- Apr 18, 2021
- 1
- 1
- 15
Great piece, but I find it remarkable that you give no attention to the difference of L3 cache size. As usual with comparing CPUs with vs. without iGPU, the one with iGPU makes a significant tradeoff by losing a lot of L3 cache so the graphics blocks can be fitted in the same die space. This is also true for the recently-announced Ryzen 5000G APU series which add Vega graphics at the cost of halving the L3 from 32MB to 16MB (and that's still better than the 11600K's 12MB).
In theory, a smaller L3 should make a difference for applications that are both memory-intensive and multithreaded (single-threaded apps will hardly be limited by DRAM access unless they are poorly written). I guess your set of benchmarks doesn't have much code of that kind. This is more often relevant for server applications that also have either a lot of RAM and/or 10G+ ethernet: database servers with huge caches, webservers that handle thousands of requests/s. These CPUs are not targeted at servers, but some desktop/consumer applications will also benefit a lot from a large L3. For example software development tools, my IDEs and build daemons typically use 10-20GB heaps and they are all written in Java which has the problem that Garbage Collection frequently scanning the whole damn thing with bad locality (~random access patterns) and with stringent latency limits. And GC doesn't do any significant work other than scanning and copying data, so performance is close to 1:1 to memory access. Benchmarks like Cinebench or POV-ray will not care much about cache since they are CPU-bound even if they use large datasets. For common users, web browsers with their multi-gigabyte heaps and complex Javascript apps are another likely example of memory-limited app once pageload ends and the poor thing has to handle your 187 open tabs.
If you're on the market for a CPU and you plan to pair it to a top-class dedicated GPU but you can't afford the latter right now, I wouldn't buy an APU, I'd rather get the best CPU that fits the budget, and I mean one that has a full-size L3 cache for its class (core/thread count), and get some entry-level/mediocre GPU (which will probably perform no worse than the best iGPUs). After the supply crisis is over and you can buy that screaming GPU, ebay the temp GPU if it's still worth anything. Do not risk the buyer's remorse of eventually being stuck with a system that has only 37% of the L3 it should have, or even 50% like the 5000G.
In theory, a smaller L3 should make a difference for applications that are both memory-intensive and multithreaded (single-threaded apps will hardly be limited by DRAM access unless they are poorly written). I guess your set of benchmarks doesn't have much code of that kind. This is more often relevant for server applications that also have either a lot of RAM and/or 10G+ ethernet: database servers with huge caches, webservers that handle thousands of requests/s. These CPUs are not targeted at servers, but some desktop/consumer applications will also benefit a lot from a large L3. For example software development tools, my IDEs and build daemons typically use 10-20GB heaps and they are all written in Java which has the problem that Garbage Collection frequently scanning the whole damn thing with bad locality (~random access patterns) and with stringent latency limits. And GC doesn't do any significant work other than scanning and copying data, so performance is close to 1:1 to memory access. Benchmarks like Cinebench or POV-ray will not care much about cache since they are CPU-bound even if they use large datasets. For common users, web browsers with their multi-gigabyte heaps and complex Javascript apps are another likely example of memory-limited app once pageload ends and the poor thing has to handle your 187 open tabs.
If you're on the market for a CPU and you plan to pair it to a top-class dedicated GPU but you can't afford the latter right now, I wouldn't buy an APU, I'd rather get the best CPU that fits the budget, and I mean one that has a full-size L3 cache for its class (core/thread count), and get some entry-level/mediocre GPU (which will probably perform no worse than the best iGPUs). After the supply crisis is over and you can buy that screaming GPU, ebay the temp GPU if it's still worth anything. Do not risk the buyer's remorse of eventually being stuck with a system that has only 37% of the L3 it should have, or even 50% like the 5000G.
Facebook
Twitter
Instagram