The rumors that Nvidia plans to use a 384-bit memory bus for its top model in the Blackwell series, the GB202 GPU are getting more solidified. This is the same memory bus size used in their current high-end model in the Ada Lovelace series, the GeForce RTX 4090. Nvidia has also started sharing its next-gen AI GPUs, known as Blackwell B100 and B200, with important partners and clients.
There were some guesses in mid-February that the RTX 50-series might support the newer PCIe 6.0 CEM standard, but there hasn't been much news on this front since the end of last year.
There were some guesses in mid-February that the RTX 50-series might support the newer PCIe 6.0 CEM standard, but there hasn't been much news on this front since the end of last year.
@kopite7kimi recently talked about what we might expect from the gaming GPUs in the Blackwell series. It looks like they might keep the same memory setup as the Ada Lovelace GPUs. This has been a bit disappointing for those who were hoping for a bigger change to a 512-bit interface.
One exciting possibility is that the GeForce RTX 50xx series could use a newer type of memory called GDDR7, following a recent announcement of this new memory standard by JEDEC. Even with this new memory, it's expected that the design will still use 16 Gbit memory chips, which means each chip would hold 2 GB of data.
The lineup could look like the following:
- GB202 - 384-bit / 32 Gbps / 24 GB (Max Memory) / 1536 GB/s (Max Bandwidth)
- GB203 - 256-bit / 32 Gbps / 16 GB (Max Memory) / 1024 GB/s (Max Bandwidth)
- GB204 - 192-bit / 32 Gbps / 12 GB (Max Memory) / 768.0 GB/s (Max Bandwidth)
- GB206 - 128-bit / 32 Gbps / 8 GB (Max Memory) / 512.0 GB/s (Max Bandwidth)
- GB207 - 128-bit / 32 Gbps / 8 GB (Max Memory) / 512.0 GB/s (Max Bandwidth)
Sources: kopite7kimi Tweet, Tom's Hardware, Wccftech
