![]() ![]() We assumed this was a far-off bit of tech, but GTC suggests it could show up sooner rather than later. Nvidia and IBM have teamed up to improve memory bandwidth and throughput by connecting an SSD directly to the GPU. That’s not relevant for a desktop GPU, but we may see a similar approach with the RTX 4080. This is to reduce latency and improve throughput to the GPU, eliminating CPU bottlenecks in servers. The last hint comes from the H100 CNX, which is a version of the H100 GPU that comes coupled with an Nvidia ConnectX-7 SmartNIC. It’s not clear if the RTX 4080 will use an MCM design, but the launch of Hopper suggests it won’t. Rumors suggest that AMD will, for the first time, leapfrog Nvidia with its RX 7000 graphics cards, and that could be due to the MCM design. Opening NVLink could lay the foundation for Nvidia to do something similar. Rumors suggest that AMD is going with a multichip module (MCM) design for RX 7000 graphics cards, essentially combining multiple separate compute clusters on a single chip. This could be relevant for Nvidia’s upcoming consumer graphics cards. ![]() With NVLink being open, Nvidia says the goal is to have other companies design semi-custom chips that work with Nvidia’s products. This is an interconnect that’s only relevant today in Nvidia’s data center, but Huang announced that it’s coming to customers and partners. ![]() Nvidia focused on scalability with the the fourth generation of NVLink. The manufacturing process is the biggest development, but Hopper holds a couple of other clues, too. It’s possible we move on to a smaller process by then, but it appears the consumer cards will still lag behind the data center ones. Nvidia looks to be leading with Hopper, featuring a more efficient architecture, in order to set up the generation after the RTX 4080. We could be seeing a repeat of the Pascal/Volta/Turing situation. This builds on leakers suggestion that the RTX 4080 will have big problems with efficiency. Based off of the rumors we’ve seen about the massive power requirements for RTX 40-series graphics cards, N5 seems more likely for the consumer range. N5 and N4 live in the same family, but N4 is slightly more efficient than N5. What’s interesting is that Nvidia is rumored to use TSMC’s N5 process for the 4080, not the smaller and more efficient process that Hopper GPUs are using. This is the first time we’ve truly seen two architectures from Nvidia live side by side.įor Hopper, we learned that it will use TSMC’s N4 manufacturing process and that Nvidia is targeting efficiency. All of this is to say, there isn’t a lot of precedent for what Nvidia is doing here. That changed with the RTX 30-series, where Nvidia unified both its product ranges under the Ampere architecture. Nvidia’s most important next-gen GPU is less than 2 weeks away Why Nvidia’s brand new GPU performs worse than integrated graphics This mysterious Nvidia GPU is an absolute monstrosity - and we just got another look ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |