NVIDIA's Blackwell B200 GPU and GB200
NVIDIA's innovations amaze at this year’s GTC
Tonight, at the GPU Technology Conference (GTC24), Jensen Huang, co-founder, president and CEO of NVIDIA, had a groundbreaking announcement for us all. His keynote speech introduced NVIDIA's latest innovations: the Blackwell B200 GPU and the GB200 superchip, set to change the landscape of computing.
The Blackwell B200 GPU offers up to 20 petaflops of FP4 horsepower from its staggering 208 billion transistors. This immense computational power tackles complex AI workloads with ease, supported by impressive efficiency, with NVIDIA claiming it can reduce cost and energy consumption by up to 25 times compared to the H100.
Huang also presented the second-gen transformer engine embedded within the Blackwell B200 GPU. This engine effectively doubles the compute, bandwidth and model size by utilising four parts for each neuron instead of the traditional eight. This breakthrough not only facilitates the 20 petaflops of FP4 performance but also paves the way for future advancements in AI computing.
NVIDIA's introduction of a next-gen NVLink switch enables seamless communication between a large number of GPUs. This switch allows an impressive 576 GPUs to converse with each other, providing 1.8 TB/s of bidirectional bandwidth. With 50 billion transistors and onboard compute capabilities, this NVLink switch minimises communication overhead, ensuring efficient utilisation of computational resources.
The implications of these advancements are sure to send shockwaves through the world of tech. Tasks that once demanded massive computing power and energy consumption can now be accomplished with unprecedented efficiency. For instance, training a 1.8 trillion parameter model, which previously required thousands of GPUs and megawatts of power, can now be achieved with a fraction of the resources.
NVIDIA's vision extends beyond individual GPUs, offering comprehensive solutions tailored for large-scale deployment. The NVL72 racks, featuring a combination of CPUs and GPUs, offer eyewatering levels of AI training and inference performance. These racks, with their 720 petaflops of AI training performance or 1,440 petaflops of inference, are ideal for cloud service providers like Amazon, Google, Microsoft and Oracle – with NVIDIA announcing that orders with the tech giants have already been placed.
NVIDIA's DGX Superpod for DGX GB200 combines multiple systems into a single powerful unit, capable of delivering 11.5 exaflops of FP4 computing. With the ability to scale to tens of thousands of GB200 superchips, NVIDIA's systems allow organisations to leverage the many benefits of AI at scale.
NVIDIA's Blackwell B200 GPU and GB200 superchip represent a significant milestone in the evolution of AI computing, and Boston greatly looks forward to the transformative impact these innovations will have on industries from healthcare to finance and everything in between, welcoming a new era of intelligent computing solutions. The Blackwell B200 GPU's incredible specifications and the GB200 superchip's potential to revolutionise AI computing leave us excited for the future of technology and its applications in real-world scenarios.
Boston can't wait for the rest of GTC, where we are certain further revelations and insights will unfold, as we explore advancements in AI, computing and the wider world of technology. Follow us on our social channels for real-time updates and deeper insights into how these innovations will shape the future of technology!
If you're at GTC, you can find us at booth 1621, you can also book a meeting with our GTC team on the ground.