FRONTPAGE Intel i9-11900K and i5-11600K Review: A Stopgap Until 10nm

[mackerel]

Ok Mack, tried it out, not sure if this is exactly what you want. Did 128K min/max first test was 5 seconds and ran around 4.0 GHz second set is 30 seconds 1 core 4100 all cores dropped to 3.7 during the complex test

Thanks. That's sufficient to answer the biggest question I had, which is, does the AVX-512 implementation improve the Prime95 IPC when not limited by other factors e.g. ram bandwidth.

The answer is yes. I'm not at home at the moment but I just did a comparison on my 7300HQ which scores around 8100 with 4 cores at 3.1 GHz. 128k FFT is small enough to run one per core so practically unlimited in performance. Normalise clocks and cores, and Rocket Lake comes out about 50% faster than Kaby Lake (and other Skylake architecture CPUs by association). I can run it against Skylake-X when I'm home again as a check. This makes it interesting enough I'll definitely buy my own sample to use once I'm home again.

Ideally we'd see up to 2x perf, but there could be many reasons (including Prime95 itself) why we don't see that, hence wanting to check on Skylake-X as another direct AVX-512 comparison point.

 

[FlanK3r]

not bad. I have 11900K and KF. KF not tested yet, but with 11900K I can not get more than 5250 MHz for hard benchmarks. And it is hot and power hungry. So propably best way for performance is
1)default setings with Intel Adaptive boost ON
2)v/f curve settings with specifics offset and ratios (maybe gold core can hold 54 for single core load)

 

[Woomack]

I was (finally) playing a bit with my 11700K during the weekend and on 360 AIO I couldn't set anything above 4.9GHz stable on all cores. Max was 4.9GHz 1.175V LLC2 in BIOS ~1.24V max in OS. This was giving ~95°C in the AIDA64 stability test, CPU+FPU. The same on 2 Z590 motherboards so far.

 

[EarthDog]

Disable avx-512.

 

[Woomack]

I just said how it works for me. I won't play games or anything like that on this CPU (or not anytime soon) so I test it like it was designed.
Btw. did you get the new SuperO/Supermicro mobo? I have one and it's ... magical

 

[EarthDog]

I just said how to get around that limit...if you choose.

No SuperO board yet. But I've had a dozen others sitting here for weeks.

EDIT: SuperO boards are sometimes a PITA. I'm thankful I worked in a server/data center environment for a decade so I know how to use the boards, lol. They are more crossover server boards (with the current state of its BIOS) than they are user-friendly. But they keep making improvements but generally solid otherwise.

 

[Woomack]

I remember your problems with the last SuperO/Supermicro motherboards ... that's why I said it's magical ... slightly better than the Z490 but not much different
I can say that the mobo has a good power design, nothing is overheating, no problems with the CPU at 350W+. It comes with 10Gbps LAN and other things which are usually in the most expensive motherboards. On the other hand, the same complicated and sometimes weird BIOS settings, and problems with memory support. Maybe they release a new BIOS as for now there are no additional IMC settings/ratios/modes which are on some top motherboards so anything above ~3600 causes huge memory performance loss.

 

[EarthDog]

At least it runs Gear1 at DDR4 3600! I have a couple here (budget boards, generally) that default to Gear 1 with DDR4 3600. One board, the writes and copies are nearly half, but this isn't a problem with gearing as no other boards drop anywhere close to what I've seen on one board when I make the switch.

For the end user, this launch I don't think is too painful... for reviewers, feels like a PITA, lol!

 

[Woomack]

On SuperO, read bandwidth is exactly half at gear 1 but write and copy look fine. I was able to run memory at 4266 CL18 but only 27GB/s read and about 58/55GB write/copy. On ASRock, I only tested HyperX 4600 at XMP so far and it gave me ~68/68/68GB/s and 52ns.
ASRock doesn't want to send me more samples as they are not releasing all their models in the EU and some are not available yet. For example, the Phantom Gaming ITX won't be released in the EU and I like to review ITX mobos The same mobos that have WiFi and no WiFi option like Extreme and Extreme E will be only available without WiFi in the EU.
There will be OC Formula but I'm not really convinced after looking at the motherboard. I wish to play with Gigabyte Z590 Tachyon but I don't want to pay for any of these new motherboards. Nothing really special and overpriced. Not to mention that in 6-8 months is expected something new from AMD and the next Intel.

 

[mackerel]

Got my 11700k today. In testing so far, it confirms the earlier results of about 50% IPC increase in the Prime95 test condition. Perf/watt running AVX-512 is about parity with AVX2 of 8086k and 11600k. My mobo default settings runs the CPU at 4.6 GHz 1.35V for all core AVX-512, hitting a peak of 216W. Running that for sustained time will hit thermal throttle under a Noctua D15. I think a power limit to say 150W will help a lot with efficiency while not impacting performance too much, so that's on the to do list. The AVX-512 perf is lower than Skylake-X's implementation though, which is >80% relative to AVX2.

I did try y-cruncher as another AVX-512 workload. The multi-thread result was faster but not much so. Single thread was as good as Skylake-X. With the faster cores, the mt perf is badly limited by low ram bandwidth. I have dual channel single rank 3600 in system and I don't think quad channel is enough either.

General observations: the CPU reminds me a lot of Zen 2 in that when idle or light load, the voltage boosts right up. I'm frequently seeing 5 GHz which practically never happens on my 8086k that is also supposed to do that.

 

[EarthDog]

What board? You may want to see if your board setup the RAM 1:1 to the memory speed (Gear 1). At 3600, MOST (not all, I have two in hand that don't), set 1:1. At 1:2, bandwidth drops and latency goes up and performance drops, particularly on latency and bandwidth-sensitive apps. I don't play P95, so no testing done that way.

 

[mackerel]

What board? You may want to see if your board setup the RAM 1:1 to the memory speed (Gear 1). At 3600, MOST (not all, I have two in hand that don't), set 1:1. At 1:2, bandwidth drops and latency goes up and performance drops, particularly on latency and bandwidth-sensitive apps. I don't play P95, so no testing done that way.

Gigabyte Z490 Elite AC. I only checked it has latest bios but not looked at ram settings beyond setting XMP. Totally forgot about the new gear thing. For Prime95, 128k FFT test size was chosen specifically as it is small enough to fit in caches, so those results are not affected by ram. For now I want to focus on the architecture, not connectivity limitations which have been a problem for many years.

 

[mackerel]

Is there a way to tell what gear is in use in software on the desktop? The mobo bios was set to "auto" so I have no idea. CPU-Z showed memory controller at 1800 MHz, as was the ram. I set it to "gear 2". No change. Set it to "gear 1". No change. I ran some Aida64 bandwidth measurements, but my copy pre-dates Rocket Lake so keeps popping up a warning saying it is unsupported. I didn't keep up the subscription on it and don't intend to. Anyway, between the two gears, no change on read/write speeds. Copy was unstable but best vs best is no change. Latency did decrease slightly in gear 1. Running y-cruncher multi-thread in gear 1 gave within 0.5% of "auto" result. I believe y-cruncher is similar to Prime95 when it comes to memory accesses in that bandwidth matters more than latency.

Edit: I have to correct a previous statement. I had stated Rocket Lake was about 50% faster and Skylake-X >80% faster than without AVX-512. That was based on a test using my Kaby Lake laptop as reference. Now I look at the numbers more, that laptop performed worse than expected relative to Coffee Lake/Comet Lake. If I use the Comet Lake results as reference, that becomes 43% and 75% improvement respectively, which is closer to what I saw with some real world testing separately.

 

[EarthDog]

There are a few ways, but in CPUz, check the memory tab and the Memory Controller Frequency. If it's 1:1 it will match your DRAM frequency (1:2 will be half).

But yes, since the FFT length all fits in cache, it likely won't make much of a difference. I am curious though as to how that board sets it out of the box.

 

[mackerel]

As written above, when specifically set to gear 1 or gear 2, both memory controller and ram clocks are the same (1800 MHz for ram at 3600). Bios bug? This is not the first time I've encountered Gigabyte bios not applying settings correctly. I did a load defaults after the CPU swap so hopefully there's no funny stuff left over from before.


In discussion elsewhere I'm left wondering, just where should the AVX-512 (specifically FP64) performance be on Rocket Lake? My only experience before was Skylake-X and by association, the related server offerings. Skylake-X had two units, which gave up to 2x the performance per clock. I never owned one, but some server versions had one unit, and I heard of others experiencing performance regressions when AVX-512 code optimised for 2 unit implementations were run on it. I don't know if that was ever resolved beyond the workaround of forcing the software to use AVX2. The comment I saw suggested Rocket Lake might have 3x256-bit FPU vector units. I have been unable to confirm this as the little available info suggests it only has 2x256-bit units that can be combined into one 512-bit unit, just like the closely related Ice Lake and Tiger Lake cores.

However, 3x 256-bit units would fit my performance observations so far for Prime95, where it has clearly more performance per clock than AVX2 CPUs, and less performance than a 2 unit AVX-512 implementation. The alternate explanation might be there actually is only two 256-bit units but improvements elsewhere lead to the increase in performance. For example, Zen 2 in this use case was 4% faster than Skylake, and Zen 3 was another 10% faster than Zen 2. All those have two 256-bit units. Still, 42% is quite a bit more than Zen 3 even.


I just had an unexplained crash on the system since adding the CPU. Mining was running as was a youtube video. Sudden reset. Running some stress tests now since I never checked before and just swapped the CPU over, although nothing unusual happened during the past Prime95 or y-cruncher runs. It was interesting when the 10600k was "new" I had some occasional weirdness but nothing recently. Almost as if the system needed time to settle, be it hardware or software. I also note my bios doesn't have the latest known microcode for Rocket Lake, so there might be something there too.

 

[Johan45]

Could have to do with the 1:1 ratio. One review I read said 1:1 @ 3800 was unstable. Check the SA and IO voltages too. Two boards I have both ASUS set the SA voltage 1.55V+ and IO over 1.3 V setting gear 2 drops these to 0.8V and 1.03 V respectively. Higher than normal voltage is needed to run 1:1, I did manage to drop the SA voltage to 1.4V and remain stable. Asking ASUS they said it's necessary for broad stability at that RAM speed.

 

[mackerel]

I'll have a look shortly. I did about an hour of aida64 stress test with cpu+cache+ram selected, plus another 30 minutes of cpu+cache only. No sustained fpu for now until I decide how to handle the power. My original plan was to find a good AVX2 stress and base it on that. Cinebench R20 runs around 175W which isn't much lower... however that data point continues the trend, Rocket Lake seems to be about the same perf/watt as Comet Lake. But, higher performing = higher power.

 

[mackerel]

Hmm... still don't like the Gigabyte bios. It lists some voltages next to the setting, but I don't believe them. SA and IO were on auto, and next to them it showed 1.050v. There's also an IO2 which I don't think I've ever come across before. The vram voltage was showing 1.2v there, but to the side was another info panel showing the expected 1.36v for the higher speed setting. In Windows, hwinfo64 reported SA to be 1.284v and IO was nowhere to be seen.

I also tried using 4096k FFT in Prime95 as a more real world ram test than aida64. Here I did see a difference between gear 1 and gear 2, with gear 1 giving about 7% higher performance. Repeating this with Auto, that was same as gear 1 so I think we can say the Gigabyte does default to gear 1. Again, I saw no difference between controller and ram clocks in CPU-Z so that can't be used to figure the gear in my case.

 

[EarthDog]

THen there is this...

In the course of our Core i5-11400F "Rocket Lake" processor review, we discovered that the Gear 2 memory mode has the potential to offer higher performance than Gear 1. The Gear 1 mode runs the memory frequency and memory controller frequency in 1:1 sync, while the Gear 2 mode runs them at 1:2, meaning that the memory controllers run at half the memory frequency, allowing you additional memory overclocking headroom. At lower, more stable, memory frequencies, it should be logical to use Gear 1. Our testing springs some surprising results.

Overall, a stock Core i5-11400F paired with DDR4-3733 MHz memory, was found to be 1.5% faster with Gear 2, when averaged across all our CPU tests, compared to Gear 1 at the same 3733 MHz frequency. Gear 2 was 3.42% faster in Cinebench R23 multi-threaded, and a staggering 6% in MySQL. Across rendering and media workloads that scale across all cores, we find Gear 2 faster by 1-3%. It's only with less parallelized workloads such as gaming, where we see Gear 2 lag behind Gear 1, though not by much. In our i5-11400F review, we show that by running your processor in Gear 2, you're making your memory controllers pull less power, freeing up power budget for the CPU cores, translating into the nT performance gains we see here. We discovered that the uncore can pull anywhere between 5 to 10 W more power in Gear 1 mode. This is valuable power eating into the already constrained power-budget of this 65 W TDP chip.

This behavior wasn't spotted in our launch-day i5-11600K and i9-11900K reviews; as the i5-11400F is our first Rocket Lake with a 65 W TDP. We expect the disparity between Gear 2 and Gear 1 at stock settings to only grow wider with higher core-count 65 W models, such as the Core i7-10700/F and the i9-11900/F. Gear 1 claws its way back to the top when you engage power limit overrides offered by motherboards. These overrides vary from motherboard to motherboard, but they generally free up more power budget for your CPU cores to sustain their boost frequencies better. For more data and commentary, be sure to catch our Core i5-11400F review.

https://www.techpowerup.com/forums/...intel-rocket-lake-11th-gen-processors.280720/

Though I'd wait for additional data sets before this is The Gospel.

 

[Woomack]

On my 11700K, there is also ~1.35V and 4.6GHz in AVX-512 tests. I guess you can go down to ~1.20-1.25V what should keep temps at about ~90-95°C. I see that results are a bit different, depends on the motherboard. Some reviews where were compared multiple motherboards mention about it. In short, can be a 100-200MHz difference at the same voltage because of only the motherboard ... and not always these most expensive are the best. Somewhere was info that Maximus XIII Hero gives worse results than some lower Gigabyte or ASRock mobos.

On ASRock, gear 2 works well at 4000+ memory clock, but I wasn't comparing many settings. At least bandwidth and latency are where they should be at a high memory clock.
The same, on ASRock or Supermicro, there is no difference in NB/IMC clock in CPU-Z and it always shows the same, but I see different results in benchmarks so I know it works.