So when AMD card, with unoptimized drivers, and in unoptimized for it software performs better than Nvidia card with optimized software and drivers we call that AMD GPUs are worse. Good to know.
Back to topic. OTOY claims that the GPUs are considered CUDA devices. If they have tested it on M395X I think it will behave like GTX 980, because both GPUs have similar core counts, and I do not think OTOY did anything to change Register File handling in CUDA from 128 Cores, to 64 cores. That would create huge change for AMD GPUs.
How come? Throughput would increase twice in the execution pipeline. There are other features that also affect performance(store and load width of instructions, etc), so in the end, we would see 40-50% performance increase clock for clock.
Let me explain how GPU architectures work, again.
Nvidia with Kepler Have designed GPU arch to have 192 CUDA Cores per 256 KB Register File Size.
With Maxwell they have jumped from 192 Cores, to 128 cores per 256 KB Register File size.
At first glance, it looks like a downgrade right?
No. What this has done is it increased throughput of the cores, because they were less starved for resources.
When we are looking at gaming and the idea of Tile Based Rasterization people thought that TBR is increasing efficiency, but actually TBR has not increased the throughput. It increased the memory balancing, and its load, and culling capabilities. This increased Geometry performance, but was topped out by throughput of the cores, which were able to do more work each cycle - increasing efficiency.
This is why GTX 980 with 2048 CUDA Cores, and roughly 4.4 TFLOPs was able to outperform 5.5 TFLOPs GTX 780 Ti, that had 2880 CUDA Cores.
Consumer Pascal GPUs have not seen per clock increase in performance, like Maxwell vs Kepler did, because they still share the same Maxwell 128 Core/256 KB RFS architecture. GP100 is proper 64 Core/256 KB RFS Architecture, so that is why we have seen per clock, and per core increase in performance over previous generation GPUs.
How does it look at AMD GPUs? Since GCN1 they were always 64 Cores/256 KB Register File Size in each Compute Unit. Actually I am just looking at GCN white paper and Registry File Size in GCN is flexible. It can vary from 128 KB to 256 KB. That is why they have had stronger throughput in even not optimized software, for this architecture. That is why AMD GCN architecture excels in Apple ecosystem, because not only it is faster per clock(but lower clocked this way), but also wider, and able to do more work, if your software is properly optimized.
What this means is: Even if OTOY will not optimize it specifically for GCN(they cannot, because of obvious reasons), the GPUs will not be worse than their direct competitors in core counts, at least from Core throughput perspective. There can be features that GCN does not have, and there is obviously the topic of the pipeline. Nvidia GPUs have longer pipeline hence is why they are able to clock higher, at lower power consumption. GCN have had always shorter pipeline, because it was able to do more work each cycle. The tradeoff for this is that it could never clock up high, and was using more power, especially in non-optimized software.
In general we both agree about the topic of optimization. But its way less work to do right now, with HIP, that it would be without it. It will be interesting to see the performance differences on GCN versus CUDA running CUDA code, but I don't expect huge difference in performance between similarly core counted GPUs. If we compare for example M395X, it will behave just like 2048 core clock Maxwell or consumer Pascal GPU, unless performance relies only on RFS, and not related to core count. Then AMD GPU should be much faster.
.
