NVIDIA’s Blackwell B200 GPUs incorporate a model new structure in comparison with Hopper but additionally devour nearly twice as a lot energy.
NVIDIA Blackwell GPUs Are Rated At Up To 1200W, Varied Configurations & All With Model New Structure
When NVIDIA’s CEO, Jensen Huang, introduced Blackwell through the GTC 2024 keynote, the reveal lacked lots of technical and architectural info. However through the subsequent few days of GTC, NVIDIA shared barely extra particulars however nonetheless with out going an excessive amount of into the technical deep-dives that we’re all awaiting. The brand new particulars have been revealed by Jonah Albe (NVIDIA SVP & GPU Architect) and Ian Buck (NVIDIA VP of Hyperscale & HPC).
Blackwell GPU – Designed For The AI Period With A Model New Structure
To start out, all of us knew that Blackwell was going to be a significant architectural improve over Hopper & it appears prefer it’s greater than that with Jonah stating that Blackwell makes use of a totally totally different micro-architecture than Hopper.
Picture Supply: NVIDIA
What we do find out about Blackwell is that it packs the 2nd Technology of Transformer Engine expertise which provides FP4 and FP6 compute codecs. These codecs and new software program optimizations are what make Blackwell the quickest AI chip of its type on the planet however that has taken a toll on its normal FP64 compute which has solely elevated by 32% versus hopper. The reasoning is obvious and easy, Blackwell is an AI chip first and that is its principal goal market. FP64 is just not that essential from an AI perspective and the decrease you go, the sooner the inferencing and coaching capabilities.
Additionally, the explanation to go the chiplet (MCM) route occurs to be the necessity to enhance total efficiency slightly than enhancing the yields. Will probably be attention-grabbing to see how NVIDIA’s first MCM method works within the discipline since we’re speaking about two GPUs working on the identical bundle. It is talked about that CUDA does a reasonably good job in dealing with the 2 GPUs & the totally different structure, requiring no main adjustments to be made for programmers.
GB200 GPU Is The Full Blackwell Specs, 500W Extra Energy Than Hopper
Throughout the launch, there was a very huge confusion surrounding all of the Blackwell GPU and platform variants. Jensen acknowledged that Blackwell is not a GPU, it is a whole platform & the platform has a variety of merchandise however they’re nonetheless based mostly on GPUs. As of proper now, NVIDIA has introduced three official Blackwell GPU variants.
These embrace the flagship and full-spec B200 which is being utilized by the GB200 Superchip platforms. This chip has the highest-rated computing capabilities and has a most TDP of 1200W. That is 500 Watts greater than the Hopper H100 which featured a 700W TDP. All the Superchip is supplied with two of those B200 GPUs and a Grace CPU for as much as 2700W energy (1200W x 2 for B200 + 300W CPU/IO).
Picture Supply: NVIDIA
Subsequent up is the Blackwell B200 utilized by the DGX & HGX platforms which is optimized round 1000W and gives nearly 90% of the efficiency of the full-spec variant. It is not identified if this variant solely has a decrease TDP or comes with cut-down specs versus the complete configuration. Lastly, there’s the Blackwell B100 which is an extra tuned variant with a 700W TDP. This variant gives round 80% perf of the B200 (1000W) and 70% perf of the B200 (1200W).
There is a chance of a single-die Blackwell GPU variant, particularly for PCIe platforms sooner or later. The Blackwell GPU structure is already being integrated in consumer-tier RTX & AI platforms with the likes of Drive Thor and the long run GeForce lineup. NVIDIA’s Blackwell GB200 GPUs will begin delivery later this yr to the primary main AI prospects adopted by quantity ramp occurring later.