Amin Aliari
LightNews LightNews
amin-3d.bsky.social
Amin Aliari
@amin-3d.bsky.social
8 followers 150 following 0 posts
3D Programmer @ Ubisoft Toronto
Posts Replies Media Videos
Reposted by Amin Aliari
vassvik.bsky.social
Let's wrap up this lovely week with a nice technical post

This is the "case study" from my Masterclass at GPC, where I apply a series of optimizations to improve the effective bandwidth of a 3x3x3 blur (a proxy for a huge set of operations on volumetric data)

Check ALT text for (a lot of) context.
A table showing the experimental results of applying 6 different compute shader versions of a simple 3x3x3 box blur on a 512x512x512 texture using either GL_R16F or GL_R32F internal format for storage for a eight different GPUs spanning several GPU architectures and vendors. The upper table shows the absolute effective bandwidth (measured as the sum of total bytes read and written divided by execution time), whereas the lower table shows the effective bandwidth relative to the theoretical bandwidth as a percentage. Each row corresponds to a specific shader variant (except for the "theoretical" row, which displays the theoretical bandwidth according to the GPU specification), and each column corresponds to a specific GPU. The color coding is per column in the upper table, and it's a single color coding on the entire lower table. Each version will be explained in detail in the subsequent posts. Version 6 applies uses half precision floating point for the shared memory cache, and the relevant extension does not exist in the Intel drivers for Windows. Likewise this version is not applied to the GL_R32F internal format benchmarks since that would destroy the precision of the backing format anyway. The code was written and initially tested on a desktop 4090 (the first column), which naturally skews the results a bit since everything was evaluated and tested on that GPU. Had I used another GPU I might have picked slightly different compromises, and the results would have been slightly different. One interesting observation is that the RTX 4000 series (Ada Lovelace architecture) significantly overperform everything else, with 7900 XTX (RDNA3) slightly behind. A large part of these overwhelmingly efficient results is due to the massive caches these devices sport (72 MiB on the desktop 4090, 64 MiB on the laptop 4090, etc.), which really helps reach peak bandwidth a lot easier.
November 17, 2024 at 10:58 PM