Ashes of the Singularity Receives Ryzen Performance Update - Up to 31% Improvement

Speaker1264 · April 1, 2017

14 minutes ago, MandelFrac said:

Why would you bet that? if you think about the problem solving process and task parallelism, it seems pretty obvious to me you launch co-dependent tasks in pairs, so as long as the chief dependencies are in threads 0-3 and 4-7 with minimal need to send things across that boundary (such as when doing the final data marshalling for the GPU), then actually it's not all that unlikely.

You don't have to plan for the design, just plan for having 4/6/8 core machines. If everything is either a power of 2 or at least has 2 as a factor, then launching dependent threads in pairs is fairly logical (and common actually if you were already optimizing for hyperthreading).

And if your requirements were for Core 2 Quad, you'd actually already be doing the same optimization pattern there as you would for Ryzen, minus SMT/Hyperthreading. So seriously, don't run your mouth when you're clueless. It's one thing to be skeptical, but to say there's no chance is flat out wrong on its face.

So you think it the developers of Ashes of the Singularity, one of the most optimized games out there just happened to be dumb enough to not have thought of this? And after the optimizations the 1800x performed nearly the same as the 6900K, which I really hadn't seen in many benchmarks prior to this, except in ones that were likely GPU bottlenecked, meaning that those games are likely not utilizing Ryzen optimally.

MandelFrac · April 1, 2017

9 minutes ago, Speaker1264 said:

So you think it the developers of Ashes of the Singularity, one of the most optimized games out there just happened to be dumb enough to not have thought of this? And after the optimizations the 1800x performed nearly the same as the 6900K, which I really hadn't seen in many benchmarks prior to this, except in ones that were likely GPU bottlenecked, meaning that those games are likely not running Ryzen optimally.

Ashes was built on a new engine with minimum requirements of Sandy Bridge. If the engine itself was not designed to be dealing with C2Q's idiosyncrasies (legacy code does take forever to take out, just see Unreal 4), then no, it wouldn't surprise me in the slightest. And AOTS is not a paragon of optimization. Until I see an AVX instruction in the disassembly (hint, just look for "ymm"), then there's a near 2x speedup for the CPU-side code available.

Or what it also means is they got the cache friendliness done well enough that Ryzen's double-sized L2 made up for its overall lower IPC just as it did for Cinebench. And before you challenge me on the definition of IPC, throughput, and performance, http://www.agner.org/optimize/instruction_tables.pdf

The results for Ryzen haven't been written up in the PDF, as Agner's been busy, but go ahead and run his test scripts here: http://www.agner.org/optimize/#testp

Several people have already found Ryzen has roughly Ivy Bridge IPC; but it's the 2x sized L2 lowering the # of 42-cycle cache miss penalties --thus keeping the pipeline better fed-- which make up for it and sometimes can even beat Kaby Lake.