Source
- URL: https://x.com/SebAaltonen/status/1611409453806403586
- Author: Sebastian Aaltonen (@SebAaltonen)
- Posted: 2023-01-06 17:08:46
Thread
1/ @SebAaltonen
Thread: A simple explanation why AMDs 144MB 3D cache makes games run significantly faster.
TLDR: The working set of most games fits inside the 3D cache. All cache lines accessed during the previous frame are still present in cache. Sequential frames only differ slightly.
2/ @SebAaltonen
If you look at any recent GPU benchmarks (RTX 4090), you notice that most games run at 240Hz+ if you remove the GPU bottleneck. Run at 1080p or similar without RTX. This is enough info to calculate how much memory we can access per frame.
3/ @SebAaltonen
DDR4-3200 provides us 50GB/s theoretical bandwidth. In reality we get roughly 40GB/s, which is proven by memory benchmarks. Divide this by 240Hz, and we get 170MB/frame. That’s how much we can access unique memory cache lines per frame in a game that runs at 240Hz = most games.
4/ @SebAaltonen
AMDs 128MB 3D cache is a victim cache, so it doesn’t have the same pages as the smaller cache levels. Thus we can use the 144MB number as our maximum working set limit. 170MB/frame considers that game is 100% mem bound all the time. In reality it’s not, thus 144MB is enough.
5/ @SebAaltonen
And this is why AMD 3D cache provides big gains for games. The CPU working set (data accessed in a single frame) fits to the 144MB cache, meaning that the next frame can read most of the data from the cache. The difference between two sequential frames is minimal.
6/ @SebAaltonen
Correction: pages -> cache lines. Victim cache doesn’t keep the same cache lines as lower level caches.
7/ @SebAaltonen
Traditional MT workloads such as compiling code doesn’t get as big benefits. Each .cpp file is compiled in a separate process. There’s zero data sharing between them. If you have a cold start (new process) for every compiled file, there’s not much a massive cache can do to help.
Branches: 2023-01-06-hkultala-caches-work-with-physical-addresses-all-the-code, 2023-01-06-isotoxin390-seb-for-the-next-gen-consoles-what-is-your-amd, 2023-01-07-morbidarne2007-well-this-is-about-to-change-in-llvm-at-least, 2023-01-07-Kodiak_73-compiling-translation-units-cpp-files-dont, 2023-01-08-GPUsAreMagic-very-nicely-put-together-but-one-big-omission-imo