In the past few days and weeks, rumors about Nvidia's upcoming GeForce RTX 5000 series have increased. Although it will probably only appear in 2025 under the code name Blackwell, the configurations of the underlying graphics chips are likely to have already been determined. The GPUs should not be confused with the Blackwell server offshoot B100. Its announcement is expected next week as part of Nvidia's GTC trade fair.
Advertisement
Most of the information about the RTX 5000 series is coming from the X/Twitter account “kopite7kimi”, who has been correct many times before. According to him, the largest gaming GPU GB202 holds 192 streaming multiprocessors (SMs), organized into 12 graphics processing clusters (GPCs). That would result in 24,576 shader cores if it remained at 128 per SM, and would be a third more than the largest graphics chip from the current Ada Lovelace generation (AD102 with 144 SMs or 18,432 shaders).
Traditionally, Nvidia likes to sell graphics cards with partially deactivated GPUs to gamers, while the fully functional chips go on more profitable professional graphics cards. In the AD102 version on the GeForce RTX 4090, 11 percent of the shader cores are deactivated.
Neues GDDR7-RAM
There was disagreement as to whether the GB202 has a 512 or 384 bit wide memory interface. Currently it looks like 512 bits with GDDR7 memory. The latter should initially not exhaust the specification and clock at a moderate rate of 28 Gbit/s per pin. This would result in a transfer rate of almost 1.8 TByte/s on 512 data lines – 78 percent more than with the GeForce RTX 4090. Nvidia could increase the RAM capacity from 24 to 32 GB.
However, all of this assumes that Nvidia does not cut the memory interface. Even crooked configurations of around 448 bits would be possible.
GB202 halved
The next smaller Blackwell GPU, GB203, is said to be significantly weaker than the GB202. We're talking about a smooth halving, i.e. 12,288 shader cores in six GPCs with a 256-bit interface.
It is questionable whether Nvidia relies on a multi-chip design for graphics cards. Then the GB202 could consist of two GB203. Until now, such constructs, in which several chips calculate on the same image, were problematic because of the latency. The GB202 could still have a monolithic structure and the GB203 could be halved for design reasons.
If Nvidia proceeds as with the RTX 4000 series and provides the GB202 for the GeForce RTX 5090 and the GB203 for the GeForce RTX 5080, a large performance difference can again be expected.
(mma)