Nvidia has announced a new GPU for data centers that may hint at what the company has in store for next-generation PC graphics cards.
The GPU is called the H100 and it’s built using the latest 4-nanometer manufacturing process from Taiwan’s TSMC, enabling Nvidia to pack more transistors on the silicon. In contrast, the current RTX 3000 series uses an 8-nanometer process from Samsung.
The other notable feature to the H100 is how it supports PCIe Gen 5 for faster data transfer speeds. Nvidia also developed a new “Hopper” architecture for the H100, which is designed to accelerate AI-base workloads, such as rendering 3D simulations or training machine learning algorithms.
The end product promises to outperform the company’s last enterprise GPU from two years ago, the A100, by sometimes three or six times, depending on the task. The H100 also has 80 billion transistors compared to the only 54.2 billion transistors found in the A100, which was built using TSMC’s 7-nanometer manufacturing process.
The company introduced the H100 and the Hopper architecture at Nvidia’s annual GTC event, a conference for AI developers. The big question is whether the features in the H100 will come to next-generation RTX GPUs for consumers.
At GTC, Nvidia said the Hopper architecture will succeed Nvidia’s Ampere architecture, which currently powers the RTX 3000 graphics card series. However, rumors suggest Hopper will only target data centers and enterprise systems. To address the consumer market, leaks have indicated Nvidia is working on a separate architecture, codenamed Ada Lovelace, which will power next-generation desktop graphics cards and laptop GPUs.
During a press briefing, we asked Nvidia if the Hopper architecture will end up in consumer products. But the company declined to comment, saying it was only prepared to discuss what it was announcing at GTC.
Nvidia is expected to talk about next-generation RTX graphics cards in September. In the meantime, the upcoming H100 GPU will use expensive HBM3 memory, instead of GDDR6X or the rumored GDDR7 video RAM. The product is set to start arriving in Q3. Customers will also be able to buy eight H100 GPUs packed together into a single server unit, called a DGX H100.
At GTC, Nvidia also unveiled a new "super chip" for servers that combines two of the company's upcoming Grace CPUs on to a single system. The resulting chip has 144 ARM-based cores across two CPUs, which can communicate a high-bandwidth "NVLink C2C" interlink.
Nvidia plans on releasing the super chip next year when the first Grace CPUs launch. The CPUs can also be integrated to work with the H100 GPUs.