We know that NVIDIA’s upcoming GPUs will be releasing in a few weeks and so more and more leaks are now getting surfaced suggesting the power of these GPUs. The leakster @_rogame managed to unearth Geekbench 5 OpenCL scores of some unknown NVIDIA GPU.
The scores of this GPU has put the best ever Titan RTX GPU to shame, the GPU scored 184,096 points in OpenCL beating the Titan RTX by nearly 39%. There are actually two cards spotted, this one comes with 118 compute units (CUs) or 7,552 CUDA cores, a maximum frequency of 1.11 GHz, along with 24 GB of VRAM.
The other GPU features 108 CUs or 6,912 CUDA Cores with1.01 GHz max clock speed and 46.8 GB of reported VRAM. This GPU scored 141,654 points in the same test.
If compared with the latest flagship GPUs, GV100 scores 142,837 while a Tesla V100 GPU scores 154,606, and a Titan RTX manages about 132,804 in the same Geekbench 5 OpenCL benchmark.
If we do believe the GPU is an Ampere GPU, it nearly boasts of 39% increased performance over the Titan RTX when considering the card with 7,552 CUDA Cores and about a mere 6.7% improvement when comparing the GPU with 6,912 CUDA Cores.
Even though we cannot deduce some solid info from the leaked benchmarks but we can see that these are not the consumer-oriented Ampere GPUs but the premium ones. Although these GPUs have high core counts this is not the full chip.
The GA100 is supposed to be the full version of Ampere GPU series with 8,192 CUDA Cores. Also, the fact remains the type of memory used has not been confirmed whether it’s a GDDR6 one or HBM2/2e.
These newly found GPUs can be the new Tesla GPUs because an RTX 3080 is assumed to have 3,480 CUDA Cores and 10/20 GB 320-bit GDDR6 memory, featuring a GA103 chip.
The new NVIDIA Ampere GPUs are scheduled to be released at the GTC 2020 which is supposed to be held on March 23 even though of the Coronavirus outbreak but it will be an online-only event this time.
We can expect NVIDIA CEO Jensen Huang giving a keynote and actually launching the new GPUs and sharing more about its new architecture.
Do check out: