Is NVIDIA reviving the RTX TITAN brand of halo-segment graphics cards with "Blackwell"? A curious-looking GB202-200-A1 spy-shot making rounds on ChilHell hints at the possibility. The upcoming GeForce RTX 5090 is the company's flagship product from the RTX 50-series "Blackwell" generation, although it does not max out the 4 nm "GB202" silicon on which it is based. The RTX 5090 enables 170 out of the 192 SM (streaming multiprocessors) physically present on the "GB202." This leaves NVIDIA with a lot of room to carve out either a halo-segment SKU such as the RTX TITAN Blackwell, or a Pro-Vis (professional visualization) product that also targets the AI research community.

NVIDIA's top Pro-Vis product tended to have more SMs enabled than the top GeForce RTX product, while having lower clock speeds, so its target users have access to more FP64-capable CUDA cores, more Tensor cores, etc. However, over the past two generations, NVIDIA discontinued the practice of giving its RTX GPUs large numbers of FP64 cores that are disabled on GeForce RTX products, to make die-space for the Tensor cores. This hence makes it more likely that a maxed-out "GB202" is the RTX TITAN Blackwell, and not a Pro-Vis product.A maxed out "GB202" would have 24,576 CUDA cores, 768 Tensor cores, 192 RT cores, 768 TMUs, and 192 ROPs, besides 88 MB of L2 cache. The RTX 5090 is configured with 170 SM, so it gets 21,760 CUDA cores, 680 Tensor cores, 170 RT cores, 680 TMUs, and 192 ROPs. Luckily, it gets to keep all 88 MB of L2 cache present on the silicon, which cushions its awesome 512-bit GDDR7 memory interface.

This brings us to the next lever NVIDIA could pull to create the RTX TITAN Blackwell—memory speed. The RTX 5090 is configured with 28 Gbps memory speed, to achieve a 1.792 TB/s memory bandwidth, although there exist 32 Gbps-capable GDDR7 memory chips. If configured with 32 Gbps memory speeds, the RTX TITAN Blackwell would rock a mammoth 2 TB/s of memory bandwidth.

Circling back to the ChipHell spy-shot, and we are shown two 12V2x6 power connectors. This could be our second indication that the maxed-out "GB202" is a halo-segment product and not a Pro-Vis one (which tend to have lower TGP values than their GeForce RTX counterparts). A maxed out "GB202" and 32 Gbps memory speed could drive TGP past the 600 W-mark, and such a product could call for two 12V2x6 inputs.

Could this blurry spy shot be from a retail-ready product? Unlikely—this could well be an NVIDIA-internal reference design, and the final product could more closely resemble the Double Flow Through architecture of the RTX 5090 Founders Edition, or perhaps even similar to the "RTX 4090 Ti prototype" we saw in 2023, which had a sideways-oriented ruler-shaped PCB, with a 4-slot thick cooling solution on top.