3 ms·

When people say "single-threaded performance" they don't mean literally one single thread will ever be used. The single-core performance is a proxy for perform

by paulmd 7y ago

When people say "single-threaded performance" they don't mean literally one single thread will ever be used. The single-core performance is a proxy for performance-per-core and that's a certainly a very important number. Intel's per-core performance is decently higher in gaming, between their lower latency giving them slightly higher IPC in gaming tasks[1] (as these are latency sensitive) and the much higher all-core clocks they are pushing (at least 10% higher than AMD at present). That still gives them an edge in gaming, although certainly much less of an edge against Zen2 than it was against Zen1 (which was much worse than people were willing to admit at the time, people at the time were benching it with a GTX 1080 (the fastest card available at the time) at 1440p and 4K and claiming there was "no noticeable difference", but faster GPUs showed the weakness). I think the 10700F looks good, 4.6 GHz all-core 8C16T part for under $300 looks compelling for gaming customers if you can keep it cool. The 10900K fills the "I want the top gaming performance first and foremost, with as many cores as I can get, and I'm not super concerned about budget" but is definitely on the expensive side. Other than that, nothing terribly impressive here IMO, the other chips fall into a range between "not fast enough" and "not cheap enough". [1] https://www.techspot.com/article/1876-4ghz-ryzen-3rd-gen-vs-core-i9/ https://www.techspot.com/article/1876-4ghz-ryzen-3rd-gen-vs-...

zozbot234 7y ago
> The single-core performance is a proxy for performance-per-core A very bad proxy then, since performance-per-core with multiple cores in use will almost certainly be limited by constraints on (1) memory bandwidth and (2) power/thermals. "Single-core performance" only applies when you're literally pegging one core at 100% cpu.
- paulmd 7y ago
  Well, all I can say is that enthusiasts expend a lot of effort getting per-core performance as close to single-core as they can ;) Pure single-core turbo is usually a bit higher than all-core turbo, but as a proxy it's surprisingly close. A 9400F (so, no hyperthreading, we are removing that variable) scores 2380 in CB R20 (AVX2 aware). As a 6C, that means its per-core performance is 397. Since single-core is 4.1 and all-core is 3.9, if we adjust the per-core by that amount, we come up with a per-core equivalent of 417. The actual single-core score is 430, so using single-core as a proxy introduces less than 3% expected error to true "per-core" performance. That is the impact of thermals, more cache available to that thread, etc - all of that adds up to less than 3% difference, in a full AVX2 heavy task. AMD boost clocks are not as neatly defined as Intel but the 3500X has let's say 4.1 single-thread boost (the advertised single-thread boosts tend to be "ephemeral" at best under actual loads) and 3.9 all-core. It scores 2650 all-core, that's 441 per-core, adjusted for speed that's a per-core equivalent of 464. The actual single-core score is 467, so basically dead-nuts on. https://www.tomshardware.com/news/amd-ryzen-5-3500x-review-expreview,40605.html https://www.tomshardware.com/news/amd-ryzen-5-3500x-review-e... AMD people just don't like "single-core" because it's not a metric they currently win in. In practice it correlates pretty damn closely to expected per-core performance, and that is a critical metric that has a world of performance impact on a variety of tasks (Amdahl's law says many real-world tasks are dominated by one bottlenecked thread, the faster you run that one single thread, the faster the whole thing runs). Once we see AMD move into the open lead in various workloads (which is very possible with Zen3), overnight we're going to see a lot of people suddenly rediscover the merits of empirical testing under non-bottlenecked conditions. The arguments about "well you would need a RTX 3070 running at 1080p" are going to fade away, just like the whining about the irrelevance of "single-core". ---- Generally speaking most home tasks are not memory-bandwidth bound. It depends, but generally they're not. As far as power/thermals, it depends on the board, if you're running a really crappy board with a weak VRM then all bets are off, but in practice it is not a problem with normal workloads. Gaming is not at all intense and can clock high even on some pretty mediocre boards if you've got some airflow. Something like video encoding (lots of AVX2) is the next step more intense, if you are going to sit there all day crushing video and you won't have airflow then yeah, you do need to buy at least something in the $170-200 tier. Prime95 is not a realistic real world workload and will definitely overheat anything.
gnulinux 7y ago
They mean game FPS will be bottlenecked by single CPU's speed. E.g. in factorio all logic is single-threaded, so you'll be bottlenecked by single CPU, even though it uses other CPUs for minor tasks.

by paulmd 7y ago

zozbot234 7y ago

> The single-core performance is a proxy for performance-per-core A very bad proxy then, since performance-per-core with multiple cores in use will almost certainly be limited by constraints on (1) memory bandwidth and (2) power/thermals. "Single-core performance" only applies when you're literally pegging one core at 100% cpu.

paulmd 7y ago
Well, all I can say is that enthusiasts expend a lot of effort getting per-core performance as close to single-core as they can ;) Pure single-core turbo is usually a bit higher than all-core turbo, but as a proxy it's surprisingly close. A 9400F (so, no hyperthreading, we are removing that variable) scores 2380 in CB R20 (AVX2 aware). As a 6C, that means its per-core performance is 397. Since single-core is 4.1 and all-core is 3.9, if we adjust the per-core by that amount, we come up with a per-core equivalent of 417. The actual single-core score is 430, so using single-core as a proxy introduces less than 3% expected error to true "per-core" performance. That is the impact of thermals, more cache available to that thread, etc - all of that adds up to less than 3% difference, in a full AVX2 heavy task. AMD boost clocks are not as neatly defined as Intel but the 3500X has let's say 4.1 single-thread boost (the advertised single-thread boosts tend to be "ephemeral" at best under actual loads) and 3.9 all-core. It scores 2650 all-core, that's 441 per-core, adjusted for speed that's a per-core equivalent of 464. The actual single-core score is 467, so basically dead-nuts on. https://www.tomshardware.com/news/amd-ryzen-5-3500x-review-expreview,40605.html https://www.tomshardware.com/news/amd-ryzen-5-3500x-review-e... AMD people just don't like "single-core" because it's not a metric they currently win in. In practice it correlates pretty damn closely to expected per-core performance, and that is a critical metric that has a world of performance impact on a variety of tasks (Amdahl's law says many real-world tasks are dominated by one bottlenecked thread, the faster you run that one single thread, the faster the whole thing runs). Once we see AMD move into the open lead in various workloads (which is very possible with Zen3), overnight we're going to see a lot of people suddenly rediscover the merits of empirical testing under non-bottlenecked conditions. The arguments about "well you would need a RTX 3070 running at 1080p" are going to fade away, just like the whining about the irrelevance of "single-core". ---- Generally speaking most home tasks are not memory-bandwidth bound. It depends, but generally they're not. As far as power/thermals, it depends on the board, if you're running a really crappy board with a weak VRM then all bets are off, but in practice it is not a problem with normal workloads. Gaming is not at all intense and can clock high even on some pretty mediocre boards if you've got some airflow. Something like video encoding (lots of AVX2) is the next step more intense, if you are going to sit there all day crushing video and you won't have airflow then yeah, you do need to buy at least something in the $170-200 tier. Prime95 is not a realistic real world workload and will definitely overheat anything.

gnulinux 7y ago

They mean game FPS will be bottlenecked by single CPU's speed. E.g. in factorio all logic is single-threaded, so you'll be bottlenecked by single CPU, even though it uses other CPUs for minor tasks.