Rumor mill: At GTC 2024 this week, Nvidia unveiled its Blackwell GPU platform for AI workloads. Nonetheless, the corporate mentioned nothing about its upcoming shopper GPUs based mostly on the identical platform, leaving many players fuming at its newfound priorities. Fortunately, a famous leaker has now revealed some alleged key particulars concerning the GB202 GPU that is anticipated to underpin the flagship Blackwell gaming card, the GeForce RTX 5090.
The most recent information comes from tipster @kopite7kimi, who claims that the GB202 GPU will use the identical TSMC 4N foundry node because the GB100 designed for AI acceleration and computing purposes. He additionally clarified that opposite to some stories, the TSMC 4N node is a by-product of the corporate’s 5nm EUV course of and never 4nm.
In the identical put up, the tipster additionally claimed that the brand new node will provide a 30 p.c improve in transistor density in comparison with TSMC’s mainstream 4NP node, which itself introduced a 6 p.c improve in transistor density and 22 p.c enchancment in energy effectivity over N5. Both method, if the report is correct, it means Blackwell can also be utilizing a variant of TSMC’s 5nm node like Ada.
GB202 will use the identical course of node as GB100. I have to make clear as soon as once more that TSMC 4N(vidia) is predicated on TSMC 5, not 4nm.
I am sorry I can’t match Jensen’s naming with TSMC’s naming. We’d like skilled chip evaluation to find out.
At the very least, there’s a 30% improve in density.– kopite7kimi (@kopite7kimi) March 19, 2024
In a separate post (since deleted), Kopite additional claimed that with Blackwell, Nvidia is rising the L1 caches of the streaming multiprocessors, suggesting that the corporate is engaged on enhancing the efficiency of its GPUs on the SM stage. In accordance with the put up, the flagship GB202 gaming GPU will convey a “vital enchancment” within the quantity of L1 cache over AD102 and GA102, each of which have L1 caches which might be 128 KB in measurement.
There is no extra new details about the RTX 5090 for now however leaked particulars in current weeks have already revealed a couple of tidbits. In accordance with an earlier report by the identical tipster, the GB202 GPU may function a 512-bit reminiscence interface in comparison with the 384-bit interface discovered on the flagship AD102 Ada GPU. The RTX 5090 can also be tipped to ship with 28Gbps GDDR7 reminiscence.
If the current leaks transform correct, it may imply the RTX 5090 can be a efficiency beast, and one which AMD may discover arduous to beat. That mentioned, these are unverified rumors and should be taken with a pinch of salt for now. In any case, the RTX 50-series is just anticipated to reach in late 2024 or early 2025, so lots can change between at times.
