SXM2 over PCIe
- Thread starter CyklonDX
- Start date
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
Yeah, luckily that is in the PDF so it shouldn't need reverse engineering at least... Though easy to get wrong and blow everything up!
Not sure if this is the right place for this, but I have two SXM2 PCIE adapters, two copper heatsinks, and two blower fans that I'm looking to part with. I also have two SXM2 P100s. I bought them from a Chinese seller last year with the intention of testing them with the P100s before buying V100s but life had other plans for me. They've been sitting in a closet for over a year now.
The whole shebang cost me around 500€ with shipping and import taxes. Looking to sell everything at cost, and I'll throw in free shipping within the EU. I'm in Germany. PM me if interested.
The whole shebang cost me around 500€ with shipping and import taxes. Looking to sell everything at cost, and I'll throw in free shipping within the EU. I'm in Germany. PM me if interested.
Nvidia Tesla V100 32GB SMX3 adapter board from ebay for $200. Doesn't support 16GB SMX2 version, it seems.
Does anyone has any experience with it?
Update (10/27/2024): Probably not worth it with SMX3 adapter approach as the total cost including SMX3 module, this adapter, heatsink and pcie riser cable might come close to $750-$800 range. Just saw in ebay that the 32GB pcie card can be had for $1000.
Does anyone has any experience with it?
Update (10/27/2024): Probably not worth it with SMX3 adapter approach as the total cost including SMX3 module, this adapter, heatsink and pcie riser cable might come close to $750-$800 range. Just saw in ebay that the 32GB pcie card can be had for $1000.
Last edited:
Hello. I'm having issues with the AOM-SXMV on Windows. I haven't tried Linux, but I intend to mainly use Windows anyway. If Linux can give better info, installing any distro is a nonissue. Sorry for the below info dump.
Description:
All four cards are connected and detected, and I can run CUDA (llama-cpp-python) and even 3D applications. But for some reason, I have to severely underclock else the driver crashes (recently the driver restarts on its own without a reboot, perhaps it's the newer drivers), and by "severely" I mean to about 660 MHz. Sometimes I can run at 840 and even 930 MHz, but it's extremely unstable and tends to crash within a minute. It's always under high load; i.e. a simple game can run at 200 fps just fine but a demanding game at 30 fps crashes, 30% usage at 1312 MHz is stable. Enc and dec can run at 1530 MHz, though sometimes crashes when stress tested for hours. CUDA and tensor cores—the power pullers—seem to be the main issue. The LEDs don't change at all (2-4-5 6-8-9 constantly on even during and after crashes).
This is tested with 1x, 2x, 3x, and 4x cards connected. I've also tried forcing x8x8 bifurcation (and forcing native x16) to no avail. All tested driver versions and earlier Windows 10/11 versions (i.e. 21H2 and Tesla driver 427.60) give the same outcome. I only have one mobo (01LM605) and one AOM-SXMV available to test. I am using the dual AOM-SXMV-PIO cables listed on Xianyu.
The temperatures seem safe (hot spot < 80C, GPU and memory < 60C, PLX feels warm to the touch). The voltage is a little concerning; I've tried multiple PSUs and cables—all tested to be able to push 12V 21A—but the reported HWiNFO rail voltage stays relatively low under stable load (~11.3V). Barring these may-be symptoms, I can't seem to find any possible causes for this. Been racking my head for over a month now, worried my board's faulty.
Screenshots:
forced x8x8 @ 660 MHz
GPU PCI bus is 136 (88:00.00)
Description:
All four cards are connected and detected, and I can run CUDA (llama-cpp-python) and even 3D applications. But for some reason, I have to severely underclock else the driver crashes (recently the driver restarts on its own without a reboot, perhaps it's the newer drivers), and by "severely" I mean to about 660 MHz. Sometimes I can run at 840 and even 930 MHz, but it's extremely unstable and tends to crash within a minute. It's always under high load; i.e. a simple game can run at 200 fps just fine but a demanding game at 30 fps crashes, 30% usage at 1312 MHz is stable. Enc and dec can run at 1530 MHz, though sometimes crashes when stress tested for hours. CUDA and tensor cores—the power pullers—seem to be the main issue. The LEDs don't change at all (2-4-5 6-8-9 constantly on even during and after crashes).
This is tested with 1x, 2x, 3x, and 4x cards connected. I've also tried forcing x8x8 bifurcation (and forcing native x16) to no avail. All tested driver versions and earlier Windows 10/11 versions (i.e. 21H2 and Tesla driver 427.60) give the same outcome. I only have one mobo (01LM605) and one AOM-SXMV available to test. I am using the dual AOM-SXMV-PIO cables listed on Xianyu.
The temperatures seem safe (hot spot < 80C, GPU and memory < 60C, PLX feels warm to the touch). The voltage is a little concerning; I've tried multiple PSUs and cables—all tested to be able to push 12V 21A—but the reported HWiNFO rail voltage stays relatively low under stable load (~11.3V). Barring these may-be symptoms, I can't seem to find any possible causes for this. Been racking my head for over a month now, worried my board's faulty.
Screenshots:
forced x8x8 @ 660 MHzGPU PCI bus is 136 (88:00.00)You are a godsend. The cables I was using were overheating, lowering the voltage. Apparently the whole pack was cheaply made. Thanks so much man.
there are plenty of pcie sxm2 adapters on ebay from china (if you don't need 4 v100's they are decent options)
// you can combine those as i hear with A100 automotive sxm2 card.
(here's gpu-z output from one of the a100 automotive cards - its not as strong as proper a100, but stronger than V100)
*but if you do not need tensor flops you would be much better with normal consumer gpus that are much faster. I also doubt you'll be able to get sr-iov on this a100. (v100 32G is decent option for sr-iov ~> but titan rtx might be even better option ~> its also cheaper, and it doesn't have messy secured bios so its possible to upgrade its memory to at least 48G - but price wise it may come down to actual price of Quadro RTX 6000 - as its literally the same card ~> just Quadro been printed by Korea Samsung if i recall and titan in Taiwan TSMC; alternatively if you don't care about it - A5000 will be around same price as rtx quadro 6000, and offer better performance just 24G of vram is a con.)
// you can combine those as i hear with A100 automotive sxm2 card.
(here's gpu-z output from one of the a100 automotive cards - its not as strong as proper a100, but stronger than V100)
*but if you do not need tensor flops you would be much better with normal consumer gpus that are much faster. I also doubt you'll be able to get sr-iov on this a100. (v100 32G is decent option for sr-iov ~> but titan rtx might be even better option ~> its also cheaper, and it doesn't have messy secured bios so its possible to upgrade its memory to at least 48G - but price wise it may come down to actual price of Quadro RTX 6000 - as its literally the same card ~> just Quadro been printed by Korea Samsung if i recall and titan in Taiwan TSMC; alternatively if you don't care about it - A5000 will be around same price as rtx quadro 6000, and offer better performance just 24G of vram is a con.)
Last edited:
I have the drive a100’s on my desk and I’m going to continue searching for an aom-sxmv/aom-sxm2. I’m 90% sure it’ll work. Also the sketchy sxm2 to pcie do work with them.
I don’t know the name or if it has a specific model name. If you search Taobao and/or XianYu for AOM-SXMV you should find it or something very similar
Just joined. I hope I'm ok to post here. I bought a 32gb sxm2 v100 and one if them 60 dollar adapter boards. It came in today. I've been pulling my hair out all night trying to get it to work. Thinking did I break it, am I an idiot. I'm a bit of a tech hoarder so I've had it in several machines. Finally it's showing up after swapping my x16 riser for an x8 riser. Any idea why? I know the x16 riser works.