Hacker News Reader

Comment by xcv123 | Hacker News Reader

Benchmarks (or any timed CPU performance tests) are precisely how we prove how fast CPUs are. No need for quotes on "prove". It's an objective measure, an actual proof. Geekbench 6 is one suite of tests that is reliable and has provided a useful measure for comparing production server VM performance at my job. It accurately predicted the performance of real server work loads.

I would rather use Geekbench or any random benchmarking suite than rely on some guy commenting that he thinks/feels his ancient quad i5 is as fast as the M1 (not even close).

View on HN

> It's an objective measure, an actual proof. Geekbench

> Geekbench 6 is one suite of tests that is reliable and has provided a useful measure for comparing production server

Geekbench is notoriously the least-useful of the lot. The big flashy numbers you share from that test are an aggregate score of almost nothing meaningful or similar being compared. Geekbench is less about hardware analysis and more about offhand comparisons to other systems. Using it to compare ARM and x86 is useless, given that both architectures are more efficient at different workloads. Worse yet, we then "average" those scores to try and give people a realistic performance impression based off wildly different CPU architectures.

In some tests, it is entirely probable that an i5 would smoke an M1. Load up an SIMD benchmark and compare AVX2 to ARM NEON. Even a Haswell CPU will come out on top if you stack the benches right.

> ... In some tests, it is entirely probable that an i5 would smoke an M1

I'm aware that there is no such thing as a perfect benchmark. Geekbench has been useful for me and is good enough to get a general idea of performance.

Check it yourself. Compare i5-4690k (fastest i5 from 2014) with my M1 Pro.

https://browser.geekbench.com/v6/cpu/compare/3356017?baselin...

M1 is faster in every single-core test in Geekbench 6. Overall 195% for single core and 427% for multi-core.

> an aggregate score of almost nothing meaningful or similar being compared

The dissimilarity between tests is intentional, by design, and is precisely what makes it useful as a general benchmark. The tests seem to cover a range of various real world use cases (C compilation, ray tracing, file compression, image processing, web browsing, etc). As I mentioned earlier, it usefully predicted real server performance at my job (JVM server doing financial batch processing).

> The dissimilarity between tests is intentional, by design, and is precisely what makes it useful as a general benchmark.

That's fine, if those numbers mean something to you then don't let me take that from you. My point is more that Geekbench has always been about cooking the books and avoiding objective comparisons. A number of their tests have been shown to instead rely on memory and SSD bandwidth, which kinda makes the whole "CPU score" concept bunk.

Most benchmarks will indeed rank the 2021 RISC CPU above the 2014 CISC CPU. My point is more that Geekbench has always been a dubious point of comparison for hardware differences. The meaning behind those scores is pretty pointless when you actually see how they're weighed and aggregated.

> A number of their tests have been shown to instead rely on memory and SSD bandwidth

These are simulating common CPU workloads, so of course they will be partly constrained by the processors memory controller bandwidth and latency, as in real workloads. The fact that this is the case makes it even more useful as a general processing benchmark. Modern processors include a memory controller in the package. What you see as "cooking the books" is actually the correct and expected approach for this purpose. It should rely on memory bandwidth for some of the tests. The entire Geekbench data set is ~1GB so would be cached by the OS in RAM on any of systems we care about benchmarking.

If you need a specialized benchmark then run a specialized benchmark. Geekbench is not pretending to be something like a LINPACK benchmark.

> It's an objective measure, an actual proof. Geekbench 6 is one suite of tests that is reliable

Did I say it wasn't useful? I said it's only useful if you're trying to measure a single set of metrics versus a broad spectrum. As is clear from the 100+ tests it doesn't compare to a generation old Ryzen architecture at. I also clearly said that it was a useful metric for what many M-series users are actually looking for ("common use cases such as JavaScript execution and visual compositing").

You should really learn to comprehend entire thoughts versus the idea/sentence/strawman you make up in your head.

> You should really learn to comprehend entire thoughts

Both of these replies from you are poorly written, hence you're getting downvoted by others. You should learn to form clearer thoughts that are worth communicating, and work on your writing.

I'm being downvoted for being a smarmy ass. I accept that.

Now feel free to introspectively accept your own faults. Particularly your pigheaded inability to argue in good faith and/or comprehend basic English.

PS - you didn't need two messages for that. All it does is reinforce your insecurity and necessity to "be right".

Next time, just edit your post. Or better yet, post and move on. It's an internet comment from an anonymous person, no need to let it live rent free in your head.

Wow. You have issues.

> PS - you didn't need two messages for that.

I only wrote one message for that.

Anyway, put that aside. The benchmark suite at Phoronix is biased to AMD due to a lot of tests being optimized for x86 instructions (AVX).

> Wow. You have issues.

Sick burn, bro.

> I only wrote one message for that.

You returned to write this later:

https://news.ycombinator.com/item?id=38121689

Because a stupid internet argument was living rent free in your head.

> Anyway, put that aside. The benchmark suite at Phoronix is biased to AMD due to a lot of tests being optimized for x86 instructions (AVX).

Which is a part of the x86 instruction set. As equally as with Apple's own simd instructions:

https://developer.apple.com/documentation/accelerate/simd

But understood, one favors one party so is invalid and the other favors your weird cult, so it's acceptable. I'll keep that in mind in the future for internet "debate": "if the benchmark doesn't favor Apple, it's invalid for <INSERT-SOME-REASON>".

You can move the goalposts or change the argument all you like. I don't really care who wins/loses an internet argument, and I'm quite well versed on both architectures to be able to acknowledge the strengths and weaknesses of each. Which I fully acknowledged multiple times, despite your inability to parse the text of that information.

> Which is a part of the x86 instruction set. As equally as with Apple's own simd instructions:

Yes M1 does not have AVX. That is x86 only. So if the application does not use M1 SIMD then it won't run as quickly. Benchmark results were obvious in many cases. Apple SIMD is simply not 10x slower than x86 SIMD. That is obvious.

Note that the Apple Accelerate API that you linked to is available on MacOS only. The benchmarks were run on Asahi Linux.

> favors your weird cult

LOL. I own an AMD Zen 3 desktop, Intel i5 Surface Pro, Apple M1 laptop. I develop on MacOS, Windows, Linux. Which cult does that put me in? Am I in all three cults?

> As is clear from the 100+ tests it doesn't compare to a generation old Ryzen architecture

I checked the Phoronix test results. Looks like AMD is ahead specifically in DSP / linear algebra tests where the code was designed for AVX instead of NEON. These are massively accelerated by specialized vector processing instructions. The results would be quite different if the code was optimized for both instruction sets.

Reply on news.ycombinator.com