Please note that this post is tagged as a rumor.
A huge monolithic GPU or a multi-chip design?
NVIDIA Graphics Device with 124 Compute Units spotted
Just a few days ago we discussed a leak from Geekbench where two graphics devices equipped with 118 and 104 Compute Units were spotted, only to learn that there is even a 124-CU device in the very same database.
Obviously, at this point, we believe that the full specs of the upcoming big core could feature even 128 CUs (8192 CUDA cores) in full form, however, such silicon might never be available for the market with full specs due to manufacturing yields.
The same thing happened with GV100 (Volta) for Tesla V100 and Quadro GV100. The GV100 GPU was only available with 80 Volta Streaming Multiprocessors enabled, despite featuring 84 in total (source).
That said, it is possible that the leaked 124 CU (SM) device is the next-gen Tesla (it would lack just as many SMs as Tesla Volta). We do not know for sure if the next-gen architecture for Tesla is codenamed Ampere, Hopper or something else. However, NVIDIA CEO will host an opening keynote at GTC 2020 (which shifted to an online event) where he might provide an update on these GPUs, next-gen architecture and future roadmaps.
The leaked device was reported with a relatively low GPU clock of 1.11 GHz, which further suggests Tesla or Quadro rather than GeForce device. It is also important to note that we do not know how many CUDA cores are in each Streaming Multiprocessor, as this number has changed in the past. See the chart below for details.
|NVIDIA TESLA (Speculation)|
|SKU||CUs / SMs||CUDA Cores||Memory|
|Next-Gen Device #1|
|Next-Gen Device #2|
|Next-Gen Device #3|
|Tesla V100 (Volta)|
|Tesla P100 (Pascal)|
|Tesla M40 (Maxwell)|
|Tesla K40 (Kepler)|
|Next-Gen: Ampere?, Maxwell SM * 128 = CUDA, Kepler SM*192 = CUDA, Volta & Pascal SM*64 = CUDA|