Rumors about Nvidia RTX 4000 and AMD RX 7000 are circulating on the web
While the current generation of graphics cards is still the main topic of conversation, that hasn't stopped speculation about what's to come. In fact, the next generation architectures from AMD and Nvidia already seem to be much more interesting than the preparations for RDNA 2 and Ampere. This, of course, is due to the return to competition from AMD, although we should not forget about Intel exit plans to the discrete graphics market. Anyway, let's get back to the Nvidia RTX 4000 and AMD Radeon RX 7000 Series rumors.
Several well-known insiders with a relatively accurate reputation have released new information about what will define the next generation of GPUs. Starting with Nvidia, we have confirmation from both Kopite7kimi, so and from greymon55, that the green team will use TSMC's 5nm manufacturing process to build more powerful GPUs with features like more CUDA cores, stream processors, Tensor cores, and beloved RT cores. Given that this would be a huge upgrade over Samsung's 8nm process, we should expect a huge generational upgrade with the Nvidia RTX 4000 Series. Further rumors from TechRadar it is speculated that Nvidia's next flagship model might even boast 18 CUDA cores, while the current RTX 432 has 3090 cores.
In addition, Greymon55 believes that Nvidia has decided to use the Lovelace architecture for the next lineup. We've known about both Lovelace and another architecture called Hopper for some time now, but it wasn't clear what exact purposes they could be used for. Nvidia's Lovelace is currently believed to lighten the RTX 4000 series, while Hopper is on hold for the RTX 5000 series. Hopper is probably far from done, but it's a multi-billion dollar company and they like to plan ahead.
Several years of epic competition
Moving on to AMD, most of the discussion ( via WCCFTech ) was related to the RDNA 3 architecture, specifically how multi-chip design can be used to spread workloads across multiple GPU dies. It's similar to how AMD builds Ryzen processors with so many cores, but for GPUs, the engineering is more complex. Latency has been one of the major problems in multi-chip designs. AMD's revolutionary Infinity Cache technology may finally be the answer to this problem.
Moore's Law is Dead , another well-known insider, also expects RDNA 3 to provide at least a 40% increase in performance per watt over RDNA 2, with 40% being a conservative estimate. The leaker claims that AMD still doesn't know if it can get a multi-chip GPU to work already in RDNA 3, but says the performance gain could be much higher if AMD is successful.
It remains unclear which production node AMD plans to use for the RX 7000 series, but there could be a good reason for this. Given current supply constraints, the fact that Zen 4 will be on TSMC 5nm, and now possibly Nvidia's RTX 4000 series, something like TSMC's 6nm might be needed to guarantee production volume. If the AMD Radeon RX 7000 Series is indeed using a multi-chip design, it can certainly afford to use an older process as a better overall business solution.