More Leaks For NVIDIA’s RTX 50 “Blackwell” GPUs Surrounding Its Memory Configurations
More Leaks For NVIDIA’s RTX 50 “Blackwell” GPUs Surrounding Its Memory Configurations
Earlier this week, we reported that leaker @kopite7kimi claimed the memory configuration for NVIDIA RTX 50 “Blackwell” GPUs is “not much different” from its predecessors, the RTX 40 series. Shortly after, various rounds of leaks came out and we are beginning to get a clearer picture of what the memory configuration may eventually look like.
Videocardz reported that the leaker has changed their tone, and is once again claiming that 512-bit memory is a possibility. As for why this is the case, the outlet theorizes that Team Green may be currently exploring all combinations, hence the conflicting information. On top of that, the leaker has said that Blackwell GPUs will be using 28Gbps GDDR7 modules, providing a bandwidth boost over current GDDR6(X) modules that maxes out at 24Gbps.
In the second round of leaks, @kopite7kimi claimed that the 512-bit bus will be used exclusively for the top GB202 silicon (likely reserved for the flagship RTX 5090), but everything below that will largely stay the same. The leaker has also alleged that GB203 (likely RTX 5080) will use a more pedestrian 256-bit bus (and CUDA cores cut in half), whereas GB204 (likely RTX 5070) will be running on a 192-bit bus.
The huge gap between GB202 and GB203 corroborates with one of the earliest leaks of Blackwell architecture, which claimed that the RTX 50 series will feature the “biggest performance leap in NVIDIA’s history”. In any case, NVIDIA may reveal the Blackwell architecture for the datacenter market in the coming GTC 2024, which will be the first in-person GTC since 2019.
Pokdepinion: I have a feeling that NVIDIA is going to boast all the big numbers exclusive to the flagship Blackwell GPU while the rest end up getting incremental upgrades. It’s not reassuring, at least.
