Already solidly within the driver’s seat of the generative AI accelerator market presently, NVIDIA has lengthy made it clear that the corporate isn’t about to decelerate and take a look at the view. Instead, NVIDIA intends to proceed iterating alongside its multi-generational product roadmap for GPUs and accelerators, to leverage its early benefit and keep forward of its ever-growing coterie of opponents within the accelerator market. So whereas NVIDIA’s ridiculously widespread H100/H200/GH200 sequence of accelerators are already the most well liked ticket in Silicon Valley, it’s already time to speak concerning the subsequent technology accelerator structure to feed NVIDIA’s AI ambitions: Blackwell.
Amidst the backdrop of the primary in-person GTC in 5 years – NVIDIA hasn’t held one in every of these since Volta was in vouge – NVIDIA CEO Jensen Huang is taking the stage to announce a slate of recent enterprise merchandise and applied sciences that the corporate has been onerous at work on over the previous few years. But none of those bulletins are as eye-catching as NVIDIA’s server chip bulletins, because it’s the Hopper structure GH100 chip and NVIDIA’s deep software program stack working on prime of it which have blown the lid off of the AI accelerator trade, and have made NVIDIA the third Most worthy firm on the planet.
But the one catch to creating a groundbreaking product within the tech trade is that it’s essential do it once more. So all eyes are on Blackwell, the following technology NVIDIA accelerator structure that’s set to launch later in 2024.
Named after Dr. David Harold Blackwell, an American statistics and arithmetic pioneer, who, amongst different issues, wrote the primary Bayesian statistics textbook, the Blackwell structure is as soon as once more NVIDIA doubling down on lots of the firm’s trademark architectural designs, seeking to discover methods to work smarter and work more durable with a purpose to enhance the efficiency of their all-important datacenter/HPC accelerators. NVIDIA has an excellent factor going with Hopper (and Ampere earlier than it), and at a excessive stage, Blackwell goals to deliver extra of the identical, however with extra options, extra flexibility, and extra transistors.
As I wrote again through the Hopper launch, “NVIDIA has developed a very solid playbook for how to tackle the server GPU industry. On the hardware side of matters that essentially boils down to correctly identifying current and future trends as well as customer needs in high performance accelerators, investing in the hardware needed to handle those workloads at great speeds, and then optimizing the heck out of all of it.” And that mentality has not modified for Blackwell. NVIDIA has improved each facet of their chip design from efficiency to reminiscence bandwidth, and every component is focused at enhancing efficiency in a particular workload/situation or eradicating a bottleneck to scalability. And, as soon as once more, NVIDIA is continuous to search out extra methods to much less work altogether.
Ahead of as we speak’s keynote (which by the point you’re studying this, ought to nonetheless be occurring), NVIDIA provided the press a restricted pre-briefing on the Blackwell structure and the primary chip to implement it. I say “limited” as a result of there are a variety of key specs the corporate will not be revealing forward of the keynote, and even the identify of the GPU itself is unclear; NVDIA simply calls it the “Blackwell GPU”. But here’s a rundown of what we all know to this point concerning the coronary heart of the following technology of NVIDIA accelerators.
| NVIDIA Flagship Accelerator Specification Comparison | |||||
| B200 | H100 | A100 (80GB) | |||
| FP32 CUDA Cores | A Whole Lot | 16896 | 6912 | ||
| Tensor Cores | As Many As Possible | 528 | 432 | ||
| Boost Clock | To The Moon | 1.98GHz | 1.41GHz | ||
| Memory Clock | 8Gbps HBM3E | 5.23Gbps HBM3 | 3.2Gbps HBM2e | ||
