Through the looking glass: A trio of next-gen Nvidia graphics cards have been discovered in the Geekbench database. One has 124 compute units and the other two have 118 and 108. Given compute units generally contain 64 cores, the cards are implied to have 7,936, 7,552 and 6,912 CUDA cores, respectively… with a catch. Geekbench counts compute units, but the structure of compute units and the cores they contain can vary from generation to generation.

Update: An earlier version of this article published on February 29 detailed only two of the three GPUs. The largest, with 7,936 cores, was uncovered only recently.

When transitioning from Pascal to Turing, Nvidia halved the number of CUDA cores from 128 to 64 per Streaming Multiprocessor (colloquially, the compute unit). However, while Pascal uses FP32 ALUs as the backbone of a CUDA core, Turing pairs an FP32 ALU with an INT32 ALU in every CUDA core, increasing the performance of each core by about one-third.

You can read more about this in our Navi vs. Turing architecture comparison.

Nvidia could boost the per-core performance again with the next generation, or, as rumors suggest, go the other way and increase the ratio of FP32 ALUs to INT32 ALUs in an attempt to increase efficiency. The bottom line is, until Nvidia tells us how they’re configuring their next-gen architecture, nothing is guaranteed. What Geekbench registers as a compute unit may be a device we’re unfamiliar with, and contain CUDA cores that perform better or worse than what we’re used to.

Model Mystery GPU 0 Mystery GPU 1 Mystery GPU 2 Quadro RTX 8000
CUs/SMs 124 118 108 72
CUDA Cores 7936 7552 6912 4608
Clock Speed 1110 MHz 1110 MHz 1010 MHz 1770 MHz
Memory 32 GB 24 GB 48 GB 48 GB

But let’s not spoil all the fun. These GPUs are, without a doubt, next-gen hardware that offer unprecedented levels of performance.

Contained within the Geekbench entries are the GPUs’ OpenCL benchmark scores. The largest card achieves 222,377 points. Two entries exist for the middle card with scores of 184,096 and 169,368, and the little one (isn’t that an oxymoron) gets 141,654. For comparison, the RTX 2080 Ti gets roughly 130,000.

It’s also a pretty safe bet that this set are underperforming members of their species. The largest pair have maximum clocks of 1.1 GHz, as recorded by Geekbench. The little one ran at 1.01 GHz. By the time the silicon graduates from engineering sample status they’ll probably reach full-blooded clocks of well over 1.5 GHz, and their performance will improve accordingly.

At a guess, I’d say that this trio are prototypes of next generation Quadro flagships. Their respective memory capacities of 48 GB, 32 GB, and 24 GB exclude them from being gaming cards. But Nvidia uses almost identical silicon for its flagship Quadro and GeForce cards, so you could estimate the sequel to the RTX 2080 Ti to have about 7,000 cores – whatever those cores are made of.