In the course of the rescheduled keynote for the GPU Technology Conference (GTC) 2020, Nvidia announced specifications for the server GPU A100 based on the ampere architecture.
Although they only allow indirect conclusions to be drawn about the performance of the desktop models, the old three-point set can be used to derive in which direction it could go.
But be careful, here we use the simplest mathematics – not to say the lumberjack method – to uncover the secrets of the RTX 3000.
What does the keynote tell us about amps?
SMs (Streaming Multiprocessors)
FP32 performance (peak)
It is particularly interesting that the The main points of the specifications do not differ or differ only slightly from the rumors of the past few months
In terms of core numbers, however, the leaks and rumors were all the more accurate: In fact, it was speculated that the GA100 chip would have 8,192 CUDA cores and 48 GB HBM2 memory. At the same time, benchmarks for an Nvidia graphics card with 6,912 cores also appeared, which – as we now know – corresponds to the A100 GPU.
What does that mean for the RTX 3000?
With regard to the desktop models, nothing can be said with definite certainty, but since the rumors about the server and professional models have proven to be correct, it seems quite realistic to assume this also for player graphics cards.
In addition, the basic specifications of the GA100 chip are now known. Of course, this will only be available as a cropped version (probably GA102, 103, etc.) for the desktop, but it seems certain that the core number will increase considerably.
Previous rumors about RTX 3080 Ti and Co:
- GA102 (RTX 3080 Ti) – 84 SMs / 5,376 CUDA cores / 12/24 GB GDDR6 384 bits (possibly 18 Gbps with a bandwidth of 864 GB / s)
- GA103 (RTX 3080) – 60 SMs / 3,840 CUDA cores / 10/20 GB GDDR6 320 bits
- GA104 (RTX 3070) – 48 SMs / 3,072 CUDA cores / 8/16 GB GDDR6 256 bits
The clock rates are to be increased by 200 to 300 MHz compared to Turing ( RTX 2000 and GTX 1600) – there is talk of a 2.2 GHz turbo clock.
All information, leaks and rumors about the RTX 3000 can be found in the following article:
more on the subject
Nvidia RTX 3000 Amp Release & Specs
What can we derive from this for the performance?
It was rumored again and again that an RTX 3080 Ti as a new flagship would have a theoretical computing power of 21 TFLOPS. For comparison, the RTX 2080 Ti comes to 13.45 TFLOPS according to its official specifications.
Based on the specifications of the Nvidia A100 GPU, you can calculate the performance per clock with the help of the set of three. This can roughly determine the performance per core and from this in turn the theoretical computing power of the entire GPU.
Depending on the assumed clock rate, the RTX 3080 Ti actually has a range between around 19 and 23 TFLOPS for suspected (and not unrealistic-sounding) 5,376 CUDA cores – an average of 21 TFLOPS.
The values determined for the RTX 3080 Ti have to be viewed with caution, but the overall picture from what we now know from the keynote with certainty and what was previously assumed is quite correct.
When should the RTX 3000 be released?
In this respect, the keynote was unable to provide any new information. So far, however Release towards the end of the year, possibly in autumn, in the room.
A product presentation could take place at the Computex 2020 computer fair in Taipei, which was postponed from early June to 28th to 30th September. As usual, appropriate teasers would be expected in the weeks or months before.