Nvidia presented the next generation of data center GPUs at the GTC with “Hopper”. Nvidia relies on TSMC’s 4N process for production. PCIe 5.0 and NVLink 4 are also on board. We took a look at what this could mean for the RTX 4000 GPUs.
During the GTC 2022 keynote (see video) on March 22, Nvidia CEO Jensen Huang officially unveiled the “Hopper” graphics architecture, which will power the company’s next generation of GPUs – at least the models for data centers. Hopper’s debut comes with the data center-focused Nvidia H100, a massive 80 million transistor chip paired with next-generation HBM3 memory and the company’s advanced NVLink interconnect technology. Hopper is “our biggest generational leap to date,” Huang proudly explains.
The presentation of new Nvidia architectures at the GTC has a long tradition: Ampere, Volta and Pascal were also in the limelight there.
Nvidia GeForce RTX 4000: Everything about performance, price and release
Just as with previous reveals, while the H100 is heavily specialized for data center use – due to the tensor-cores-laden hardware configuration designed to optimize deep learning tasks – that doesn’t mean that for PC gamers can’t glean some information from Hopper’s AI-centric vision. Even if a possible RTX 4080 and its ilk should be based on the “Lovelace” architecture, as some current rumors suggest.
These are the most important points, which is why “Hopper” is also interesting for PC gamers:
1. Hopper relies on TSMC’s 4N process
Nvidia’s Turing RTX 20-series GPUs were built at TSMC, the contract manufacturer that also makes chips for AMD, Apple, Intel, and virtually every leading technology company. The Ampere GPUs of the GeForce RTX 30 series are instead manufactured in Samsung’s 8nm process. The GPUs deliver impressive performance, but also consume a lot of power, as our test of the RTX 3090 shows.
For Hoppers – or at least the H100 – Nvidia returns to TSMC. TSMC manufactures the chip in a 4N process specially optimized for Nvidia, which is ultimately a more energy-efficient version of the N5P manufacturing process. A look at our graphics card comparison shows that AMD’s current RX 6000 graphics cards, which TSMC manufactures using the 7 nm process, are more energy efficient than Nvidia’s RTX 3000 GPUs. It would therefore not be unlikely that Nvidia would want to use the same manufacturing process for the RTX 4000 models as with Hoppers, the current rumors definitely point to that.
2. PCIe Gen 5 (and NVLink 4)
AMD’s Radeon RX 5000 series graphics cards were the first models to rely on PCIe 4.0, but from the looks of it, Nvidia should have the edge when it comes to PCIe 5.0. The H100 is the first GPU to feature the even faster connectivity technology, and with Intel’s 12th Gen Core processors recently bringing PCIe 5.0 support to mainstream desktops, there’s every reason to believe that Nvidia will also bring the Next-gen GeForce GPUs will rely on PCIe Gen 5.
How much the higher bandwidth really brings in the end has yet to be seen in the tests. The differences between a PCIe 3.0 x16 and a PCIe 4.0 x16 interface are extremely small, even with a current top model.
Hopper also marks the debut of the fourth generation of NVLink, Nvidia’s proprietary high-bandwidth interconnect. For the RTX 30 series, multi-GPU configurations are strictly limited to the flagship GeForce RTX 3090. If Nvidia wants to continue offering multi-GPU support in the GeForce flagship cards, NVLink 4 will bring them from a wider range can benefit. According to Nvidia, the NVLink 4 implementation of the H100 achieves a transfer rate of 900 GB/s, compared to 600 GB/s for the Ampere-based A100 with third-generation NVLink.
3. A lot of performance, but also high power consumption
The H100 is highly optimized for data center tasks, and Nvidia hasn’t (curiously) disclosed any information on CUDA core counts, clock speeds, or GPU die sizes – all things that would give a lot more insight into what to be expected from future GeForce cards. But the specs the company has revealed show that Nvidia isn’t afraid to make it big with Hopper.
The hopper GPU in the flagship H100 has a whopping 80 billion transistors – a whopping leap from the 54.2 billion transistors in the Ampere-based A100 and nearly four times the number of transistors compared to the 21.1 billion in the Volta-based V100.
However, in order to be able to drive all these transistors to maximum performance, a lot of power is required. Nvidia states that the H100 requires up to 700 watts, which is 300 watts more than the fastest A100 chip. That doesn’t necessarily mean that GeForce RTX 4000-series GPUs will also be power hogs, but recent rumors suggest that Nvidia’s next-gen graphics cards will be heavy users of power.
The leaks currently speak of 600 watts and more for the top model. The fact that the H100 draws so many watts backs up those claims, although we’ll have to wait for the official reveal of next-gen GeForce graphics cards to know for sure.
What else would we have wished for?
Significantly more information. The GTC reveals of past Nvidia architectures provided much more detail on core counts and clock speeds, giving a much deeper look into what to expect from the consumer-oriented GeForce offshoots. There are many rumors about the RTX 4080 and its sibling models, which are said to be based on the “Lovelace” architecture, but they sometimes contradict each other. The H100 is slated to launch in Q3, and we expect to hear more about next-gen GeForce GPUs by then as well.
This article first appeared on our English-speaking PCWorld colleagues under the title “Nvidia’s next-gen Hopper GPU revealed: 3 things PC gamers need to know” .