Notebookcheck Logo

Rumor | Nvidia RTX 50 GB102 Blackwell early internal performance gains only in range of Pascal to Turing and not 2.6x from AD102, increased SM counts on the anvil

Nvidia is targeting substantial performance increases with RTX 50 Blackwell GB102. (Image: GH100 Hopper via Nvidia)
Nvidia is targeting substantial performance increases with RTX 50 Blackwell GB102. (Image: GH100 Hopper via Nvidia)
Nvidia's next gen RTX 50 Blackwell with purported 144 SMs isn't reaching ambitious performance gains in internal testing, according to latest rumors. Apparently, Blackwell's performance gains vis-à-vis AD102 are in the range of what we've seen from Pascal to Turing, but Nvidia wants more. This could imply that Nvidia might go for higher SMs counts with GB102 and possibly even a multi-chip module architecture similar to the purported GB100 datacenter GPU.

Nvidia hasn't fully completed its RTX 40 series Ada lineup yet — we still expect the RTX 4060, RTX 4060 Ti and possibly, the RTX 4050 to launch in the coming weeks and months. However, rumors of Nvidia's next big consumer GPU have been trickling now and then. 

Paul from the YouTube channel RedGamingTech, who has been putting out quite a bit of information about RTX 50 codenamed Blackwell, has shared a few additional pointers about what's cooking in Nvidia's labs. According to Paul, RTX 50 Blackwell indeed be based on a TSMC 3 nm process that is customized for Nvidia. This, along with some of the information that Paul had shared earlier, continues to hold water as of now.

While the grapevine is predicting up to a 2.6x performance increase for Blackwell over Ada, Paul's sources seem to indicate that things aren't exactly rosy in internal testing. Early engineering samples don't seem to be hitting close to the anticipated performance targets for the consumer GPU despite achieving high frequencies. Apparently, this is on the lines of a performance uplift between Pascal to Turing in raster but Nvidia is aiming higher. This could mean that Nvidia may choose to increase the SM count if needed.

Nvidia is currently testing the GB100 GPU, which is likely to be targeted at the datacenter with 256 SMs, 5120-bit HBM3, and a 128 MB L2 cache. This is expected to be a multi-chip module (MCM) architecture.

Nvidia has been toying with the idea of an MCM for quite some time now. The GH100 Hopper datacenter GPU was expected to be an MCM back in the day, but it launched as a monolithic die with 144 SMs. With increased SM counts expected for GB100, we might finally get to see an MCM package from Nvidia.

The consumer RTX 50 GB102 Blackwell is rumored to sport 144 SMs together with 384-bit GDDR7, 96 MB L2 cache, and support for PCIe Gen 5 x16. Information is scarce at this point, but assuming Blackwell sports a similar SM count as AD102, there would be obvious changes such as next generation Tensor and RT cores together with a host of new features including advanced ray tracing pipelines that directly integrate denoising.

That being said, it is still very early days for RTX 50 Blackwell. We are likely to come across contrasting reports for quite a foreseeable time at least till things get on production lines, so just savor the leak with a pinch of salt.

Buy the MSI Gaming GeForce RTX 4070 12 GB on Amazon

Read all 6 comments / answer
static version load dynamic
Loading Comments
Comment on this article
Please share our article, every link counts!
> Expert Reviews and News on Laptops, Smartphones and Tech Innovations > News > News Archive > Newsarchive 2023 05 > Nvidia RTX 50 GB102 Blackwell early internal performance gains only in range of Pascal to Turing and not 2.6x from AD102, increased SM counts on the anvil
Vaidyanathan Subramaniam, 2023-05-12 (Update: 2023-05-12)