Hacker News Reader

Comment by seminatl | Hacker News Reader

I don't really get how they reached their conclusion. It seems like on most of the tests this new part gets beat by a cheaper one from Intel. It seems like a kinda unfair approach to use handbrake without AVX-512 support? Also not sure why they include the 3-d particle thing without AVX ... I guess because Intel is just too fast on that?

If you look through the results, the things most people want to do with a computer, like browse the web and start their applications, are noticeably faster with the Intel i9-9900K at half the price. And the only game where the CPU makes a difference in these benchmarks is also a lot faster with the Intel part.

View on HN

You could just buy a CPU for a third or fourth of the price if you want to just browse. You don't need a $750 16-core CPU. You can buy a $200 9600KF or a $240 3600x or something even cheaper. For anything multi-core the 3950X blows the 9900k out of the water, and for 1440p gaming the 3950x trades blows with the 9900k. If you want 1080p gaming, or web-browsing, this is not the CPU for you. That said, even for 1080p gaming and single-threaded benchmarks the 3950 is usually within a few percent of the 9900K.

Ok but if it’s “trading blows with” and “within a few percent of” a CPU costing half the money, why is it rational to prefer it?

I know why I want AMD parts and that is because of ECC memory support but this article doesn’t even mention it.

First of all, the 9900k doesn't cost half the money. Second, you can also buy a 3900x if you want 12 threads with similar ST characteristics and pricing closer to the 9900k. There are other options.

If you want to game @ 1080p and do mostly single threaded work (without parallelism), this CPU is not for you. If do a lot of single-threaded jobs in parallel (for example, gaming + being a server or gaming and streaming) or anything multi-threaded, this CPU destroys the 9900k, while being just a bit weaker when doing just one or two single-threaded tasks.

Look I'm not arguing about which CPU is better. I'm arguing that taking the geometric mean of a bunch of completely unrelated benchmarks doesn't make any sense. There are clearly workloads in this very article where a cheaper part is better, sometimes even a cheaper AMD part.

And the article states clearly: "[the Core i9-9900KS] does pull out ahead in a number of ST tests as well as in low resolution (CPU-bound) gaming."

Regarding the geomean, they clearly state that their benchmarks prioritize multi-core benchmarks: "This metric also puts the 3900X above the 9900KS, because despite the 5.0 GHz all-core on 8-cores, moving to 12-core and 16-core at almost the same performance per core gives more of an advantage in our test suite's MT-heavy workloads."

They're not hiding behind anything. Everything is clearly stated.

avx2 performance between the new amd cpus and intel is comparable. ive done some detailed comparisons with my own code that is tuned for intel and the ryzen is quicker in some places, slower in others.

AVX-512 is not AVX2.

I'm pretty well versed in the SIMD instruction sets, thank you. You might surmise that from the fact that I'm writing and benchmarking SIMD code....

I was responding to:

"Also not sure why they include the 3-d particle thing without AVX"

in which the op was not clear which AVX instruction set they were referring to. It seemed to be a difference benchmark from the one which had avx-512 disabled, due to the word "also"

So aside from your irrelevant comments about AVX2 being about even between the two parts, what is a reasonable explanation for the second chart on page 4 of the article, assuming you can bother yourself to go read it. Their analysis seems to be essentially that if the Intel part wasn't 300% faster, then the AMD part would be 15% faster. But it is a fact that the Intel part is really three times faster on this benchmark, because of AVX-512. The only reason to include the chart with AVX disabled is to fit the article into a prejudiced narrative.

AVX-512 severely downclocks all cores of a processor.

If I'm using some of my cores to run AVX-512 code in the background, it's going to drastically slow down everything else I'm doing.

This can be a significant issue even if you are only running the one task. If you have intermittent AVX-512 code, any time any one of the cores starts running that code, everything else also slows down too which can result in a net loss in performance. Even if you only ran a single core and it used AVX-512, this could still be the case because it takes precious time before and after to lower and raise the clocks that much and everything else runs slower that whole time in addition to running slowly while the AVX stuff executes.

None of your statements are correct.

1) Intel has per-core frequency and voltage control. To the extent that AVX-512 inhibits higher clock speeds that effect is local to the core or cores on which AVX-512 is active, and only those cores.

2) Not all AVX-512 instructions have this effect. `vpermb` for example does not.

3) Some other AVX instructions that are not AVX-512 have this effect. Also anything else that makes a core hot has this effect.

4) Most importantly, Cannon Lake microarch and later don't have this effect.

Reply on news.ycombinator.com