Welcome to the next generation of GPUs. One day of his GTC 2022 keynote on Tuesday morning, Nvidia CEO Jensen Huang formally revealed “Hopper,” the graphics architecture powering the company’s subsequent generation of products, and welp, it looks esteem an absolute monster—no longer lower than in data-heart earn.
Hopper’s debut takes shape in the facts heart-centered Nvidia H100, a beastly chip wielding a whopping 80 million transistors, essentially the most significant use of subsequent-gen HBM3 memory, and advances in the company’s blazing-instant NVLink interconnect expertise. Hopper is “our ideally suited generational jump ever,” Huang proudly declared.
Introducing new architectures at GTC is par for the direction for Nvidia; Ampere, Volta, and Pascal moreover stepped into the highlight there. And lawful esteem with these previous unearths, whereas the H100 is closely in actuality good for data-heart uses, with a tensor core-laden hardware configuration designed to optimize deep-discovering out duties, that doesn’t suggest we humble PC avid gamers can’t infer some knowledge from Hopper’s AI-centric point to, although the RTX 4080 and its ilk stop up working on a particular “Lovelace” architecture esteem some most trendy rumors suggest.
Here are three key things that Nvidia’s Hopper architecture diagram for the next-gen GeForce lineup.
1. Hopper jumps to TSMC 4nm
Nvidia’s RTX 20-series “Turing” GPUs were constructed at TSMC, the fab that moreover builds chips for AMD, Apple, Intel, and virtually all bleeding-edge expertise firms. However the Ampere GPUs right by the GeForce RTX 30-series switched to Samsung’s 8nm direction of as a change. They delivered ferocious performance, however moreover drew ferocious quantities of vitality.
For Hopper—or no longer lower than the H100—Nvidia returns to TSMC. This chip is constructed the use of a custom model of TSMC’s 4N direction of, which is a more vitality-pleasant model of TSMC’s N5 direction of. AMD’s RDNA 2-primarily based Radeon RX 6000-series graphics playing cards swiped the vitality effectivity crown from Nvidia, and the next generation for these competitors gaze esteem they’ll all be crafted by TSMC. That’s no guarantee—the Ampere-primarily based A100 became once moreover constructed at TSMC, unlike its RTX 30-series siblings—however essentially the most trendy spate of rumors all display cloak it. Game on.
Nvidia
2. PCIe Gen 5 (and NVLink 4)
AMD’s Radeon RX 5000-series playing cards were first to the PCIe 4.0 punch, however Nvidia raced previous it for the PCIe 5.0 pole situation. The H100 is basically the most significant graphics processor constructed with the even-faster interconnect expertise, and since Intel’s 12th-gen Core processors no longer too long ago brought PCIe 5.0 attend to mainstream desktops, there’s every motive to factor in that Nvidia will embody it when subsequent-gen GeForce GPUs appear.
Hopper moreover marks the debut of the fourth generation of NVLink, Nvidia’s proprietary excessive-bandwidth interconnect. The RTX 30-series limits multi-GPU configurations to the flagship GeForce RTX 3090 alone. Assuming Nvidia continues to present multi-GPU attend in GeForce flagship playing cards, you’ll perceive a broad flee boost with NVLink 4. The H100’s NVLink 4 implementation offers 900GB/s of bandwidth, Nvidia says, in contrast to the 600GB/s that the Ampere-primarily based A100 and its third-gen NVLink offers.
The H100 is amazingly optimized for data-heart duties, and Nvidia (oddly) didn’t whine any knowledge about CUDA core counts, clock speeds, or GPU die sizes—all of which would present vital deeper clues into what to await from futuristic GeForce playing cards. However the specs the company did point to display cloak that Nvidia isn’t timid to head broad with Hopper.
The flagship Hopper GPU right by the flagship H100 bristles with an astonishing 80 billion transistors—a enormous jump previous the 54.2 billion transistors in the Ampere-primarily based A100, and almost four instances as many transistors because the 21.1 billion lurking right by the Volta-primarily based V100.
Nvidia
Pushing all these transistors to peak performance requires a lot of vitality even though. Nvidia says the H100 draws as much as 700 watts, a staggering 300 watts more than the quickest A100 chip. Holy moly. That doesn’t essentially suggest that RTX 4000-series GeForce GPUs would be gasoline guzzlers too, however most trendy rumors suggest that Nvidia’s subsequent-gen graphics playing cards will relish a form of vitality. Seeing the H100 scheme so vital wattage lends credence to that claim, even though we’ll want to now stay awake for a formal point to of subsequent-gen GeForce graphics card to clutch for sure.
What we aloof want to find
Slightly vital every little thing. The GTC unearths of previous Nvidia architectures supplied vital elevated factor about core depend and clock flee, which gave vital deeper perception into what to await from their person-centered GeForce cousins. Rumors of the RTX 4080 and its siblings working on a particular “Lovelace” architecture moreover muddy the waters. The H100 is slated to hit the streets in the third quarter, and we’d demand to listen to more about subsequent-gen GeForce GPUs sooner or later round then besides. Set tuned!
Show: While you bought something after clicking links in our articles, we might perhaps perhaps just compose a puny price. Be taught our affiliate hyperlink coverage for more facts.
