3 ms·

I have been incredibly skeptical of frequency numbers. Case in point, the AMD FX-9590 had a 4.7 GHz base an 5.0 GHz boost back in 2013 [0]; and the Ryzen 9 5950

I have been incredibly skeptical of frequency numbers. Case in point, the AMD FX-9590 had a 4.7 GHz base an 5.0 GHz boost back in 2013 [0]; and the Ryzen 9 5950x has a 3.4 GHz base an 4.9 GHz boost [1]. To me, the frequency has become a "blob". It means nothing. If anything, I see more frequency and I think more power draw and less stability. Real performance is a function of IPC and frequency and that is the only thing that really matters. [0]: https://www.amd.com/es/products/cpu/fx-9590 https://www.amd.com/es/products/cpu/fx-9590 [1]: https://www.amd.com/en/products/cpu/amd-ryzen-9-5950x https://www.amd.com/en/products/cpu/amd-ryzen-9-5950x

leetcrew 6y ago
none of those numbers fully describe the performance of a CPU, but idk why you would be "skeptical" of frequency numbers in particular. IPC itself is somewhat arbitrary. in the real world, IPC constantly varies with the mix of instructions, cache-friendliness of the program/data, memory performance, etc.
- josalhor 6y ago
  > none of those numbers fully describe the performance of a CPU, but idk why you would be "skeptical" of frequency numbers in particular Because I have seen quite a few consumers get guided too much by the "more frequency, more better" motto. Moreover, it doesn't stop there. I have seen papers where the only thing cited is the frequency of the processors, and no other reference to the manufacturer/architecture/version of the processors used. > IPC itself is somewhat arbitrary. in the real world, IPC constantly varies with the mix of instructions, cache-friendliness of the program/data, memory performance, etc. Absolutely, I would also be skeptical of an "IPC" metric. Performance is the only thing that counts.
tenacious_tuna 6y ago
> To me, the frequency has become a "blob". It means nothing. That seems reasonable across microarch versions, but presumably is totally reasonable within a product-release-cycle? i.e. to compare the FX-9590 to an FX-8350 based off clock speed. If it's the same core count and microarch, presumably clock is about the only difference, aside from other features like PCIE lanes, cache, etc.
- josalhor 6y ago
  > That seems reasonable across microarch versions, but presumably is totally reasonable within a product-release-cycle Sure, but the performance doesn't scale linearly with the frequency. So how can we account for that? Also, the i9 10900F and i9-10850K have the same turbo. How long can they sustain in for? On how many cores? What frequency is left afterward? This gets even weirder with more recent technologies that push the frequency higher depending on power delivery stability and thermal constraints. I agree; the frequency is indicative of performance within a product lifecycle. But I find it too meaningless to give an approximate delta on the performance. And this assumes the process you are running is not too much memory bottlenecked or I/O bound and this delta in performance can be measured. (And the setup can sustain higher thermals). While not perfect, I prefer to look up benchmarks that resemble my workload and call it a day. Processors are too complicated for me to try to compare with simple numbers.
  - BuckRogers 6y ago
    Well said. That’s exactly your I feel about it at this point. The thermal and power properties and how they interact with newer boost algorithms is unpredictable. I try to find my workload benchmarks and judge off that. For me that would be VS but any compilation comparison suffices.

leetcrew 6y ago

none of those numbers fully describe the performance of a CPU, but idk why you would be "skeptical" of frequency numbers in particular. IPC itself is somewhat arbitrary. in the real world, IPC constantly varies with the mix of instructions, cache-friendliness of the program/data, memory performance, etc.

josalhor 6y ago
> none of those numbers fully describe the performance of a CPU, but idk why you would be "skeptical" of frequency numbers in particular Because I have seen quite a few consumers get guided too much by the "more frequency, more better" motto. Moreover, it doesn't stop there. I have seen papers where the only thing cited is the frequency of the processors, and no other reference to the manufacturer/architecture/version of the processors used. > IPC itself is somewhat arbitrary. in the real world, IPC constantly varies with the mix of instructions, cache-friendliness of the program/data, memory performance, etc. Absolutely, I would also be skeptical of an "IPC" metric. Performance is the only thing that counts.

tenacious_tuna 6y ago

> To me, the frequency has become a "blob". It means nothing. That seems reasonable across microarch versions, but presumably is totally reasonable within a product-release-cycle? i.e. to compare the FX-9590 to an FX-8350 based off clock speed. If it's the same core count and microarch, presumably clock is about the only difference, aside from other features like PCIE lanes, cache, etc.

josalhor 6y ago
> That seems reasonable across microarch versions, but presumably is totally reasonable within a product-release-cycle Sure, but the performance doesn't scale linearly with the frequency. So how can we account for that? Also, the i9 10900F and i9-10850K have the same turbo. How long can they sustain in for? On how many cores? What frequency is left afterward? This gets even weirder with more recent technologies that push the frequency higher depending on power delivery stability and thermal constraints. I agree; the frequency is indicative of performance within a product lifecycle. But I find it too meaningless to give an approximate delta on the performance. And this assumes the process you are running is not too much memory bottlenecked or I/O bound and this delta in performance can be measured. (And the setup can sustain higher thermals). While not perfect, I prefer to look up benchmarks that resemble my workload and call it a day. Processors are too complicated for me to try to compare with simple numbers.
- BuckRogers 6y ago
  Well said. That’s exactly your I feel about it at this point. The thermal and power properties and how they interact with newer boost algorithms is unpredictable. I try to find my workload benchmarks and judge off that. For me that would be VS but any compilation comparison suffices.