Through the trying glass: A pair of next-gen Nvidia graphics playing cards have been found within the Geekbench database. One has 118 compute models, and the opposite 108. Given compute models usually comprise 64 cores, the 2 playing cards are implied to have 7,552 and 6,912 CUDA cores, respectively… with a catch. Geekbench counts compute models, however the construction of compute models can fluctuate from era to era. Nvidia can be identified to change core configurations between generations.

When transitioning from Pascal to Turing, Nvidia halved the variety of CUDA cores from 128 to 64 per Streaming Multiprocessor (colloquially, the compute unit). However, whereas Pascal has FP32 ALUs because the spine of a CUDA core, Turing pairs an FP32 ALU with an INT32 ALU in each CUDA core, growing the efficiency of every core by about one-third.

You can learn extra about this in our Navi vs. Turing structure comparability.

Nvidia might enhance the per-core efficiency once more with the following era, or, as rumors counsel, go the opposite approach and enhance the ratio of FP32 ALUs to INT32 ALUs in an try to extend effectivity. The backside line is, till Nvidia tells us how they’re configuring their next-gen structure, nothing is assured. What Geekbench registers as a compute unit could also be a tool we’re unfamiliar with, and comprise CUDA cores that carry out higher or worse than what we’re used to.

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

But let’s not spoil all of the enjoyable. These GPUs are, surely, next-gen {hardware} that provide unprecedented ranges of efficiency.

Contained inside the Geekbench entries are the GPUs’ OpenCL benchmark scores. The huge one reaches 184,096 factors and the baby (isn’t that an oxymoron) will get 141,654. For comparability, the RTX 2080 Ti will get roughly 130,000.

It’s additionally a reasonably secure guess that this pair are underperforming members of their species. The huge one had a most clock, as recorded by Geekbench, of 1.11 GHz. The baby ran at 1.01 GHz. By the time the silicon graduates from engineering pattern standing they’ll in all probability attain full-blooded clocks of nicely over 1.5 GHz, and their efficiency will enhance accordingly.

At a guess, I’d say that this pair are prototypes of subsequent era Quadro flagships. Their respective reminiscence capacities of 48 GB and 24 GB exclude them from being gaming playing cards. But Nvidia makes use of nearly an identical silicon for its flagship Quadro and GeForce playing cards, so you might estimate the sequel to the RTX 2080 Ti to have about 7,000 cores – no matter these cores are fabricated from.