Yes, past a few cores there is little gain, cooling is challenging if you actually use all the available power, which is hard if not impossible - BUT, with the bios control mentioned, and the realization your never going to get full throttle anyway - there is a way out!

First, disable HT. If you have a need for the cores, build in more! you cannot use all the cores you can have -speaking XP32 here. Where I use handbrake this is still true, though not XP32. Disabling HT is good for smoothness, and provides some clocking headroom, and gains a 10-15% thermal headroom. Again, we are re purposing 2nd hand here, we can afford the obvious blasphemy of literally castrating a formerly high end cpu...

Next, you don't have to run all the cores. Same thought as the HT, but a little more rude to the pride of a high end xeon. Most bios's on a 5500 series X8 will allow 1,2,3,4,5,6 cores per socket. Again, adjustable thermal advantage. My floppy mentioned XP is what I ended up with after building on X8's with linux, multiple X8DTHxxx variants. I bought a X8DTi to recycle cpu's and memory after I noticed the chipset drivers on the CD and did a test nlited XP install on a 7 slot dual 5520 beast (X8DTH-i) that came up clean, though not really useful. The X8DTi-F install was clean too (qd3nodrv.inf), and correct sata from boot, and offers the floppy port and PCI. With dual 5687's as 2 core it will run with supermicro 4 pipe coolers and 900-1200rpm quiet 4pin replacement fans almost silently. Also in the bios you can force a LOWER multiplier for 100% duty NON-turbo. This really help the thermals yet will not restrict full turbo speed, that will game 4@3.9Ghz all day without noise. It sports an AMD Sky 500 16lane 4GB ECC working at full capability on XP32, and also is passively cooled. The swap is part of the PAE ram drive, 20.99GB of the 24 installed, it has passed with 48GB, 24x4, triple channel, fastest is 6 dimms only for full 1333 speed iirc. I think PAE can handle 64GB.
Usually I run all 4(x2) cores, no HT, and 18 cmos ratio. That's 2.4 limit non turbo, and like I said that does not limit the turbo, it just limits the mid clock, so it holds 2.4, or pegs 3.9ish. not 3.6 as rated. I never hear it.
Turn on HT and compile a kernel with t-16 and you will hear the OH alarm in about 20-30 minutes when the sockets hit 96C. In the 'archival fun' configuration as explained the cpu's range in the 70-80's, as does the gpu. It's in a wide body tower from the 90's that once housed my P3TDe-6. I made the 80pin backplane, ripped it out, modded the front cage to clear the board, etc.

No to the special power supply, not an issue, it runs a miniATX 500w PS that does have a 24 pin but I needed to fake it on the extra cpu connections, One 4 pin in each 8 pin is good enough, again, it takes effort to actually load these things to their max and an XP32 can't come close. As configured I measured 288W at the wall with effort to do so, I recorded mid 90C cpus and 97C GPU - and QUIETLY! In real use that peak is rare. With Debian or some other proper 64 bit OS it could maybe melt...The lack of a rack of scsi drives makes a huge difference, no, I don't miss that 'genuine' mess.

My favorite to date for my general use is still X9's. I have a X9SRA for single tasking and a X9DAi (3 pci-e x16 gpu's)for server crunching. That X9D is in a 19" rack case I modded into my desk when it had a P6SGU or DGe something...also in the 90's, I still have an AT rear plate for it) Nothing has excited me since. Past favorites of the single cpu variety; C7Q67 (E3 Xeon modded) and X9SCV-Q run XP32 well. C2SBC-Q's were fantastic workhorses that could overclock some and should be very compatible for older OS's.

Fun!

ruthan wrote on 2021-05-24, 01:35:

there is not place for 2 water coolers in 1 machine and some custom water cooling is complicated.

I forgot this part...maybe I made the case that Xeons can be turned into passive quiet over performers much easier than clocking their little brothers... I'm telling you, Xeons are better cpu's!

Anyway, in my past life I built many custom rigs. The most challenging factor with cooling is building for 12k+ altitudes. I even potted a few things, one big block of heat! I think water cooling sucks and is generally a bad idea. One thing that gets skipped with water cooling are the many hot spots that then get ignored. Chipsets of various purpose get hot too, vrm's, and often the largest power sucker and heat maker is memory (assuming Solid State storage). Water is simply a mess, and incomplete. Case airflow is paramount and more effective with all massive and passive coolers. There are also custom heat pipe kits that work well when combined with billet aluminum cases (at least one case wall). High dollar yes, but far superior to the point of silence!

So my 5687's run nearly silent and I don't think a few OC's Ghz make a lot of difference, especially when the stressed lesser option certainly isn't running silent!

Thanks, i checked all these boards, but there is every time something wrong for retro use.. One is PCI-E only and that one with PCI slots has bad slots placement.. so even if would accept complications with casing is not worth.

These make sense, if need cheap multicore performance first and some retro pc usage in second.. As i have other computers and want to invest money to build them, for me make sense make dedicated retro pc standalone were enough performance in more OSes is nice bonus and good for some backup pc / lan party etc.

ruthan wrote on 2020-06-17, 12:00:

Solution is probably replace it with AMD card - not tested yet, but it has some flaws. I personally dont like AMD cards from this era, because of worse power/performance ratio.. but mainly they are worse for Linux KVM, i maybe, maybe will make it working for WinXP virtual machine, but not for Win98 (videopass through is just not working for them)- i plan use for it other Nvidia card - Geforce 6xxx. There is chance that MacOS will ignore such one card, but i would not bet on it..

It does vary! I've found amd to vfio pass fine on one board, then fail migrated to another board. The 7 slot X8DTH's I mentioned will pass anything, seriously, everything I ever tried, gpu's, scsi, everything. Half the stuff wouldn't pass on any other board I tried. A scsi card revealed the trick to why I think....All those 16 slots are really 8's. Why that makes a difference I really don't know - but a brief test a few years back confirmed this theory when I cabled a 16 slot to a 16 gpu with a 8 wide cable and it worked. Not helpful, but it worked while it did not directly plugged in.

The X8DTi did pass the sky500 to XP. I was testing srv-io on that board with the gpu awhile back. The same does not pass on the X9DAi. Quadro's pass everywhere (FX1500/4500, K6xx). The FX1500 is a nice card to pass for 2k, but I put a passive cooler on it and it's too huge, from my C2SBC-Q. Late pci nvidia's usually don't, my pci 5450 passes. I'd expect that oddball sky500 to run Mac, BeOS, XP32. Although now only XP32 on a single ssd with my last optical doesn't get used much, it is a wide ranging box - you're coaxing me to play with it. The onboard matrox 200 variant could work with early OS's maybe? Why is the PCI slot poorly placed? It would need to have a sound card without a hdmi+ gpu. With the sky a single slot, a sound card and a pci gpu bring it up to passing 2 hardware assisted vm's! There even room for some more usb cards!

I had a call for some serial things to work on a modern system with a guest OS being 98, maybe this board is what I should try it on. I'll try XP first...I skipped 9x then, so...

ruthan wrote on 2021-05-23, 16:58:

No, but it would be probably same as with VIA 880.

Thanks ruthan!

Please, try our new cache control utility that allows you to slowdown your system with activated cache. It works with Pentium Pro+ CPUs and it would be nice to get your feedback about X58 and Core i7/Xeon compatibility.

Here is the official thread Feel free to post in english. Any questions and feedback are welcomed.

@White What is reason to make other cache on / off utility when there are others? Others are flawed or why?

If you read utility description carefully, you will see that it is not only on and off options. This utility allows to switch cache operating mode between writeback and writethru. Check it. I do not know any other utility that could do this. Or I’m wrong?

Description should be better why i would want to writeback and writethru, what is difference, what is expected performance effect etc?

White wrote on 2022-01-07, 12:04:

Please, try our new cache control utility that allows you to slowdown your system with activated cache. It works with Pentium Pro+ CPUs and it would be nice to get your feedback about X58 and Core i7/Xeon compatibility.

Here is the official thread Feel free to post in english. Any questions and feedback are welcomed.

Thanks, I have not tested in on newer Intel architectures but it works on Core 2 and also on AMD Athlon XP, Athlon 64 and Phenom II.
It's also worth mentioning that the utility does not work in protected/v86 mode when EMM386 loaded but it works with JEMM386 since it emulates privileged instructions.

ruthan wrote on 2022-01-09, 19:36:

Description should be better why i would want to writeback and writethru, what is difference, what is expected performance effect etc?

Practical translation:
1. Write-back cache enabled means full cache speed on modern architectures. That is both memory reads and memory writes are executed at maximum speed.
2. Write-through cache means different cache speeds for reads and writes. That is memory reads are executed at full cache speed but all memory writes are executed at only main memory speed.

So performance reduction depends on the application's memory usage pattern. If the application uses mainly memory reads then performance in write-through mode is more close to write-back/full cache speed mode.
But if the application uses mainly memory writes then performance in write-through mode is more close to no-cache/cache disabled mode.

For the purposes of DOS games compatibility, you want the read cache disabled to mainly make sure instruction reads are also slow, which best simulates a CPU running at slow clock rate and taking a while to execute each of those instructions. Slowing other memory operations is nice but secondary to the above. I've been successfully using SetMul.exe for my X58-based machine and it does take down my Xeon 2.67GHz to a stupid 386 33Mhz, which fixes most speed-sensitive games/programs for me (e.g. SB detection in Sierra games).

Practical translation: 1. Write-back cache enabled means full cache speed on modern architectures. That is both memory reads and […]
Show full quote

Practical translation:
1. Write-back cache enabled means full cache speed on modern architectures. That is both memory reads and memory writes are executed at maximum speed.
2. Write-through cache means different cache speeds for reads and writes. That is memory reads are executed at full cache speed but all memory writes are executed at only main memory speed.

So performance reduction depends on the application's memory usage pattern. If the application uses mainly memory reads then performance in write-through mode is more close to write-back/full cache speed mode.
But if the application uses mainly memory writes then performance in write-through mode is more close to no-cache/cache disabled mode.

Thanks, but this is still theoretical, to be practical we would need some numbers, what to expect 😀

Great read ruthan.

I'm building a similar rig using the Evga x58 3x SLI Classified. One added advantage to this board are hardware jumpers to disable PCIe lanes. In the end hope to mount an external switch to select which GPU is active depending on the OS I intend to boot.

In any case, I've hit a strange performance "ceiling" of sorts with the Evga board. I'm curious if you've encountered anything of the like. Here's my test:

I've built 2 x58 platforms:
1) Evga X58 3x SLI Classified
2) Asus Rampage II Gene
I've mounted the same Core i7-920 in each, a single 1GB stick of DDR3 (1333), an ATI X850XT and an Audigy 2 ZS. Both machines have a fresh Windows 98 SE installation with DX9, the ATI Catalyst 6.2 drivers (VSync disabled) and VxD drivers for the Audigy 2. No rlowe or any other patches installed (non needed!)

My result for the Quake 2 timedemo (map demo001) at 640x480 OpenGL:
Evga X58 3x SLI Classified: 4.4 seconds, 151 fps (capped? limited?)
Asus Rampage II Gene: 0.7 seconds, 948 fps (expected)
The Evga X58 is performing on par with a Pentium III 600/GeForce256 🙁

Have you experienced any sort of performance ceiling or limiting like this? BIOS options are where I'm constantly seeking a solution including disabling CPU features (Speed Step, HT, Virtualization, etc) but none have helped. I don't often build with equipment this "modern" or configurable. PIII performance for a machine that can boot everything is absolutely fine of course but a bit underwhelming given the hardware.

Any other X58 builder's run into this?

Hi. Try to set everything like described here: https://www.modlabs.net/articles/installing-w … tel-x58-chipset

@astigmatism - I find out other pro thing, you can disable whole PCI-E root on OS level to disable some PCI-E slot.. Its working at least WinXP+ in Win98 you dont need to disable things, because they are just unsupported.

Well i dont met exactly this sort of performance problems, but because sorting performance problems is my daily job. I can give at least some tips, or at least more questions to help find rootcause.
1) Its problem only with Quake 2, its performance similar /low in other games? I saw lots, sometimes on some machine 99% games work fine, but only 1 behave strange, its not probable, but.. How performance scale with resolution.
2) Are you sure than on both machines you have enabled DMA for good storage performance? If you rerun Quake 2 timedemo and you have much more FPS in second run, it could point to this problem.
3) Try some CPU only and mainly GPU benchmarks, to find about if problem is CPU or GPU related.
4) Try to disable sound card, if there would be significant difference.
Its also good ideal run same benchmark on same HW with WinXP.. its Win98 problem only, probably is but you never know. In my package is also Dos CPU testing utility, but it would be really strange would have problem even there, it would be some HW, bios problem for sure.
5) Again is not probable but even rlowe patches with some bad luck, can generate some performance problems too.

I think R.L. memory patch should have nothing to do with performance issues and he wrote that he didn't use it anyway...

I would try some nVidia VGA like GF6xxx or 7xxx on both MBs if perf. diff. would change...

Thanks for all the encouraging suggestions

Thanks also for the info regarding disabling PCIe devices at the OS level. I incorrectly assumed multiple GPUs in newer OS's posed a problem for builds of this type and considered a hardware solution an advantage. I suppose it's actually the opposite! 😀

Performance limiting is not isolated to Quake II but any of the games I use for Windows 9x benchmarking: Unreal Tournament, Expendable, Half-Life, Quake III even 3DMark 2000. I used Quake 2 as an example as the chasm between expected and realized performance best demonstrates my issue.

Here's another oddity I forgot to mention: there is no performance scale with resolution. The performance obtained at 640x480 is identical at 1600x1200. This is why I termed my problem as a performance ceiling. I certainly have a bottleneck in play here but where or what I have not discovered yet. I get a sense that PCIe bandwidth could be a factor here? Adjustments to the BIOS setting PCIe "Maximum Payload Size" made no difference. I was also to a point where I was pondering the possibility that Evga didn't implement the chipset or QPI pathing correctly.. ugh

DMA is always enabled. Its usually the first thing I do after a Win98 install 😀 Thanks for checking.

I haven't performed the same tests in Windows XP but this is a good suggestion. I began this build by configuring the oldest OS I intend to run and which I also expected to be the most troublesome. Perhaps I should move up.

For my platform testing I removed as much complexity as possible to isolate the problem so no rlowe patches (although I will typically use them). Trying a GeForce card is also a good idea. ATI's last Windows 98 devices and drivers are demonstrably more compatible than Nvidia's last devices and drivers. I do happen to have one of the FX PCX cards (Nvidia 5xxx series with a PCIe interface) so I should certainly try that - it won't explain why the Asus board is fully functional and the Evga is not however.

Regrettably, I've lost my motivation with this board and brought out another Asus X58 board from the collection (a Rampage III) so I'll be attempting a Win98 build with it instead.. mainly in light of OS configuration of PCIe device disabling. I suppose I figured in the end that the Asus X58 boards have more BIOS configurability than the Evga and that was ultimately where my troubles were rooted.

Well if still have 155 FPS with 1600x1200 isnt performance already good enough? Unless you have some 150 Hz+ monitor, it should be enough, unless other games are running slowly i mean <60 FPS.

Its not 1 to 1 but you can try Q2Dos same benchmarks in Dos if will able to get more fps. If yes problem would be some ATI drivers cap in Windows 98, but does not make sense that is only with one board.. it pointing more to some ATI graphic panel settings or some performance monitoring utility overlay etc.

Absolutely ruthan, the performance is still acceptable. It became more about understanding why the two platforms were behaving differently despite identical accessories and software.

As a final follow-up to this curiosity - It turns out I experienced the same bottleneck to performance on both my X58 Asus boards (Rampage II Gene and Rampage III BE). Here's the interesting bit however: only when I increased the physical RAM capacity to 4GB or greater. This configuration necessitated my use of Loew's patchmem. I'm fortunate to have a ton of 1GB sticks of DDR3 1333 (from dumpster server finds no less) and happened to discover this exact limit:

3GB of DDR3: Quake II timedemo 900+ fps
4GB of DDR3: Quake II timedemo 150 fps

Again, this is despite rloew's PATCHMEM or the typical changes to system.ini (MaxPhysPage and MaxFIleCache): Physically including a fourth 1GB module to increase the system total to 4GB with no other changes suddenly performance limited my configuration. This is also just behavior I've discovered on the Asus boards. The Evga X58 board was performance limited no matter what the RAM capacities were. Of course I'm fully aware that 4GB of RAM is a theoretical limit for a 32-bit OS and I'd be willing to accept my shortcomings based on this understanding alone but of course many of you are using much more than 4GB in your X58 builds without issue.

I often treat Vogons as a knowledge archive. These posts serve future hobbyists and enthusiasts as the same problems are discovered anew and a google search reveals an old Vogons thread. Perhaps a future hobbyist will encounter these exact issues and have the same questions. I'd highly encourage any new information for anyone able to reproduce this issue. At this point I'm going to move forward with my Asus Rampage III BE with 3GB as Windows 10's minimum requirement is 2GB and this is the final OS I intend to run on this build.

Thanks everyone 😀

astigmatism wrote on 2022-03-17, 15:48:

As a final follow-up to this curiosity - It turns out I experienced the same bottleneck to performance on both my X58 Asus boards (Rampage II Gene and Rampage III BE). Here's the interesting bit however: only when I increased the physical RAM capacity to 4GB or greater. This configuration necessitated my use of Loew's patchmem. I'm fortunate to have a ton of 1GB sticks of DDR3 1333 (from dumpster server finds no less) and happened to discover this exact limit:
3GB of DDR3: Quake II timedemo 900+ fps
4GB of DDR3: Quake II timedemo 150 fps

I give you a link with description about how to fix this issue. Don't you read it? Described method is possible to apply not only for 6k, but for other cards too. Please, read carefully.
Patchmem or other utilites will NOT help you.