Latest CPU recommendations

Can confirm that with my i9-12900K at over 4000 Thread Mark single core performance, loading large samplesets like Nancy and Billerbeck takes 2 - 3 minutes and is mainly bottlenecked by CPU speed, not waiting for I/O on the SSD that much. Moving cache to a slower external SSD at one time (for shuffling directories around) only made a small difference in load time, so a faster SSD won't be a great investment if that's the only purpose, while a higher Thread Mark processor does help both load times and run time performance.

Curious. I expected a (much) greater improvement in sampleset loading with that new PC configuration. The 85GB of the Nancy sampleset loaded at 2'30 equates to a throughput of 570MB/s, which is just a little more than double the performance (~250MB/s) I can get on a very modest system based on an Intel mobile i7-6500U CPU (CPU Mark of 3277; Single Thread Rating of 1700).

Anyway, Martin announced and feeds us an expectation of a new loading engine that will improve (a lot?) the loading performance of samplesets. So, even without knowing whether this improvement will come soon or not, I will perform some future-proofing choices in this regard by purchasing a faster SSD with just a little more investment.

Whatever the medium term means. That can be 5 years. By then your PC will be old again

I'd say you can load Groningen (85 GB), and it should run okay on either of those processors using historically informed registrations.

At about 3400 Thread Mark and 25000 CPU Mark, you may have to set polyphony lower than 8000 voices, which SP recommends for that sampleset, to use HW 7 highest audio quality at 96k and/or a low latency buffer size. Just something to be aware of. You'd need to test and set the polyphony limit appropriately for the CPU.

That should only be a problem if you couple all the plenums together or try to draw a modern full organ -- things the organ probably wasn't built to handle gracefully anyway. If you need to do that (as a wind sag experiment for a documentary on herd animal noises), you could always drop one of the 4 mic perspectives temporarily by routing ranks to nowhere.

I don't have Groningen, so maybe somebody who does can give a better idea how it works on their CPU model.

In the last days, I received contradictory information about the compatibility of the 48/96GB memory modules with that Mini PC based on the AMD Ryzen 7735HS CPU, or even with the CPU itself, so I postponed the acquisition a little. It is not known for sure that compatibility already exists, or if it depends on a BIOS update or, if it will never exist at all. Meanwhile, information came about an imminent launch of Mini PCs based on the new Ryzen 7840HS/7940HS, which, in addition to being 10 to 20% faster than the previous 7735HS, have a greater chance of already being compatible (or acquiring the desired compatibility in the near future) with such non-binary memories. Anyway, someone did the that yesterday, bought the 96 GB kit and installed them on a NUC based on a Intel 13th gen i5 mobile CPU. It worked!

https://williamlam.com/2023/05/96gb-sod ... -esxi.html

Trying, therefore, to avoid a disappointment with those purchases, I decided to look for a NUC/MiniPC with both 13th gen Intel CPU and DDR5 support, and I found a model, manufactured by chinese Minisforum, named NPB7:

https://store.minisforum.com/collection ... 0152211701

Comparing the performance scores of these three CPUs, the i7-13700H ties with one of them and loses to the other by a small percentage of 3%:

https://www.cpubenchmark.net/compare/52 ... -i7-13700H

Assuming these scores might be reliable, one thing intrigues me. The Intel CPU has 14 cores and 20 threads: 6 "performance" cores at 2.4 GHz and 8 "efficiency" ones at 1.8 GHz, with respective turbo clocks of 5.0 and 3.7 GHz. Minimum, typical and maximum TDPs are 35, 45 and 115 W, respectively. As for the AMD Ryzen 7840HS, it has 8 (equal) cores and 16 threads. However, the base clock is 3.8 GHz (4.0 GHz on the 7940HS), much higher than Intel's, with the turbo reaching 5.1 GHz (5.2 GHz on the 7940HS). Minimum and typical TDP of 35 and 54 W respectively.

All three CPUs are between 29600 and 30600 CPU Mark. However, Ryzen's single thread ratings are about 10% higher, in contrast with the total number of cores of Intel's CPU, which is 75% higher.

So, the golden question: which of these CPUs would tend to provide the highest overall performance for Hauptwerk?

I'd be concerned that only 2 and 3 test samples (one sample = one run of the benchmark suite, then published by the user) means the AMD numbers will probably change as more people run the tests on more different PC builds. The confidence interval would be very wide if you had all the outcomes to do a statistical test, comparing populations of 2, 3, and 43. Not that tests on different PC models are IID anyway...

Right now, that's not a very solid comparison, though it's still more useful than trying to interpret base and turbo clock speeds and number of cores, which rarely makes sense unless it's within the same close family of CPUs.

Since it has more test samples, the i7-13700H already has a pretty firm score.

I do know that HW 7 does very well for me with mixed-speed P and E cores and variable clock multipliers on my i9-12900K, so that shouldn't deter you from buying an Intel gen 13.

Hello, can anyone tell me, which Priority order is for Hauptwerk most important (GHz, Single Thread Rating, Cores, Cache, AVX)?

GHz
Intel Core i9-12900KS 3.4 GHz (CPU Mark 44348)
Intel Core i9-13900KS 3.2 GHz (CPU Mark 62561)
AMD Ryzen 9 7950X3D 4.2 GHz (CPU Mark 63627)

Single Thread Rating
Intel Core i9-12900KS 4365
Intel Core i9-13900KS 4796
AMD Ryzen 9 7950X3D 4229

Cores (P/E)
Intel Core i9-12900KS 8 / 8
Intel Core i9-13900KS 8 / 16
AMD Ryzen 9 7950X3D 16

Cache L1 / L2 / L3
Intel Core i9-12900KS 1,152KB / 6 MB / 30 MB
Intel Core i9-13900KS 1,536KB / 16 MB / 36 MB
AMD Ryzen 9 7950X3D 1,024KB / 16 MB / 128 MB

AVX
Intel Core i9-12900KS AVX2
Intel Core i9-13900KS AVX2
AMD Ryzen 9 7950X3D AVX-512

https://www.cpubenchmark.net/compare/48 ... -9-7950X3D

If Martin answers, please ignore mine.

I'd say CPU Mark (aggregate multicore performance on parallel apps) and Thread Mark (single thread/core performance) -- these are about equally important. And never buy a CPU without AVX2 or AVX512 (either one is fine) if you can avoid it.

You have to have a good minimum of single thread performance, so don't go chasing extremes just for aggregate performance. A 40000 CPU Mark 4000 Thread Mark processor can do almost anything Hauptwerk samplesets can demand today. A 60000 CPU Mark 2000 Thread Mark processor will be *worse* for Hauptwerk because of slow single core performance, despite the higher multicore rating. I would go for at least 3000 Thread Mark and prefer 4000 or more.

GHz, cache, and core count are physical descriptions of (usually) dissimilar hardware, so they are bad proxies for real, broad scope benchmark tests. They just leave out too many performance factors. They're good for marketing literature.
Each contributes to performance, but in different degrees on different machine families, even from the same company.

Yes, comparing GHz, cache, and cores between two Intel Core gen 13 CPUs is meaningful because they are all variants of the same design, but it's not good between gen 12 and 13 or between Intel and AMD.

Thanks for your assessment. I also see it that way, that CPU mark and Thread mark are important. Martin has always emphasized that the thread performance is very important.
AVX is also an important factor because it allows multiple calculations in one step due to the respective bit width. But how this affects the other factors can only be answered by Martin.

I wouldn't underestimate the cache now, because according to gamer tests it also has a big impact on the performance. It is also mentioned in the HW manual ("16 MB"):
"On both Mac and PC platforms, recent Intel processors in the i9, i7, or equivalent Xeon ranges of processors (with AVX2 or AVX-512 CPU instruction set support), with at least eight physical CPU cores, at least 16 MB of CPU cache, and a high base clock speed, are particularly recommended for best performance with large or complex organs in Hauptwerk, and for large numbers of audio channels and real-time impulse response reverbs".

Only the assessment of which of the three mentioned CPU's is the best-suited CPU by the rating order is not so definitively clear to me.

I would be interested to know what effect AVX-512 has in the overall picture (whether the -512 is really used, or only the instruction set of AVX2).

The effect of cache and its interaction with the particular processor is fully covered by the two benchmarks, so it shouldn't need to be considered separately.

A previous thread (I'll try to find it) said AVX512 wasn't better than AVX2 for Hauptwerk. I think that was because vector calculations were mostly on the up to 4 perspectives per rank, of which mostly we use 1 or 2, and probably convolution. Also, I'm sure the PassMark suite includes vector computations.

EDIT: It was an answer to you: viewtopic.php?f=16&t=20719&p=155080&hilit=avx%2A512#p155080

Single thread rating is not important for Hauptwerk. HW is a multi threaded program and uses all available cores.
So, multi thread rating must be compared (Cinebench R23 MT).

In case of single thread performance: Core i9-13900KS clock is 6 GHz but the limits are the usage of max. 2 cores and a max. CPU temperature of 70 degrees Celcius. Otherwise, the clock is lowered to 5,4 GHz and lowered further if temperature becomes too high (the CPU trottles).

HW requires that the CPU clock to be constant to avoid audio hickups. Disable this in the BIOS.
for Intel processor:
- Intel Turbo Boost Technology
- Enhanced Intel SpeedStep Technology (EIST)
- Intel C-State C1 - C8 if that state is available in BIOS
for AMD processor:
- AMD Cool N Quiet
- Precision Boost (X-Boost)
- Core Performance boost
- Global C state control
Performance tests don't do that.

Ryzen 9 7950X3D has a large L3 cache, but is only available for 8 of the 16 cores because the cache is stacked on top of one of the CCD's. Reason is the cooling of the CPU cores.

A fine comparison with some technical background between Ryzen 9 7950X3D and Intel Core i9-13900KS is written (in Dutch language) on: https://www.ct.nl/achtergrond/ryzen-9-7 ... 9-13900ks/
This article was earlier published in their magazine but now online.

Single thread rating is not important for Hauptwerk. HW is a multi threaded program and uses all available cores.
So, multi thread rating must be compared (Cinebench R23 MT).

In case of single thread performance: Core i9-13900KS clock is 6 GHz but the limits are the usage of max. 2 cores and a max. CPU temperature of 70 degrees Celcius. Otherwise, the clock is lowered to 5,4 GHz and lowered further if temperature becomes too high (the CPU trottles).

HW requires that the CPU clock to be constant to avoid audio hickups. Disable this in the BIOS.
for Intel processor:
- Intel Turbo Boost Technology
- Enhanced Intel SpeedStep Technology (EIST)
- Intel C-State C1 - C8 if that state is available in BIOS
for AMD processor:
- AMD Cool N Quiet
- Precision Boost (X-Boost)
- Core Performance boost
- Global C state control
Performance tests don't do that.

Ryzen 9 7950X3D has a large L3 cache, but is only available for 8 of the 16 cores because the cache is stacked on top of one of the CCD's. Reason is the cooling of the CPU cores.

A fine comparison with some technical background between Ryzen 9 7950X3D and Intel Core i9-13900KS is written (in Dutch language) on: https://www.ct.nl/achtergrond/ryzen-9-7 ... 9-13900ks/
This article was earlier published in their magazine but now online.

Martin has repeatedly explained the importance of single thread speed to Hauptwerk's several single-threaded models, so I won't repeat all that here.

About forcing fixed clock speed, I've found the exact opposite on my last three PCs, an i9-9900, an i9-12900K, and an AMD 5700U, where allowing the multiplier to change (I don't enable overclocking, so Turbo Boost speed) seriously improved dynamic polyphony over the base clock speed, and caused no glitches at all. I've stress-tested that thoroughly on all three machines. I do understand why Martin recommends the safe approach of using the fixed base speed, but I lose 50% of my polyphony on the i9-12900K when I do that. You have to make a decision on a case-by-case basis.

Until we have a database of uniform polyphony tests collected by users for Hauptwerk, publicly available benchmarks are more valid than trying to interpret architectural features. I was a capacity planner/performance modeller for years, and I wouldn't rely on comparing brochure data.

Yes, I know what Martin is saying. In my opinion he means the importance of the speed for all threads and not only a single thread.

What I try to say above is that comparing the single thread performance in benchmarks gives you wrong info. The single-thread test only uses a single core and single thread, the amount of cores or hyper threading ability doesn't count. And that is not the case with Hauptwerk. For one core the CPU clock can be 'very' high which is not the standard situation; there are some limitations. Therefor I have given the example of the new i9-13900KS.

Cinebench R23 MT uses all available cores and threads. CPU clock is not set at a fixed speed but varies.
For an overview and difference, see https://nanoreview.net/en/cpu-list/cinebench-scores

Because you don't overclock when a fixed CPU clock is used, there is a great difference in polyphony in your case.
But in a modern BIOS you can directly set the clock 0,5 GHz higher than it's base clock without problems.

I didn't suggested *only* comparing single threaded performance, just that it's about as important as aggregate performance and you need both for HW. I don't think you actually disagree with that, and you make some good points.

Anyway, the only accurate way to compare is to test HW on each CPU model, and we can't. All we're talking about is which tests and/or marketing numbers might make decent proxies in the absence of HW testing, so there's room for differing answers and not enough direct evidence to be sure of the best proxies.

If HW came with an automated performance test that measured polyphony and reported the user settings, machine specs, and results back to MDA, we could know how to compare machines in some cases. I don't imagine Martin has time to do that.

About overclocking, you're right that I could set a fixed overclock multiplier. I've never had any reason to use a fixed clock rate on the i9-12900K at all, so I don't. It does a great job with Turbo Boost and Speed Step both on, cools down quickly when I'm not playing full organ using a lot of polyphony voices. Speed Shift kills Hauptwerk performance on this machine, though. (Oddly, the Dell XPS 8950 BIOS doesn't allow for Turbo Boost with Speed Step, perhaps because of cooling. I have the liquid cooled model.)

Thank you for your opinions and links!

You both have good and coherent arguments. It would be very helpful if Hauptwerk speed test samples existed like on cpubenchmark.net.

It is indeed difficult with the heat the CPU's generate.
I am reluctant to use a water cooler because again there is a lot of technology that will break faster than an air cooler like Noctua's NH-D15. This one offers a MTTF of 150000 hours with a 6 year warranty.
https://noctua.at/en/products/cpu-cooler-retail/nh-d15

I have a cooling tower from Noctua in my old PC with the CPU 3930K (6 core @ 3.2 GHz), which has been running for over 10 years until today without any problems and very quiet. My CPU temperature is always below 50 degrees (around 45 in winter and below 55 in summer).

Which RAM do you use? DDR4/5 with what throughput? Mainboard?

128GB DDR5, 4x32GB, at 3600MHz

I don't know what mb they use.

ASUS Prime X570-PRO
Ryzen 9 5950X (running on 4 GHz fixed clock speed) with a Scythe Mugen 5 PCGH Edition Rev. B Cooler
G.Skill Ripjaws V F4-3600C18Q-128GVK – 128 GB DDR4 (https://www.gskill.com/specification/16 ... cification)
MSI GeForce GT 730 2GB LP V1 graphical cart
Samsung 980 Pro 2 TB
4x 12 cm PWM PST fan in a Corsair midi tower case; speed control defined in BIOS