Nvidia Ampere A100 GPU scores 446 points in OctaneBench

by Mark Tyson on 27 July 2020, 10:11

Tags: NVIDIA (NASDAQ:NVDA)

Quick Link: HEXUS.net/qaem7v

Add to My Vault: x

PC enthusiasts and gamers are eager to learn more about the next generation consumer graphics cards from Nvidia. Many have already built up their hopes based upon official green team boasts, and various leaks, about the performance of Ampere GPUs. Those information nuggets are still rather thin on the ground though, so the publishing of OptaneBench results from an Nvidia Ampere A100 GPU system user ahead of the weekend was very welcome.

Jules Urbach, CEO of OTOY, the developer and maker of OctaneRender software, shared the Nvidia Ampere A100 GPU benchmark scores on Friday. OctaneRender, and thus OptaneBench, only works on Nvidia CUDA platforms but it is useful for comparing different generation and model GPUs form the green team.

So, what did Urbach find? The Nvidia Ampere A100 GPU scored 446 points in OctaneBench OB4. From this headline result Urbach exclaimed that "Ampere appears to be ~43% faster than Turing in OctaneRender - even w/ RTX off!"

If you look at the OB4 results chart, partly reproduced above, you will see that the best consumer Turing graphics card, the GeForce RTX 2080 Ti, scores 302 points. This is the likely reference for the claim but the A100 provides a 47 rather than 43 per cent performance boost. Actually, the best performing Turing GPU in the results table is the Grid RTX 8000-24Q, scoring 328 points.

In the current OB4 results table you can see Nvidia Volta results take the top places with this GPU architecture behind the leading Titan V, Tesla V100, and Quadro GV100 accelerators. The Nvidia Ampere A100 GPU is still ahead of these, but only by between 11 to 33 per cent.

As a reminder, the Nvidia Ampere A100 GPU is built on TSMC 7nm with 3D stacking, it contains 54 billion transistors, 3rd gen Tensor cores, 3rd gen NVLink and NVSwitch, , supports Multi-instance GPU (MIG), and is paired with 40GB of super-fast HBM2e memory.

Source: Jules Urbach via VideoCardz.



HEXUS Forums :: 9 Comments

Login with Forum Account

Don't have an account? Register today!
But isn't this the absolute cream of the crop top end silicon for the top inferencing cards whereas what will realistically be available is far smaller and lower performance. That 40GB of HBM2e is also going to be literally bank breaking.
What does it says?
I have no idea what the benchmark measures, and how it would translate to normal gaming performance. Will that mean at least 4k@60Hz on ultra with 99,9% time or maybe close to 4k@120 ?
If anything less, then it is failure, no matter the numbers in benchmark.
DevDrake
What does it says?
I have no idea what the benchmark measures, and how it would translate to normal gaming performance. Will that mean at least 4k@60Hz on ultra with 99,9% time or maybe close to 4k@120 ?
If anything less, then it is failure, no matter the numbers in benchmark.
AFAICT it measures raytracing, i would say realtime raytacing but IMO whether something is realtime sort of depends on the detail level and frame rates, technically my old i7 920 could do realtime raytracing if you'd be happy with 1 frame per day. :D
Tabbykatze
But isn't this the absolute cream of the crop top end silicon for the top inferencing cards whereas what will realistically be available is far smaller and lower performance. That 40GB of HBM2e is also going to be literally bank breaking.
would you rather see the intel way of 5% gain in generation?
Actually, Octane renderer is an unbiased pathtracing renderer.

With RTX disabled (Octane 4 - and therefore Octanebench 4 doesn't support RTX - the 2020 version does), the benchmarks are basically about pure FP32 grunt.

So it doesn't really equate to gaming, and with RTX disabled, not realtime raytracing either. Otoy does have tech for that, but Octane is more about ‘final’ quality for TV and film etc rather than realtime, currently.

With RTX enabled, the Turing gen was getting 1.5-4x speed improvement, especially in scenes with more raytracing of course. So, for people doing CGI, it will be also very interesting to see what the Ampere RTX speedup is.