Post your Ryzen Blender Demo Scores! (AMD clarifies Blender Benchmark Confusion, Run @ 150 Samples)

[Shivansps]

Is not rare to include new features or optimizations to a B version? I whould have called that 2.79.

 

[itsmydamnation]

Is not rare to include new features or optimizations to a B version? I whould have called that 2.79.

https://lists.blender.org/pipermail/bf-committers/2017-January/048000.html

i guess you could call it an unplanned release.

 

[nvgpu]

https://www.youtube.com/watch?v=2XNMYV_1pAw

38.58 seconds on 2.78a
25.66 seconds on 2.78b

Yeah, all 2.78a benchmarks are invalid and should be ignored, they cripple Intel's real performance.

 

[itsmydamnation]

Yeah, all 2.78a benchmarks are invalid and should be ignored, they cripple Intel's real performance.

Not sure if cereal?

Spoiler: hint

CON cores gain more from b the intel cores do ( os much it looks like they become 100% FPU bottlenecked) and the Stilt has said then generally speaking Zen gains more from using ICC with AVX/2/FMA options then intel cores do . This is going to be a prime Zen benchmark with B because it has 4x128bit SSE/AVX/AVX2 pipes vs intels 2

 

[jpiniero]

You'd think AMD would have shown Blender running Linux if that was the case though. I guess we won't really know the full impact of 4x128 vs 2x256 until the benchmarks come out.

 

[itsmydamnation]

You'd think AMD would have shown Blender running Linux if that was the case though. I guess we won't really know the full impact of 4x128 vs 2x256 until the benchmarks come out.

Look at the workload..... tell me what part of rendering uses either FP64 or 8 packed FP32. Very little because we are dealing with data structures aligned to 128bits. AVX2 doesn't have scatter+gather the compiler just can't grab any data from anywhere and smash it into a single operation.

This is different to say encoding where there is no long lived internal data structure your reading in data serially and generating key + delta/incremental frames where more throughput over less pipelines can shine.

 

[HilbertSpace]

For future reference:


Blender 2.78c with 150 samples

MacBook Pro (15-inch, 2016) - 45.31 Seconds
Intel i7-6700HQ (2.6-3.5 GHz) - 8 Threads
macOS 10.12.4


ASUS Z270i - ~32 Seconds
Intel i7-7700K (5.0 GHz OC) - 8 Threads
Win 10 x64 (Creator's update)


Dell 7910 - 10.24 Seconds
Intel Xeon E5-2695 v4 x2 (2.1-3.3 GHz) - 72 Threads
Ubuntu 14.04 LTS

 

[jhu]

Now that OpenCL and CUDA are at feature parity, here's my reference RX480 for comparison:

It's about 23 seconds for a $200 video card. For Blender Cycles rendering, GPU is the way to go.

 

[NesuD]

Ryzen 1700x @ 3.8 ghz 150 sample setting