Just as we're all enjoying the weekend, rumors of NVIDIA's next generation Ampere GPU have surfaced -- this time, teasing some basic specifications, purported performance, and more.
We know that NVIDIA's next-gen GPU architecture after Turing will be Ampere, and then Hopper -- with Ampere finding a new home on the 7nm node and offering some truly kick ass performance improvements as well as power efficiency. Pascal and Turing were already quite power efficient GPUs, but Ampere really amps things up... pun intended.
This is where Twitter user '_roggame' comes in, where he has discovered some Geekbench results for "two unknown NVIDIA GPUs". The first of which packs 7552 CUDA (so Ampere) cores, 118 compute units (CUs), GPU clock speeds of 1.11GHz and 24GB of memory. We could be looking at GDDR6 or HBM2/2e, that is not confirmed. The next one has 6912 CUDA cores, 180 compute units, and a GPU clock of 1.01GHz while it packs much more memory at 48GB.
NVIDIA Ampere GPUs Teased
- NVIDIA Ampere GPU 1: 6912 cores
- NVIDIA Ampere GPU 2: 7552 cores
- NVIDIA Ampere GPU 3: 8192 cores (full version of Ampere)
We're probably looking at a Tesla card here, and not a traditional gaming-focused GeForce or professional-focused Quadro. With this level of performance, the purported Ampere-based GPU destroys the Tesla V100 by a good amount.
Geekbench 5 results from the new card seeing it smash out an OpenCL score of 184,096 -- a chunk ahead of the Tesla V100 with 154,606, the GV100 with 142,837, and the TITAN RTX with 132,804. In the middle of that we have the Ampere GPU with 6912 CUDA cores and a Geekbench 5 result of 141,654 -- beating out the TITAN RTX easily in its early form.
Geekbench 5 Results (Incl. Ampere GPUs)
- Mystery Ampere GPU 1: 184,096
- Mystery Ampere GPU 2: 141,654
- NVIDIA Tesla V100: 154,606
- NVIDIA GV100: 142,837
- NVIDIA TITAN RTX: 132,804
NVIDIA's flagship TITAN RTX costs $2499 and finds itself as the fastest consumer-focused Turing-based graphics card the company sells, with a $1000+ gap between it and the flagship GeForce RTX 2080 Ti graphics card which comes in Founders Edition form, but as well as many colorful editions with custom cards offered by various AIBs.
The TITAN RTX rolls out with 24GB of GDDR6 memory, a huge 4608 CUDA cores, 1770MHz boost clock and its 24GB of GDDR6 on a wide 384-bit memory bus that offers a super-fast 672GB/sec of memory bandwidth. We don't even have the results from the fastest version of Ampere with its huge 8192 cores and it's already knocking Volta and Turing around the ring.
If you were looking at this post wondering if the Ampere GPU here was going to be powering the upcoming GeForce RTX 3080 and GeForce RTX 3070, then I'm sorry I've disappointed you. I have an article on that here with the latest purported specs from mid-January 2020.
We know to expect both the Ampere GPUs inside of the new GeForce RTX 3080 and GeForce RTX 3070 to be GA103 and GA104, respectively. As for specs, here's what we've heard so far:
GA103 (GeForce RTX 3080)
- 10/20GB GDDR6
- 320-bit memory interface
- 60 SMs
- 3480 stream processors
GA104 (GeForce RTX 3070)
- 8/16GB GDDR6
- 256-bit memory interface
- 48 SMs
- 3072 stream processors