Aaron Vegh
Human Brain EnthusiastIs there single/multicore benchmark for CPUs that makes current-generation CPUs comparable with stuff from say the mid-90s onwards?
I'd like to specifically compare how fast a G4 (e.g. PowerPC 7455) is versus current Apple M5.
(I get that this is a somewhat hard problem because of various fuckeries that benchmarks and compilers do.)
Empterdose@thomasfuchs It looks like a few people have submitted Geekbench 2 scores from M4 Mac minis. https://browser.geekbench.com/geekbench2/search?q=PowerMac10%2C2 vs. https://browser.geekbench.com/geekbench2/search?q=Mac16%2C10 would seem to show that a 10-core M4 is ~30 times faster than a 1.5 GHz G4. That sounds impressive at first blush, but that would mean that it's only three times faster per core, which is obviously wrong. Rosetta 2 would account for some of that, but not all of it.
Raven667@MrDOS @thomasfuchs i dont have numbers but what is your expectation of single thread performance difference over 20y? I think all the improvements have been in adding cores, and splitting work, afaik single-core perf is only incrementally better, maybe 5-10% per major generation bump, or less, so 3x faster doesnt sound entirely crazy to me.
Edit, i see some better estimates downthread that suggest 10-15x is probably a better ballpark for M-series vs G3/G4
@raven667 @thomasfuchs My gut feel was also 10-15x.
Trying to make an apples-to-apples comparison over the same timeline, an AMD Athlon X2 4200+ from August 2005 gets a single-threaded PassMark CPU rating of 803. Just shy of twenty years on, a Ryzen 9 9950X3D from early 2025 gets 4742 – nearly 6x faster. Not sure how you'd count the generations between those two parts, but if we say there's a generation every ~two years, that's more like a 15% generation-on-generation improvement, which is pretty incredible.
@raven667 @thomasfuchs (Looking at a 4200+ here rather a 4400+, because the PassMark sample size is too low. And comparing with the 9950X3D rather than an Apple M part because I don't see an M5 ranking yet, and the M4 is more like 18 years newer than the 4200+, and I didn't want to have to figure out the corresponding top-performing CPU from 2004.)
@raven667 @thomasfuchs Aside from the multi-core scaling, I/O has also gotten so much faster in the last 20 years, and I think that contributes as much to the “feel” of speed as anything. My modest NVMe drive from five years ago can sustain ~1.2 GB/s continuous serial read. What would spinning rust do in 2005, 80 MB/s? There's a 15x multiple right there.
@raven667 @thomasfuchs Socket 939 was used with first-gen DDR, which topped out at 3.2 GB/s transfer. DDR5 _starts_ at 32 GB/s, and the sort of high-end/enthusiast system we're talking about could see double that. So there's another 10-20x multiple.
@raven667 @thomasfuchs Back to the original point, I still suspect in a fair fight, an M5 is still about 10x a G4, but probably not much more than that: I was definitely overestimating the scale factor.
If anyone has an empirical test to try, I have a 1.25 GHz G4 Mac mini close to hand that I'd be happy to run benchmarks on.
Fabian Transchel@thomasfuchs I found 7zip* to be a reasonably time-stable benchmark as it runs on almost any architecture and almost any CPU.
Of course, it doesn't really check anything besides (un)zipping things...
* e.g. 7z b -mmt1
@thomasfuchs Incidentally, I use it to "prove" that running #gentoo -march=native optimizations is worth about 15% performance-wise on workloads I never do 😂
Markus Osterhoff
Arakin@thomasfuchs someone ran a custom single threaded Dhrystone benchmark against various computers he used since 1976 - not sure if this is of interest? It isn't optimised and so gains from this and multi threading would be much better these days.
(Updated with a better source which includes a list of results and link to the GitHub repo for the benchmark)
https://hothardware.com/news/dhrystone-benchmarks-1976-today
☮ ♥ ♬ 🧑💻
David Smith@thomasfuchs @joe really janky way of estimating, but: the Cell PPE/Xenon cores (same uarch) were “essentially” dual G4s at half the clock frequency, since they were in-order and alternated two threads.
So if you can find a usable single threaded comparison between a modern system and the Xbox 360 or PS3, that’ll get you a hypothetical “1.6GHz G4 with less anemic frontside bus”.
Then you could try to estimate the remaining scaling by comparing G4s and extrapolating.
@thomasfuchs @joe as a calibration: the Switch 1 gets ~300-350 on geekbench 6, the M5 gets ~4300, and the Switch was a noticeable upgrade from the Wii U’s “1.24GHz G3 with weird non-AltiVec vectors” (should be vaguely G4-like), so any numbers you get showing less than 15x or so are probably suspect.
Alex Pretzlav 🥨@thomasfuchs back in 2021 I tested the time to compile a specific version of git on a PowerBook G4 and an M1 Pro iMac. The M1 was 37 times faster.
@pretz yeah that tracks