
...
NVIDIA quickly shifts gears and brings us AI performance once again where its Blackwell GB200 GPU platform once again reigns supreme with a 30x gain over H100 in GPT (1.8 Trillion Parameter). The GB200 NVL72 platform enables up to 30x higher throughput while achieving 25x higher energy efficiency and 25x lower TCO (Total Cost of Operation). Even putting the GB200 NVL72 system against 72 x86 CPUs yields an 18x gain for the Blackwell system and a 3.27x gain over the GH200 NVL72 system in Database Join Query.
But NVIDIA isn't stopping any time soon as the company is anticipated to start production of its next-gen Rubin R100 GPUs by as early as late 2025 and the initial details sound insane.
Jim said this ironically. I doubt it that CowoS L cost as much as 10 billions.
wccftech.com
never miss an opportunity for a good price increaseNVIDIA Blackwell GPUs Estimated To Cost Up To $35,000, AI Servers Up To $3 Million As The Firm Gears Up For The Next "Gold Rush"
NVIDIA Blackwell GPUs are going to be the fastest AI chips the world has ever seen when they become available & also the most expensive yet.
Except that's less than a GH200.
Sure. That's why all hyperscales are developing their own silicon.
