128Gb Eight Channel Workstation Madness Paired With ThreadRipper 3995WX & 24Gb GeForce 3090Ti In Windows 8.1 Operating System
You will see Windows 7 benchmarks in a while if i’ll be able to launch this beast under the Windows 7 wing
¶ PREFACE NOTE
Alright, this is the final multicore beast, operating on the mighty Windows 7.
While it is possible to run a Threadripper 7995WX or even a 9995WX, I am not going to do that.
This isn’t just because of the insane prices, but also due to the enormous TDP (>350W) of modern
high-performance CPUs and my sheer lack of desire to mess around with water cooling solutions.
Even though the 5995WX operates in the 280W range, I’ve decided to stick with the 3995WX because
I scored a pretty good deal on a brand-new Gigabyte GA-WRX80-SU8 mobo. Unfortunately, this mobotherboard
does not support the 5000 series, at least not officially. Support was present in the F5 version
of the BIOS, but it was rapidly removed due to compatibility and stability issues.
You can still acquire that older BIOS from Eastern servers or web archive backups,
but taking all of the above into account, common sense prevailed. I decided that saving money is a smart move,
especially when I’m not sponsored to do these kinds of builds.
A quick note on RAM: 2400MHz was chosen simply because I got a great deal on it.
AMD’s recommended speed is 3200MHz, which would theoretically provide 25% more memory bandwidth.
However, in practice, the performance increase would only be around 10–15% due to general system overhead.
Keep these figures in mind to get the complete picture.
As for cooling, a top-of-the-line Noctua NH-U14S TR4-SP3 will handle the 3995WX relatively easily.
Stock BIOS dated back 2021 had CSM issues, so it is recommended to update it to the latest F6 revision (2023).
█ COMPUTER CONFIGURATIONS
[1] ► ThreadRipper 3995WX / Gigabyte WRX80 SU8/ 128Gb DDR4 2400Mhz [8-channel] / GeForce 3090Ti 24Gb / Noctua NH-U14STR4-SP3
[2] ► Ryzen 9 7950X / MSI X670 / 64Gb DDR5 5600Mhz [2-channel] / GeForce 3090Ti 24Gb / Noctua NH-D15 SE AM5
PSU: Seasonic 1300Watt Gold [in both cases]
GFX SETTINGS:
2560x1600 32Bit color [where possible], ULTRA+ settings, Anti-aliasing is turned OFF
SOFTWARE:
Operating system: fully patched Windows 8.1 x64. All software were patched to the latest versions.
▒ FINAL REALITY [1996]
| SYSTEM | [1] | [2] |
|---|---|---|
| ALL | 109/13/21 | 198/16/38 |
| 3D | 5000/348/254/429 | 5353/490/84/562 |
| 2D | 815/200 | 962/546 |
| BUS | 122/691 | 598/899 |
| AGP | 536 | 972 |
Obviously due to it’s age benchmark can’t comprehend the real power of Threadripper
▒ 3DMARK 1999 MAX
| TEST | [1] | [2] |
|---|---|---|
| 3D | 25 | 25 |
| CPU | 1024 | 632 |
Threadripper is valued a bit higher!
▒ 3DMARK 2000
| TEST | [1] | [2] |
|---|---|---|
| SCORE | 2300 | 2303 |
Tie! Looks like the measuring limit for this benchmark.
▒ 3DMARK 2001 SE
| TEST | [1] | [2] |
|---|---|---|
| SCORE | - | 128796 |
Testing application could not start, even patched version fixed to work with modern CPUs
▒ 3DMARK 2003
| TEST | [1] | [2] |
|---|---|---|
| SCORE | 144599 | 7805 |
Surprise! 2003 benchmark has a capability to evaluate Ripper.
▒ 3DMARK 2005
| TEST | [1] | [2] |
|---|---|---|
| SCORE | 46212 | 14998 |
ThreadRipper is a winner again.
▒ 3DMARK 2006
| TEST | [1] | [2] |
|---|---|---|
| 3D | 39815 | 18386 |
| SM2 | 13670 | 7200 |
| SM3 | 18522 | 5998 |
| CPU | 15027 | 20941 |
Threadripper here is substantially slower than consumer Ryzen, but GPU shines with 3995WX, quite weird results
▒ 3DMARK VANTAGE [2007]
| TEST | [1] | [2] |
|---|---|---|
| GPU | 81479 | 20471 |
| CPU | 64365 | 100900 |
GPU showed a little better performance with Threadripper, CPU-wise Ripper performs very badly comparing to Ryzen.
Just for your information: Vantage is a DirectX 10.1 benchmark.
GPU performance on 7950x looks like a glitch of a hacked 3090Ti driver.
To observe Windows 7 results of 7950X machine click [here]
▒ 3DMARK 2011
| TEST | [1] | [2] |
|---|---|---|
| GS | 17230 | 13607 |
| PS | 17217 | 18901 |
| CS | 15796 | 12900 |
| GT1 [fps] | 85 | 86 |
| GT2 [fps] | 82 | 81 |
| GT3 [fps] | 104 | 101 |
| GT4 [fps] | 52 | 50 |
| PT [fps] | 55 | 119 |
| CT [fps] | 74 | 100 |
DirectX 11 benchmark, handles Threadripper noticeably better than older ones
▒ 3DMARK 2021
| TEST | [1] | [2] |
|---|---|---|
| Fire Strike Ultra | 24202/23314/14592 | 26800/48004/14820 |
We can clearly see that Threadripper is not a gaming platform at all
▒ UNIGINE SANCTUARY [2007]
| TEST | [1] | [2] |
|---|---|---|
| FPS | 755 | 636 |
| MARK | 32000 | 26981 |
Despite benchmark’s age noticeable victory of Threadripper here
▒ UNIGINE TROPICS [2008]
| TEST | [1] | [2] |
|---|---|---|
| FPS | 476 | 567 |
| MARK | 12001 | 13304 |
Ryzen is a winner, curious to see that more modern benchmark handled Threadripper platform in a different way
▒ UNIGINE HEAVEN [2009]
| TEST | [1] | [2] |
|---|---|---|
| FPS | 224 | 241 |
| SCORE | 5628 | 5866 |
| MIN FPS | 10 | 12 |
| MAX FPS | 467 | 471 |
Ryzen is a winner again, it looks like the win of Threadripper in Sanctuary test was a glitch
▒ UNIGINE VALLEY [2013]
| TEST | [1] | [2] |
|---|---|---|
| FPS | 163 | 236 |
| SCORE | 6800 | 9893 |
| MIN FPS | 33 | 58 |
| MAX FPS | 260 | 389 |
Ryzen destroys Ripper here, not leaving any chance
▒ UNIGINE SUPERPOSITION [2017]
| TEST | [1] | [2] |
|---|---|---|
| SCORE | 22810 (7800) | 8018 (7541) |
Threadripper wins here almost in a triple rates, note that shaders were set to 4K optimized and resolution was set to 2560x1600. Extreme shaders setting delivers only 7800 score points. It seems that benchmark is well written and optimized for high core count CPUs.
▒ STALKER BENCHMARK [2009]
| FPS [Avarage/Maximum] | [1] | [2] |
|---|---|---|
| DAY | 276/604 | 411/667 |
| NIGHT | - | 448/760 |
| RAIN | 355/815 | 497/775 |
| SUN | 266/557 | 460//510 |
Night section was skipped due to a program glitch
► CPU-Z
| v2.20/17.01.64 | [1] | [2] | i9-285K |
|---|---|---|---|
| Single | 517 | 760 | 898 |
| Multi | 34120 | 15530 | 18855 |
Ripper destroys both Ryzen and i9 in multi-threading, single core performance is weak due to Zen3 architecture
► MEMORY PERFORMANCE
| AIDA64 v | [1] | [2] | Geforce 3090Ti |
|---|---|---|---|
| Single-Precision FLOPS | 5972 GFLOPS | 2657 GFLOPS | 42283 GFLOPS |
| Double-Precision FLOPS | 2977 GFLOPS | 1324 GFLOPS | 654 GFLOPS |
| 24-bit Integer IOPS | 1983 GIOPS | 2085 GIOPS | 21337 GIOPS |
| 32-bit Integer IOPS | 1980 GIOPS | 2120 GIOPS | 21337 GIOPS |
| 64-bit Integer IOPS | 450 GIOPS | 167 GIOPS | 5003 GIOPS |
| AES-256 | 450000 Mb/s | 356703 Mb/s | 114000 Mb/s |
| SHA-1 Hash | 183700Mb/s | 73584 Mb/s | 239844 Mb/s |
| Single-Precision Julia | 2213 FPS | 1108 FPS | 6517 FPS |
| Double-Precision Mandel | 1181 FPS | 592 FPS | 172 FPS |
No commentaries here, just compare yourself
► CACHE PERFORMANCE
| AIDA64 v | [1] | [2] |
|---|---|---|
| L1 Cache READ | 13120 Gb/s | 4903 Gb/s |
| L1 Cache WRITE | 7352 Gb/s | 2635 Gb/s |
| L1 Cache COPY | 13937 Gb/s | 5100Gb/s |
| L1 Cache LATENCY | 0.9ns | 0.7ns |
| L2 Cache READ | 6541 Gb/s | 2575 Gb/s |
| L2 Cache WRITE | 6576 Gb/s | 2388 Gb/s |
| L2 Cache COPY | 5800 Gb/s | 2472 Gb/s |
| L2 Cache LATENCY | 2.8ns | 3.4ns |
| L3 Cache READ | 580 Gb/s | 1476 Gb/s |
| L3 Cache WRITE | 1410 Gb/s | 1238 Gb/s |
| L3 Cache COPY | 1540Gb/s | 1318 Gb/s |
| L3 Cache LATENCY | 10.5ns | 10.3ns |
The memory subsystem on Threadripper just rips Ryzen apart with it’s consumer oriented heritage
▒ CINEBENCH R20
| - | [1] | [2] |
|---|---|---|
| CINEBENCH R20 | 23128 | 13645 |
▒ CINEBENCH R23
| - | [1] | [2] |
|---|---|---|
| CINEBENCH R23 | 59404 | 35153 |
As you can see CPU rendering is a relax zone for CPUs like 3995WX.
Ripper destroyed Ryzen despite more modern architecture, DDR5 and advanced lithography process.
This is where high core matters!
▒ RENOISE PERFORMANCE
Renoise testbed conditions.
- track:257 and without any problems reached [Ta! Amiga Rulez!!!]
- FX : > 1500
- No audio cracks at all
- interface: ASIO/48Khz/~5ms
- CPU overload protection: 98%
7950X 32CPU reached: Max Load: 24%
3995WX 128CPU reached: Max Load: 24%
So, as we can observe the load is the same as with 7950X.
But if we look into windows task manager we’ll see slightly different figures.
I think that is a more reliable source, because it is operating system, rather than just application.
Looking at these graphs we can note ZERO load on threads 64-128, so it is safe to conclude than even modern
version of Renoise 3.5.4 with multicore improvements just can’t manage such amount of cores/threads.
32 cores (64 threads) is an absolute maximum CPU that tracker can handle and not with 100% efficiency of course.
Couple of words about the quirks of the 3995WX system, loading module and closing application take almost
triple time compared to consumer Ryzen 7950X. It looks like that program tries hard to comprehend this insane amount of cores.
If you are interested in using latest Renoise 3.5 on Windows 8.1 despite the restriction of the installer,
head over to [Renoise forums] , where i explain how to do this in an easy way.
█ FINAL THOUGHTS
Peculiarities and features of the WRX80 platform that all legacy users must know:
- quite buggy CSM implementation even with the latest version of the BIOS
[but we all know that this board was not made for legacy compatibility] - very long boot times i presume this is happing due to octal memory configuration and complex training algorithm
[even longer than modern Ryzens, sometimes it drives me nuts] - the single-tower Noctua air cooler cooled the system very well,
thanks to its excellent design and the large silicon surface area of the Threadripper CPU
[actually, this was a surprise for me]
Finally, I have completed what I’ve always dreamed of.
It was a fun experience, but we can to push things much further [if we want to].
If you for some reason need to stay legacy, here is an example on how to do this.
As you can see you are not limited by performance at all.
If your budget allows you to spent so much money and stay on Windows 8, you can freely install even 9995WX,
since Asus motherboard based upon WRX90 has CSM option in the BIOS.
Please note that things will not work out-of-the-box, you will need to know that WRX90 BIOS
can be picky in terms of BIOS settings and hardware environment.
And it can be even more pickier than Gigabyte WRX80!
So, we are already on the uncharted territory and things can go pretty wild and unexpected.
One more thing to remember is to purchase water cooling for such a beast.
Ah, and almost forgot, take a note that PCIe PS/2 controllers are not working on the WRX80 chipset.
Motherboard debug LED just showed error 27 and prevented system from booting,
it looks like that it’s an issue with octa-chanel memory configiration, but i didn’t have any desire to dive into the topic.
But what is more important is that top-of-the-line Zen 7 CPU will soon meet the timeless Windows 7.
This is going to be a really cool showcase. Stay tuned for Zen 7 comparisons.
Someday we’ll pit it against the 3995WX, and maybe even something more powerful,
to squeeze absolutely all juice from the computer silicone.
If somebody has a screen-shot of task manager showing 192 threads running Windows 7 on a Threadripper 9995WX
please share it with me and will pinpoint provided information here with a reference to the original author.
