AMD Radeon RDNA2 Navi (RX 6500, 6600, 6700, 6800, 6900 XT)

[Deleted member 2197]

Is there a link to the TechPowerUp results?
Edit: Got it from the other thread.

 

[Flappy Pannus]

Trigger Warning: these results have no raytracing or DLSS enabled AFAICT. It's just rasterization without upscale or reconstruction.

Please do move along if games without DLSS or raytracing offend you so much that you feel the need to derail the discussion with endless DLSS and raytracing propaganda.

Dude come on, you're really not the one on this site to use the 'triggered' accusation for other people. You need to dial it back a tad.

 

[PSman1700]

Dude come on, you're really not the one on this site to use the 'triggered' accusation for other people. You need to dial it back a tad.

Reporting doesnt help either. This user can do whatever he wants unfortanately. Try contacting Rys (the owner of this board).

 

[BRiT]

Dude come on, you're really not the one on this site to use the 'triggered' accusation for other people. You need to dial it back a tad.

Reporting doesnt help either. This user can do whatever he wants unfortanately.

Not accurate. User is banned.

 

[PSman1700]

Not accurate. User is banned.

We didnt know, thanks

 

[Lurkmass]

Blender 3.0 beta Cycles benchmarks using the new HIP backend by Thomas Dinges. Brian Savery made it all possible ...

 

[CarstenS]

Very nice results too:

https://twitter.com/x/status/1460275115455885317

 

[Deleted member 2197]

Very nice results too:

https://twitter.com/x/status/1460275115455885317

The Classroom scene usually depicts good performance for AMD, so will be interesting to see reviews using the full Blender testing suite.

 

[trinibwoy]

Very nice results too:

https://twitter.com/x/status/1460275115455885317

What does it mean that the 3090 is faster vs Optix and CUDA. The 3090 is still running CUDA right just compiled from HIP?

 

[CarstenS]

3090 is running CUDA and Optix, as before. Cycles X got faster. It does not, for instance, subdivide the picture in small tiles anymore.

 

[trinibwoy]

3090 is running CUDA and Optix, as before. Cycles X got faster. It does not, for instance, subdivide the picture in small tiles anymore.

Oh nice. Impressive gains for the 6900. OpenCL really does suck.

 

[xpea]

Oh nice. Impressive gains for the 6900. OpenCL really does suck.

I would not call it impressive when you are more than 3 times slower than your competition... Even worst, half priced NV card is also faster...

 

[trinibwoy]

I would not call it impressive when you are more than 3 times slower than your competition... Even worst, half priced NV card is also faster...

That's a whole other story. Looking forward to a proper review but it does seem Ampere's abundance of flops is coming into play. Or maybe it's just due to maturity of CUDA vs HIP.

 

[nAo]

That's a whole other story. Looking forward to a proper review but it does seem Ampere's abundance of flops is coming into play. Or maybe it's just due to maturity of CUDA vs HIP.

It's likely the ray tracing HW also contributes to better perf..

 

[trinibwoy]

It's likely the ray tracing HW also contributes to better perf..

For Optix sure but the CUDA path doesn't use RT hardware right? Is the HIP implementation using RT hardware on RDNA2?

 

[Lurkmass]

For Optix sure but the CUDA path doesn't use RT hardware right? Is the HIP implementation using RT hardware on RDNA2?

There are no ray tracing APIs in HIP that I'm currently aware of. There used to be plans for an OpenCL ray tracing extension before it was shelved. Now that HIP is their official compute API even on Windows, they're looking into exposing ray tracing for it ...

 

[AlexV]

There are no ray tracing APIs in HIP that I'm currently aware of. There used to be plans for an OpenCL ray tracing extension before it was shelved. Now that HIP is their official compute API even on Windows, they're looking into exposing ray tracing for it ...

Whilst not necessarily ideal, we do expose (and have been exposing) the relevant builtins for a while: https://reviews.llvm.org/D104946. They map directly to the instructions described in 8.2.10 here https://developer.amd.com/wp-content/resources/RDNA2_Shader_ISA_November2020.pdf, and can be used as-is in HIP __device__ code.

 

[Lurkmass]

Whilst not necessarily ideal, we do expose (and have been exposing) the relevant builtins for a while: https://reviews.llvm.org/D104946. They map directly to the instructions described in 8.2.10 here https://developer.amd.com/wp-content/resources/RDNA2_Shader_ISA_November2020.pdf, and can be used as-is in HIP __device__ code.

Ah but nonetheless I still don't think it's been used yet in blender. What other improvements is your team planning to make ray tracing in a better state for HIP ?

 

[Deleted member 2197]

RDNA2 For Workstation: AMD Radeon Pro W6600 & W6800 Review – Techgage
November 23, 2021

AMD has so far released two Radeon Pro models built around its RDNA2 architecture – W6600 and W6800 – and compared to the previous-gen, there’s a lot of improvement. In addition to faster performance in general, we see additions like hardware RT, as well as support for resizable BAR and variable rate shading. Let’s see how both cards fare in our gauntlet of tests.

 

[Man from Atlantis]

https://twitter.com/x/status/1463893815991975936