Skip to main content
Newsletter IconGoogle IconFacebook IconX IconInstagram IconYouTube Icon

NVIDIA H100 GPU has nearly 13,000 AI-designed arithmetic circuits

NVIDIA details its new Hopper H100 GPU has nearly 13,000 arithmetic circuits designed by AI, something humans were only capable of.

Facebook IconX IconReddit Icon
Comments
Published
Updated
1 minute & 45 seconds read time

NVIDIA's next-gen Hopper GPU architecture is one of the most monsterous pieces of technology the human race has ever created... but it wasn't just humans... artificial intelligence (AI) helped in a huge uway, too.

In a recent post on NVIDIA's own Developer website, the company underlines how it used AI to design the greatest GPU it has created so far: the new NVIDIA Hopper H100 GPU. NVIDIA normally designs most of its next-gen GPUs using Electronic Design Automation (EDA) tools, but AI helped out with Hopper H100 through the PrefixRL methodology, which is an optimization of Parallel Prefix Circuits using Deep Reinforcement Learning. This allows NVIDIA to design smaller, faster, and more power-efficient GPUs... all while having more performance.

Popular Now: The Legend of Zelda: Twilight Princess 3DS port is getting closer to release as New 3DS performance improves

Rajarshi Roy, Applied Deep Learning Research at NVIDIA, tweeted: "Arithmetic circuits were once the craft of human experts, and are now designed by AI in NVIDIA GPUs. H100 chips have nearly 13,000 AI designed circuits! How is this possible?" and then posted a link to the blog that he co-wrote on NVIDIA's website.

The authors explain: "Arithmetic circuits in computer chips are constructed using a network of logic gates (like NAND, NOR, and XOR) and wires. The desirable circuit should have the following characteristics:"

  • Small: A lower area so that more circuits can fit on a chip.
  • Fast: A lower delay to improve the performance of the chip.
  • Consume less power: A lower power consumption of the chip.

NVIDIA has said that it used this method to design close to 13,000 AI-assisted circuits, which offer a 25% area reduction when put up against EDA tools (which are just as fast, and just as functionally the same as the new method powered by AI).

PrefixRL is said to be a very computational demanding task, where it uses 256 CPUs and 32,000 GPU hours for each and every physical simulation of each GPU. NVIDIA worked on this too, of course, developing "Raptor", which is NVIDIA's new in-house distributed reinforcement learning platform that taps NVIDIA hardware for this very thing.

  • Raptor's ability to switch between NCCL for point-to-point transfer to transfer model parameters directly from the learner GPU to an inference GPU.
  • Redis for asynchronous and smaller messages such as rewards or statistics.
  • A JIT-compiled RPC to handle high volume and low latency requests such as uploading experience data.
NVIDIA H100 GPU has nearly 13,000 AI-designed arithmetic circuits 01
Animated GIF related to NVIDIA H100 GPU has nearly 13,000 AI-designed arithmetic circuits
Photo of the MSI Gaming GeForce RTX 3090 Ti 24GB GDRR6X

Best Deals: MSI Gaming GeForce RTX 3090 Ti 24GB GDRR6X

Prices last scanned 1 hour and 1 minute ago

* Prices may be inaccurate. As an Amazon Associate, we earn from qualifying purchases. We earn affiliate commissions from Newegg and PCCG sales.

Join Our Newsletter

Join the TweakTown Newsletter for daily tech updates delivered to your inbox.

See previous giveaways.

Comments

Stay Updated

Follow TweakTown for breaking tech news, reviews, and daily updates.

Follow TweakTown on GoogleAdd TweakTown as a preferred source on Google
Newsletter Subscription