Nvidia ‘Blackwell’ GPU leak points to significant memory speed upgrades

Contrary to previous rumours, it looks like the memory buses of the Nvidia RTX 50 series won't be much different from what we've seen in the RTX 30 and 40 series. Based on recent reports, the flagship GB202 GPU will feature a 384-bit memory bus, just like its predecessors. However, the memory will be considerably faster, running at 28 Gbps.

According to Kopite7kimi, the RTX 50 series will not feature a 512-bit memory bus, which had previously been hinted at. Instead, the flagship RTX 50 GPU, known as GB202, will retain the same 384-bit memory bus configuration of the RTX 3090 and RTX 4090. In fact, the entire ‘Blackwell' family of GPUs is expected to have memory bus specifications similar to those of Ampere and ADA.

Although I still have fantasies about 512 bit, the memory interface configuration of GB20x is not much different from that of AD10x.

— kopite7kimi (@kopite7kimi) March 9, 2024

Although the memory bus specifications for the RTX 50 series may not see a significant upgrade, it's expected that the memory will. We already know Nvidia will use GDDR7 memory on the upcoming GPUs, but it's still unclear at what speed it will operate. However, according to Kopite7kimi, the VRAM will run at 28 Gbps, up from 21–23 Gbps from the RTX 40 series cards using GDDR6X memory. Assuming a similar memory configuration on the RTX 50 series, maximum memory bandwidth would increase by up to 33%.

The Blackwell GPUs are said to feature TSMC's N3 node and DisplayPort 2.1. The line-up is expected to be led by the GB202 GPU, which is rumoured to feature up to 24,000 CUDA cores.

Discuss on our Facebook page, HERE.

KitGuru says: We are still quite a ways off from next-gen GeForce graphics cards from Nvidia but the leaks have begun. Assuming these are on track for late 2024 or early 2025, we should hear much more throughout the year. 

Become a Patron!