Intel: i7-11700 vs. i9-10900 for wavefront P-1 or PRP

[techn1ciaN]

I've been thinking about building a second PC to replace my Blu-ray player. I would obviously run Prime95 on it and have enough spare cash that I can throw in an overpowered processor for that purpose. I have narrowed it down to the two models in the thread subject but am not getting further because it doesn't seem obvious to me which one should pull ahead in primality testing. The 11700 has AVX-512, but at only one operation per cycle. The 11700 is reported to have higher IPC, but the extent of the benefit seems to be task-dependent. The 10900 has two more cores.

Assume a 280 mm water cooler, a decent Z[xxx] chipset motherboard, the fastest dual-channel RAM one can get before the ratio of price paid to speed gained starts getting less reasonable*, and around 10–12 GB allocated to Prime95 in the case of running P-1**. I have access to a Micro Center location and the two models are priced similarly (11700 $290; 10900 $330). I would remove the stock 65 W power limit in either case and let Turbo Boost (or whatever Intel is calling that functionality these days) do its thing, but would probably not tune anything more finely than that.

If the performance difference is a wash or close to it, which should have the lower load power consumption?

All input appreciated. I checked the benchmark thread and that one old "dedicated PC for Prime95" thread but could not see anything relevant.

* I am aware of the benefit of dual-rank DIMMs but have not researched how much more they tend to cost. Input would be welcome on which optimization — dual-rank DIMMs or higher speeds — will tend to give a greater benefit when spending the same amount of money.

** I could look into 32 GB of RAM if the P-1 throughput boost would be significant, but the speed would probably be lower.

[JWNoctis]

Current best seemed to be i7-12700 combined with a motherboard that let you enable AVX-512 in excharge for disabling E-cores, and the fastest DDR5 memory you could get and overclocked to maximum stable frequency, assuming Intel or MB manufacturer doesn't take out that unofficial AVX-512 support in a future microcode or BIOS patch, and that would actually produce reliable result.

But since that's irrelevant to your case, I could say that if an octa-core Zen 3 chip - otherwise apples versus oranges but had AVX2/FMA3 instructions of largely comparable throughput - would start to hit memory bandwidth limit of a dual-channel dual-rank DDR4-3200 setup at just above 2GHz all-core frequency on a wavefront PRP test, an octa- or 10-core Ice Lake/Rocket Lake could probably do something similar, since none of them had nearly enough cache to run those without hitting main memory.

That same setup also achieved some 80% of the performance at ~75W running at 25W and ~2.2GHz, as reported by the processor itself: There's not much to gain after hitting that memory bandwidth bottleneck, except maybe as an impromptu space heater for winter.

As for AVX512 and Rocket Lake, there are some interesting reads.

The only additional cost for dual-rank DRAM is they tended to be larger - With 16Gb chips, the minimum dual-channel dual-rank DDR4 setup would be 64GiB. Actual performance benefit is about 5% in terms of 7-Zip benchmark on mine.

[techn1ciaN]

Originally Posted by JWNoctis

But since that's irrelevant to your case, I could say that if an octa-core Zen 3 chip - otherwise apples versus oranges but had AVX2/FMA3 instructions of largely comparable throughput - would start to hit memory bandwidth limit of a dual-channel dual-rank DDR4-3200 setup at just above 2GHz all-core frequency on a wavefront PRP test, an octa- or 10-core Ice Lake/Rocket Lake could probably do something similar, since none of them had nearly enough cache to run those without hitting main memory.

Reading between the lines, this would seem to mean that an i7 or i9 would not be worth buying at all*, since I could get the same or nearly the same throughput by purchasing a six-core i5 (for example) and letting it run at a higher frequency. I will consider an i5-11400 and do some very bad napkin math.

I'll say that >3 GHz on your chart looks like the point of diminishing returns (as opposed to >2 GHz, since I care less about power efficiency). I will account for the 11400 having two fewer cores than your 5800H but supporting AVX-512 (I'll use mackerel's "13%" figure for the latter). My extrapolated "frequency of diminishing returns" with 3200 MHz dual-channel RAM is then around 3.5 GHz, which is comfortably below the 11400's expected sustained boost once I take the 65 W power cap off. So the obvious conclusion seems to be that if Prime95 can't fully load down even a lower-end i5 unless you use expensive specialty RAM, then an i7 or i9 probably wouldn't get close in any realistic configuration.

It would of course be a waste of such a powerful processor, but one wonders just how fast the 10900 or 11700 (for example) could run PRP-CF — since the FFTs for the current PRP-CF wavefront do fit in a 12, 16, or 20 MB cache easily. In fact, that gives me the idea that in a hypothetical full-frequency 10900 or 11700 system, one might figure out how many threads can be working on PRP-CF before the throughput of the main PRP worker is significantly impacted. This is unless I'm misunderstanding how large FFTs interact with a small cache; I believe that if an FFT is too large and goes to RAM, something else that wants to take up a large amount of cache and can fit will be allowed to, but have nothing to back that up. The obvious hole in this idea even if my understanding is correct is that PRP-CF tests being small also makes the wavefront advance rapidly. I have no idea how long it might take (with, say, a 16 MB cache) before wavefront PRP-CFs start eating memory bandwidth just like wavefront PRPs.

* Unless I were very concerned with (iters/sec)/watt, which I'm not unless power consumption gets into "cooling problem" territory.