Two Unknown Nvidia GPUs Spotted With Insane Core Counts

(Image credit: Shutterstock)

Twitter user @_rogame has shared his discovery of two test results for very interesting Nvidia graphics cards that seemingly pack an enormous amount of cores and memory. These might be Ampere or even Hopper, The benchmarks date as far back as October of last year. As with any unverified test submission, it's important to approach the specifications with a bit of caution.

The first graphics card allegedly features 118 Compute Units (CUs), which is equivalent to what Nvidia calls Streaming Multiprocessors (SMs). Assuming that Nvidia retains the 64 CUDA cores per SM setup for its next-generation products, the unknown graphics card should have 7,552 CUDA cores. The graphics card is listed with a 1.11 GHz base clock and 23.8GB of onboard memory.

The second graphics card reportedly sports 108 CUs, amounting to 6,912 CUDA cores. The graphics card ticks with a 1.10 GHz base clock and has 46.8GB of memory at its disposal.

The base clocks seem fairly low and the memory configuration looks a bit off for both graphics cards. It is worth noting that these are likely early engineering samples, so the final specifications can vary. Also when it comes to unreleased hardware, the software probably isn't picking up the specifications correctly. It's rational to expect the two graphics cards to actually have a 24GB and 48GB configuration, respectively.

Swipe to scroll horizontally
ModelArchitecture (GPU)Streaming MultiprocessorCUDA CoresBase ClockMemory CapacityTransistor CountDie Size
Nvidia GPU 1?1187,5521,110 MHz23.6GB??
Nvidia GPU 2?1086,9121,010 MHz46.8GB??
Tesla V100SVolta (GV100)805,1201,245 MHz32GB HBM221.2 billion815 mm²
Quadro RTX 8000Turing (TU102)724,6081,395 MHz48GB GDDR618.6 billion754 mm²
Titan RTXTuring (TU102)724,6081,350 MHz24GB GDDR618.6 billion754 mm²
GeForce RTX 2080 TiTuring (TU102)684,3521,350 MHz11GB GDDR618.6 billion754 mm²

It's pretty obvious that the two puzzling graphics card aren't part of the GeForce family. The GeForce RTX 2080 Ti, which is the current flagship, comes with 4,352 CUDA cores and 11GB of GDDR6 memory. We doubt that its successor will arrive with over 50% more CUDA cores. It's entirely possible because of Ampere's die-shrink to the 7nm node, but it's highly unlikely that a GeForce graphics card will come that many CUDA cores.

For some context, the generation-over-generation upgrade from the previous GeForce GTX 1080 Ti to the current GeForce RTX 2080 Ti only brought around a 21% increase in CUDA cores. Besides, it would be useless to have 24GB or 48GB on a gaming graphics card because such high-density memory will only inflate pricing unnecessarily.

The 24GB graphics card matches that of the Nvidia Titan RTX in terms of memory capacity. Historically, the Titan graphics cards are similarly specced to their GeForce counterparts. Since GeForce is out of the question, logically the graphics cards could only belong to Nvidia's Quadro or Tesla lineup. At any rate, the mysterious Nvidia graphics cards will deliver crazy levels of performance.

Strangely enough, the obscure graphics card popped up inside Z370-based systems along with the Core i7-8700K. It's weird to see a Tesla graphics card paired with a mainstream system, we would normally expect to see them paired with EPYC or Xeon systems. However, we can't discard the possibility that someone tested those graphics card with whatever system they had at hand. In any case, it's healthy to take the results with a pinch of salt.

The graphics card with the 7,552 CUDA cores put up an impressive score of 184,096 points while the one with 6,912 CUDA cores scored 141,654 points in Geekbench 5's OpenCL benchmark. The highest OpenCL scores for the Quadro RTX 8000 and Tesla V100 are 143,438 points and 171,075 points, respectively.

Nvidia will be hosting its annual GPU Technology Conference (GTC) very soon. Let's cross our fingers that the chipmaker surprises us once more.

Zhiye Liu
RAM Reviewer and News Editor

Zhiye Liu is a Freelance News Writer at Tom’s Hardware US. Although he loves everything that’s hardware, he has a soft spot for CPUs, GPUs, and RAM.

  • escksu
    Given the ongoing trend esp. in deep learning, I won't be surprise if these are meant for that. I seriously doubt they are meant for gaming.