Rumor mill: At GTC 2024 this week, Nvidia unveiled its Blackwell GPU platform for AI workloads. However, the company said nothing about its upcoming consumer GPUs based on the same platform, leaving many gamers fuming at its newfound priorities. Thankfully, a noted leaker has now revealed some alleged key details about the GB202 GPU that’s expected to underpin the flagship Blackwell gaming card, the GeForce RTX 5090.
The latest news comes from tipster @kopite7kimi, who claims that the GB202 GPU will use the same TSMC 4N foundry node as the GB100 designed for AI acceleration and computing applications. He also clarified that contrary to some reports, the TSMC 4N node is a derivative of the company’s 5nm EUV process and not 4nm.
In the same post, the tipster also claimed that the new node will offer a 30 percent increase in transistor density compared to TSMC’s mainstream 4NP node, which itself brought a 6 percent increase in transistor density and 22 percent improvement in power efficiency over N5. Either way, if the report is accurate, it means Blackwell is also using a variant of TSMC’s 5nm node like Ada.
GB202 will use the same process node as GB100. I must clarify once again that TSMC 4N(vidia) is based on TSMC 5, not 4nm.
I’m sorry I cannot match Jensen’s naming with TSMC’s naming. We need professional chip analysis to determine.
At least, there is a 30% increase in density.– kopite7kimi (@kopite7kimi) March 19, 2024
In a separate post (since deleted), Kopite further claimed that with Blackwell, Nvidia is increasing the L1 caches of the streaming multiprocessors, suggesting that the company is working on improving the performance of its GPUs at the SM level. According to the post, the flagship GB202 gaming GPU will bring a “significant improvement” in the amount of L1 cache over AD102 and GA102, both of which have L1 caches that are 128 KB in size.
There’s no more new information about the RTX 5090 for now but leaked details in recent weeks have already revealed a few tidbits. According to an earlier report by the same tipster, the GB202 GPU could feature a 512-bit memory interface compared to the 384-bit interface found on the flagship AD102 Ada GPU. The RTX 5090 is also tipped to ship with 28Gbps GDDR7 memory.
If the recent leaks turn out to be accurate, it could mean the RTX 5090 will be a performance beast, and one that AMD could find hard to beat. That said, these are unverified rumors and need to be taken with a pinch of salt for now. In any case, the RTX 50-series is only expected to arrive in late 2024 or early 2025, so a lot can change between now and then.