AI/DirectML FSR coming to AMD GPUS/RDNA 2?

[UnboxmyLunch]

Recently more information becomes public about RDNA 2, DirectML/DeepLearning benchmarks and AI based FSR.  Could you please suggest if these

 

information is legit? RDNA RX 6000 GPUs really are now capable of competitive to RTX cards, DeepLearning/DirectML features?

 

Does it mean, we can expect that RDNA 2 / Radeon 6000 GPUs, will benefit from the upcoming AI based FSR/Upscaler?
 
 
Sources:

https://www.google.com/amp/s/wccftech.com/amd-microsoft-bring-tensorflow-directml-to-life-4x-improvement-with-rdna-2-gpus/amp/

 

and

 

https://segmentnext.com/amd-fidelityfx-super-resolution-ai/

 

 

btw DLSS 1.0 was not great but was not too horrible and was not using Tensor Cores, from 1.9/2.0 they started to use tensor cores, so in theory even with ZERO directml utilization, AMD could still match DLSS 1.x? http://forum.notebookreview.com/threads/tensor-cores-and-dlss-non-rasterised-performance-of-nvidia-ampere-vs-turing.834939/

 

So with little of DirectML they could potentially add error correction as DLSS 2.x has and improve FSR with Deep Learning?

[igormp]

On 9/18/2021 at 8:08 AM, UnboxmyLunch said:

Could you please suggest if these

 

information is legit?

It is.

On 9/18/2021 at 8:08 AM, UnboxmyLunch said:

RDNA RX 6000 GPUs really are now capable of competitive to RTX cards, DeepLearning/DirectML features?

No, they're not.

On 9/18/2021 at 8:08 AM, UnboxmyLunch said:

btw DLSS 1.0 was not great but was not too horrible and was not using Tensor Cores, from 1.9/2.0 they started to use tensor cores, so in theory even with ZERO directml utilization, AMD could still match DLSS 1.x?

DirectML is an API that allows easy use of hardware for some operations that are widely used in ML. Think of it as DirectX, but for machine learning.

It doesn't change the fact that AMD's consumer GPUs lack any tensor hardware (some Instinct GPUs to feature such hardware), which means that it'll be slower no matter what when compared to RTX GPUs.

 

Also, DirectML's performance still sucks a lot, it's around 10x slower than an equivalent Nvidia GPU running under CUDA, and sometimes even slower than using your CPU.

 

I have run/collected some benchmarks comparing their performance on ML:

On 6/14/2021 at 8:03 PM, igormp said:

Throwing some values in a table with some AMD GPUs for comparison:

+-------------------+---------------+----------------+---------------+----------------+
|    GPU-Imgs/s     | FP32 Batch 64 | FP32 Batch 128 | FP16 Batch 64 | FP16 Batch 128 |
+-------------------+---------------+----------------+---------------+----------------+
| 2060 Super        | 172           | NA             | 405           |            444 |
| 3080              | 396           | NA             | 900           |            947 |
| V100              | 369           | 394            | 975           |           1117 |
| Radeon VII (ROCm) | 288           | 304            | 393           |            426 |
| 6800XT (DirectML) | NA            | 63             | NA            |             52 |
+-------------------+---------------+----------------+---------------+----------------+

 

 

[RyzenFlame]

On 9/19/2021 at 11:49 PM, igormp said:

It is.

No, they're not.

DirectML is an API that allows easy use of hardware for some operations that are widely used in ML. Think of it as DirectX, but for machine learning.

It doesn't change the fact that AMD's consumer GPUs lack any tensor hardware (some Instinct GPUs to feature such hardware), which means that it'll be slower no matter what when compared to RTX GPUs.

 

Also, DirectML's performance still sucks a lot, it's around 10x slower than an equivalent Nvidia GPU running under CUDA, and sometimes even slower than using your CPU.

 

I have run/collected some benchmarks comparing their performance on ML:

 

 

Well -  now Intel and AMD said that XESS will benefit RDNA2 GPUs and use DP4a.... so its gonna be close to DLSS, as they say....

[igormp]

3 hours ago, RyzenFlame said:

 

Well -  now Intel and AMD said that XESS will benefit RDNA2 GPUs and use DP4a.... so its gonna be close to DLSS, as they say....

We need to see how well it'll perform with DP4a. It will surely be slower than using dedicated matrix FMA units, but will it still be good enough for most cases?