12.5.1 : Keynote

• Jensen Huang : NVIDIA
• Slides
• RTX Blackwell 5090 is 30
• 1 pixel -> 15 generated and has to stay temporary stable
• 5090 30 percent smaller than 4090
• Agentic AI -> Reasoning
• The only way to have more people at GTC is to grow San Jose
• chain of thought, Best of end, consitency checking
• The amount of computation needed for inference is 100x compare to what we expect
• numpy -> cupy
• cuLitho : ASML, TSMC
• cuOPT : mathematical optimisation (for flight, worker, drivers, plants, writers, etc)
• cuOPT : will be open source soon
• Parabrick : gene sequencing
• EARTH 2 : weather simulation
• cuQuantum : quantum computing simulation
• cuEquivariance, and cuTensor
• cuDSS, AMGX, cuFFT, cuSparse
• cuDF, cuML (acceleration for Pandas and Spark)
• Warp for physic
• GM choose NVidia to build their autonomous vehicles fleat
• Automotive safety : NVIDIA Halos, 7 millions lines of code safety assist
• Two Blackwell GPU into on Blackwell package : HGX
• Swiches in the middle of the Rack
• Liquid cooled => compressed all the Computing into one single rack
• 60000 components to 600000 components => 1 EFlops computer in one Rack
• One Rack has as many part as 20 cars
• Inference at scale : x axis : token per second, y axis : Throughput in token
• 400 tokens for a classic model, 8000 for a Reasoning model. 20x more token, 150x more computing
• Pipeline Parallelism, tensor Parallelism, expert Parallelism
• Prefill : read pdf or web site, watch a video to get information and learn
• You ingest the KV cache and you produce one token, and redo the same for the next token
• NVIDIA DYNAMO : operating system of an AI factory : find which data and computation to do on which GPU on which rack
• NVIDIA DYNAMO : open source -with perplexity as partner
• Token per second per megawatt
• NVLink 8 and FP8, and than FP4 for quantization
• 25x in one generation at the same consumption power
• In this context Blackwell is 40x the potential of Hopper (for reasoning model)
• 25x at iso power in one generation (between Hopper and Blackwell)
• Full Production of Blackwell
• Blackwell ultra : upgrade second half of this year
• Vera Rubin : Vera 2x Grace (the chassi if the same), NVLink144 => connected to 144 GPUs
• Second half of the year, Rubin Ultra
• Spectrum-X : SuperNIC and supercharge Ethernet
• Copper for local communitation and photonic for large scale stadium data center communitation
• Mark Zander techonlogy (Transivers + Laser)
• Transivers from the GPU, to the switch, to the next switch
• Standard 30 Watts and 1000$ if you buy it in high volume, 6 per GPU => 180 W/GPU, + 6000$ per GPU => 30 W per transivers
• Every GPU would have 6 transivers slots : 100 watt just for communitation
• 1M GPU => 6M Transivers of 30 Watts => 180 MW just for Transivers
• 6 MW is 10 Rubin Ultra Racks
• World First MIR Micro Mirror and micro Mirror (TSMC : cupe)
• 3.5x less consumption : from 30 W to 8.57 W per Transiver
• 100% of NVidia developer will be AI assisted by the end of this year
• Grace Blackwell personnal computer : 20 PFlops, liquid cooled
• For the very first time your storage system will be GPU accelerated
• DGX SParks (Gigits) : 20 CPU, 128 GB memory, 1PFlop
• DGX Station : Grace Blackwell, workstation : 72 CPU Cores, 20 PFlops
• Storage accelerated GPU
• NVIDIA Omniverse with Cosmos
• Verifiable physics rewards
• NVIDIA + Disney Research + Google DeepMind
• Groot N1 is open-source