• Sashleycat

SCV: Testing 2950X, 5800X and 1660S, 2080 Ti in Bricklink Stud.io photorealistic renderer.

You may remember a while back I posted on my blog saying that Stud.io, Bricklinks Lego CAD software's built in photorealistic rendering software (called "Eyesight Renderer") doesn't work that great on Nvidia's Turing cards? As in, not as fast as one might expect - especially vs the CPU. Well, in this video I made a test to kinda show that. But also, since I don't have any non-Turing NVIDIA GPUs (the GTX 770 died, it will n̶o̶t̶ be missed. Ahem), I can't show an apples to apples comparison between them.


That said, you can see the 2950X crushing the GTX 1660 SUPER, and the RTX 2080 Ti isn't that much faster (though it is the fastest of the bunch, I should also hope so, the Chungus GPU). By the way, the renderer doesn't use the RT cores, obviously. Just CUDA.


Also in this video, I cried a bit because the stock 8-core 5800X almost freakin' matches my 16-core 2950X at 4 GHz all-core locked, in a test that can use all 32 threads on the 2950X. MY PRECIOUS ZEN+, YOU CAN'T BE THAT OLD ALREADY, RIGHT?


Uh, no. Couple of pointers here. Firstly, Ray Trace rendering is pretty sensitive to latency and branch handling. In fact, probably quite a bit sensitive to latency. Not only does the 5800X have much higher clock speeds and thus lower L1,2 (and even L3, despite more cycles needed per request) latency; but it also has the same amount of L3 cache despite having half of the cores. Uh, on top of that, my 2950X's memory is eye-wateringly slow with DDR4-2666 C16-18-18-36 in UMA (distributed mode). That's 110ns hard latency (Measured with AIDA). Quad Channel with eight Dual Rank DIMMs though! I heard you liked memory parallelism! .. Of which the Zen3 IMC is better ay anyway! >_>.


Sash, why? WHY SO SLOW?


Cuz 128 GB and 2DPC, Dual Rank is like putting an overgrown marrow in the IMC's anus. That is, assuming the IMC's anus is roughly human-sized. So yeah, it's not going to be fast. Also the second kit is rated for 2666 MHz, and I need UMA mode for certain workloads that tank with Local memory access. So you get latency. Lots of it. But I don't need to justify myself to you! Nuh uh.


Just keep in mind the RAM on the 2950X is far from the highest performance. Incidentally, L3 hit rate in Studio Renderer is 70-80% on 2950X so it is missing L3 a bit, so RAM is being hit here somewhat. Anyway, I digressed again.


Probably should share the video now. Here it is.