NVIDIA's Blackwell B200 GPUs incorporate a brand new architecture compared to Hopper but also consume almost twice as much power.
NVIDIA Blackwell GPUs Are Rated At Up To 1200W, Various Configurations & All With Brand New Architecture
When NVIDIA's CEO, Jensen Huang, announced Blackwell during the GTC 2024 keynote, the reveal lacked a lot of technical and architectural information. But during the next few days of GTC, NVIDIA shared slightly more details but still without going too much into the technical deep-dives that we are all awaiting. The new details were revealed by Jonah Albe (NVIDIA SVP & GPU Architect) and Ian Buck (NVIDIA VP of Hyperscale & HPC).
Blackwell GPU - Designed For The AI Era With A Brand New Architecture
To start, we all knew that Blackwell was going to be a major architectural upgrade over Hopper & it looks like it's more than that with Jonah stating that Blackwell uses a completely different micro-architecture than Hopper.
What we do know about Blackwell is that it packs the 2nd Generation of Transformer Engine technology which adds FP4 and FP6 compute formats. These formats and new software optimizations are what make Blackwell the fastest AI chip of its kind on the planet but that has taken a toll on its standard FP64 compute which has only increased by 32% versus hopper. The reasoning is plain and simple, Blackwell is an AI chip first and that's its main target market. FP64 is not that important from an AI perspective and the lower you go, the faster the inferencing and training capabilities.
Also, the reason to go the chiplet (MCM) route happens to be the need to improve overall performance rather than improving the yields. It will be interesting to see how NVIDIA's first MCM approach works in the field since we are talking about two GPUs running on the same package. It's mentioned that CUDA does a fairly good job in handling the two GPUs & the different architecture, requiring no major changes to be made for programmers.
GB200 GPU Is The Full Blackwell Specs, 500W More Power Than Hopper
During the launch, there was a particularly big confusion surrounding all the Blackwell GPU and platform variants. Jensen stated that Blackwell isn't a GPU, it's an entire platform & the platform has a range of products but they are still based on GPUs. As of right now, NVIDIA has announced three official Blackwell GPU variants.
These include the flagship and full-spec B200 which is being used by the GB200 Superchip platforms. This chip has the highest-rated computing capabilities and has a maximum TDP of 1200W. This is 500 Watts more than the Hopper H100 which featured a 700W TDP. The entire Superchip is equipped with two of these B200 GPUs and a Grace CPU for up to 2700W power (1200W x 2 for B200 + 300W CPU/IO).
Next up is the Blackwell B200 used by the DGX & HGX platforms which is optimized around 1000W and offers almost 90% of the performance of the full-spec variant. It isn't known if this variant only has a lower TDP or comes with cut-down specs versus the full configuration. Lastly, there is the Blackwell B100 which is a further tuned variant with a 700W TDP. This variant offers around 80% perf of the B200 (1000W) and 70% perf of the B200 (1200W).
There's a possibility of a single-die Blackwell GPU variant, especially for PCIe platforms in the future. The Blackwell GPU architecture is already being incorporated in consumer-tier RTX & AI platforms with the likes of Drive Thor and the future GeForce lineup. NVIDIA's Blackwell GB200 GPUs will start shipping later this year to the first major AI customers followed by volume ramp happening later.
