Ok was a good read until NVME was mentioned. Going back to A useless short analysis on PCIX to PCIe bridge performance on Tualatin
I do not consider a G41 and PCIe part of when reading the original post....
what does "Supermicro 370DLE with two LIN-LIN and a pair of P-IIIS 1400." have to do with all that ??

Sphere478 wrote on 2023-08-19, 03:30:

I put it right into a pcie slot if I recall on the core2 setup

Okay, but getting only a few kBps is definitely not normal.

Did you try this on a more recent Linux distro? Debian Jessie is quite dated. I think it was around late 3.x kernel and I'm not sure how NVMe worked with that generation of kernels.

Horun wrote on 2023-08-19, 03:34:

Ok was a good read until NVME was mentioned. Going back to A useless short analysis on PCIX to PCIe bridge performance on Tualatin
I do not consider a G41 and PCIe part of when reading the original post....
what does "Supermicro 370DLE with two LIN-LIN and a pair of P-IIIS 1400." have to do with all that ??

Sorry for going off-topic. Was discussing about the existence of a PCI-X to PCIe adapter based on PEX8114 and whether NVMe would work well on those PCI-X systems with it.

And I ended up discussing about possible issues with such setup on some other old systems we tested (and yeah, had nothing to do with PCI-X). My apologies.

LSS10999 wrote on 2023-08-19, 03:39:

Okay, but getting only a few kBps is definitely not normal. […]
Show full quote

Sphere478 wrote on 2023-08-19, 03:30:

I put it right into a pcie slot if I recall on the core2 setup

Okay, but getting only a few kBps is definitely not normal.

Did you try this on a more recent Linux distro? Debian Jessie is quite dated. I think it was around late 3.x kernel and I'm not sure how NVMe worked with that generation of kernels.

Horun wrote on 2023-08-19, 03:34:

Ok was a good read until NVME was mentioned. Going back to A useless short analysis on PCIX to PCIe bridge performance on Tualatin
I do not consider a G41 and PCIe part of when reading the original post....
what does "Supermicro 370DLE with two LIN-LIN and a pair of P-IIIS 1400." have to do with all that ??

Sorry for going off-topic. Was discussing about the existence of a PCI-X to PCIe adapter based on PEX8114 and whether NVMe would work well on those PCI-X systems with it.

And I ended up discussing about possible issues with such setup on some other old systems we tested (and yeah, had nothing to do with PCI-X). My apologies.

No man, kbps on the k6

The core2 was mbps but still much slower than that same sm.2 tick on a modern today conputer. It was usable on the core 2 but slower than sata on the same system if I recall correctly

Which led me to what I am trying to say that the nvme is not good for retro and seems to be heavily cpu dependent.

Unlike sata, scsi and ide with its dma and dedicated processing.

Debian jessie was what I had to use on the k6 any newer won’t run.

On the core 2 I used windows 10

Ok, it seems I was a bit too optimistic on CL2 timings. While the board is stable in single threaded applications, it randomly crashes in Quake 3 with enabled SMP. So there's two option - SMP is just a much heavier abuse on RAM stability or two working CPUs somehow affect voltage drop on RAM.

Tried to boot with low profile X700 with 128-bit GDDR2 memory, didn't even start. Probably not enough power. Although I need to confirm that it works on PCIe system.
Another one is coming, so if I make it work, that should be ultimate low profile solution for Win9x.

Update: I've got another low-profile X700 and it works fine. Is it worth it? Eh, probably not. Apparently, it has only 4 active pixel pipelines and, strangely enough, all 6 vertex units active. So only mildly better than X550, mostly due to GDDR2. Something like a buffed Radeon 9600XT.

Now about Quake 3 SMP. I can't make it stable on ServerWorks chipset or there's something wrong with my Windows XP image - still randomly crashes even when I've tried ASUS TR-DLS. After 3 hours of Prime I conclude that CL2 is stable.

Two instances of Quake 3 without r_smp 1 also work just fine. Difference in PCI performance is actually astonishing under such conditions. I've got only 53 fps in Demo001 on normal PCI and 96 fps on dedicated PCI 66 slot, when another Quake 3 instance was running side-by-side with an active botmatch.

Update,
Next step will be to repeat numbers on a 370DE6 (again with adapters for Tualatin).
See if dual channel memory and higher PCI64 channels will have an impact.
Plus, the presence of AGP pro (2x) could tell me if there is any speed loss between pci at 66Mhz (with not shared bandwith) and the AGP.
Guess not at all or very minimal, as I cannot put anything powerful on that AGP 2x pro port.