Windows7ge

I chop those up and re-use them for patch cables between switches, patch panels, servers and end-clients. Whatever length I need it's nice to be able to just make one, two, or twenty... Good example:

Oh just a EPYC 7551P, 128GB DDR4, and 7 PCI_e x16 slots. Rigging a shroud from the front of this box with a 120mm fan wouldn't be hard but creating the necessary shape would put me through hell.

Is your bracket half-height or full height? I have my for sale store on the forum. If I remember correctly you were interested in 2.5" SAS drives but both you and I think it was @BondiBlue had unresolved money issues for how many you each wanted. If you want, look over my shop. We can probably work something out. So about 62% the performance for almost but not quite 33% the power draw for about 20% less money to buy. Still a better long term deal if you have the means to cool it.

If I asked you to send me your Tesla P4 @da na what would you like in exchange? With my wimpy 2 watt fan that's pretty quiet all considered and even the type one proto a Furmark burn-in wouldn't go over 75C and F@H wouldn't go over 65C. How is yours so hard to control where you couldn't use it for your rendering/encoding or whatever the task was you said? So this is a neutered 1080Ti? performance per watt there's a clear winner for F@H. Tesla P4 ~900K PPD, ~65W 1080Ti 2.2M PPD, ~275W from the wall. So in terms of pro's and cons the P4 draws about 25% the power for 40% of the performance coming in at about 50% the price of a 1080TI right now in a single slot form factor and a much much lower dB with the right fan setup. I could throw three of these in this box. Outperform a GTX 1080Ti and have it consume less power quieter. Hilarious.

Well this is disappointing but it tells us something valuable. Sun Oct 22 14:39:02 2023 +---------------------------------------------------------------------------------------+ | NVIDIA-SMI 535.104.06 Driver Version: 535.104.06 CUDA Version: N/A | |-----------------------------------------+----------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+======================+======================| | 0 Tesla P4 On | 00000000:22:00.0 Off | 0 | | N/A 75C P0 68W / 75W | 7519MiB / 7680MiB | 98% Default | | | | N/A | +-----------------------------------------+----------------------+----------------------+ This is with the new slope design. I'm not seeing a higher power utilization (like 75C target and boost until that temp) but I'm currently experimenting with F@H. It maxed out around 65C with the old model so to see if a lower load might show if the new slope is having any influence at all I ran this for about 20mins and came to the exact same results. 65C. I would expect even if the slope is not optimal that if it didn't hurt the performance it should at least show, I don't know. 1C difference? but here under two work loads no slope vs slope showed no difference at all. I can still explore trying to tighten the fit of the cooler more including adding that lip to the blocker so it channels a little better into the heatsink and we could explore internal baffles to try and re-direct more air down the center but improving the deign from here ins't off to a good start.

So not only do you need a vGPU license for each instance but you're supposed to have a different license depending on what you're using it for? Talk about you will own nothing and be happy. I've tested a B & Q profile. It's worked in both instances and the license claims I'm good until January. I guess this means should I explore bigger more powerful GPU's I might very well run into conflicts between the effectively hacked license/server and what the workload is I have it running.

In F@H how does one know if GPU tasks are failing? BOINC has an event log that tells you everything but I'm a little lost with F@H. It says in 14hrs Ive completed 10WU but I don't know if those were CPU only. Right now I'm running on profile GRID P4-8Q. F@H website says I should be around 182K PPWU, 704K PPD. Right now I'm only seeing around 147K estimate PPWU for GPU but I'm seeing an estimate of 949K PPD. So lower than average for WU but higher than average for PPD. Hard to read if that's good or if I should expect higher. If I had an A40 I might ask so I could see what performance I'm supposed to be getting. It appears the Linux performance is supposed to be just a little higher than Windows so I think I'm gonna go ahead and try to set that up.

Just BOINC or F@H too?

I'm not working with the best software for creating slopes. It was basically done by hand using circles. Did the best I could for a v2. Less you happen to get that software working you were talking about. Tomarrow I'll have the time to swap this in and see what difference it makes. It will at least tell us if we're on the right track or wasting our time. As for diffuser we know the hot spot is in the middle. Wouldn't be out of the question to try and channel more air towords the center. I might also still be able to tighten up the fit so air isn't flug around as soon as it reaches the lip of the card. Really direct the air into the fins could give us another 1 or 2 degrees.

Could be right. Those projects may have just ran their course. About that... This was when I tried connecting to them tonight. They were my first go-to. I guess not anymore. I don't know what happen to them. They split from IBM and joined some other agency? Took them forever and now this is what I come back to.

Whooo! Utilization and temps are both looking pretty good on the host hypervisor. +---------------------------------------------------------------------------------------+ | NVIDIA-SMI 535.104.06 Driver Version: 535.104.06 CUDA Version: N/A | |-----------------------------------------+----------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+======================+======================| | 0 Tesla P4 On | 00000000:22:00.0 Off | 0 | | N/A 67C P0 53W / 75W | 7519MiB / 7680MiB | 94% Default | | | | N/A | +-----------------------------------------+----------------------+----------------------+ +---------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=======================================================================================| | 0 N/A N/A 2345908 C+G vgpu 7488MiB | +---------------------------------------------------------------------------------------+ I think I was talking to you Levent when you brought up that after pass-through you had a GPU folding fine but you couldn't really get BOINC going right. It might be worthy of note @leadeater I've been hearing that BOINC as a whole is falling. I tried about six different previously popular projects trying to get a GPU load going. Could not and at least two of the six were offline/out of commission but they were a big part of the pentathlon years ago. The future for BOINC is looking kinda bleak. May be time I finally transition over to folding. Would be nice to earn at least the contributor badge with my tiny P4. I am also going to have to wind up a Ubuntu Server VM and see if I can get F@H running without a UI. If I were to take this P4 and split it in half I'd be curious to see what if any performance difference shows between the two. For compute I expect Linux to edge ahead.

Software is being buggy so don't ask me how the .STL is going to look for you but the slicer will chop it up into gcode all the same. It happens. I anticipate that ever slight point will disappear in the actual print. Even if it doesn't I can probably cut it off with an exacto-knife. tesla_p4_blower_proto_v2.stl

Not perfect but what do ya'll think?

Given it depends on the workload. I plan to experiment later with actual compute tasks on Windows and I wanna try Linux but for now this will suffice to find the best duct design. Noise level is low outside a faint whine that comes with tiny fans. Full tilt it only draws 2.16W. It's not an incredibly powerful fan. Really good performance/noise balance.

@da na @FloRolf @Schnoz Sorry for the wait. NVIDIA is a bitch. I had to trick the card into thinking there was a vGPU License Server running on my network and I only just got it at least working enough to start performing tests. Without an activation server. This happens: That 20min degradation was the real bitch. So with that out of the way everything else went pretty smoothly. I created one VM with 8GB VRAM at 1920x1080 and ran furmark, the fuzzy donut. After 30 minutes this was the result with the first iteration prototype. Sat Oct 21 18:09:41 2023 +---------------------------------------------------------------------------------------+ | NVIDIA-SMI 535.104.06 Driver Version: 535.104.06 CUDA Version: N/A | |-----------------------------------------+----------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+======================+======================| | 0 Tesla P4 On | 00000000:22:00.0 Off | 0 | | N/A 75C P0 65W / 75W | 7519MiB / 7680MiB | 97% Default | | | | N/A | +-----------------------------------------+----------------------+----------------------+ +---------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=======================================================================================| | 0 N/A N/A 2176591 C+G vgpu 7488MiB | +---------------------------------------------------------------------------------------+ 97% utilization ~65W from the board and 75C. My evening plans got cancelled so I might get to doing a re-design tonight.

Why is it AFTER I contact someone that I figure it out. PS C:\Windows\system32> & 'nvidia-smi' -q | Select-String "License" vGPU Software Licensed Product License Status : Licensed (Expiry: 2024-1-19 21:4:37 GMT) There was a configuration file I created at step Create config file and I was supposed to fix it to suit my setup. Because I'm working with a fanless GPU Da na, Florolf, and Shnoz are interested in helping me overclock an airflow guide for optimal performance. While i burn-in the GPU I'll let you know if it starts mis-behaving but I expect the issues shouldn't go beyond fps limiter if it loses contact with the DLS for over 24hrs.

So far the only windows side issues I'm seeing are the ones Leadeater outlined. After 20mins I'm capped at 15FPS for anything load intensive.

Well I'm confused because it looks as though there's two things to execute: docker run -e TZ=EST -e DLS_URL=`192.168.0.163 -i` -e DLS_PORT=443 -p 443:443 -v $WORKING_DIR:/app/cert -v dls-db:/app/database collinwebdesigns/fastapi-dls:latest sudo -u www-data /opt/fastapi-dls/venv/bin/uvicorn main:app --app-dir=/opt/fastapi-dls/app First one throws an error saying the service is already running: docker: Error response from daemon: driver failed programming external connectivity on endpoint intelligent_khorana (67abf4e4000a5f4b49ab3b1626c6c79646caf407f1a8a77af8895cff79b1a07c): Error starting userland proxy: listen tcp4 0.0.0.0:443: bind: address already in use. ERRO[0000] error waiting for container: Second one looks half normal except it's running on localhost and I see no option to set it to another interface: INFO: Started server process [576] INFO: Waiting for application startup. INFO:main: Using timezone: EDT. Make sure this is correct and match your clients! Your clients renew their license every 13 days, 12:00:00. If the renewal fails, the license is 90 days, 0:00:00 valid. Your client-token file (.tok) is valid for relativedelta(years=+12). INFO: Application startup complete. INFO: Uvicorn running on http://127.0.0.1:8000 (Press CTRL+C to quit)

I made it to the end where I try to curl.exe the token over from the server but it claims it can't reach my server. I can ping the license server just fine so it has to be a service configuration error.

Alright, I might enlist your assistance @Levent here's where we're at. Hypervisor driver sees the card: +---------------------------------------------------------------------------------------+ | NVIDIA-SMI 535.104.06 Driver Version: 535.104.06 CUDA Version: N/A | |-----------------------------------------+----------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+======================+======================| | 0 Tesla P4 On | 00000000:22:00.0 Off | 0 | | N/A 66C P0 55W / 75W | 7519MiB / 7680MiB | 62% Default | | | | N/A | +-----------------------------------------+----------------------+----------------------+ +---------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=======================================================================================| | 0 N/A N/A 1928341 C+G vgpu 7488MiB | +---------------------------------------------------------------------------------------+ We have vGPU profiles: 0000:22:00.0 nvidia-157 Available instances: 0 Device API: vfio-pci Name: GRID P4-2B Description: num_heads=4, frl_config=45, framebuffer=2048M, max_resolution=5120x2880, max_instance=4 nvidia-214 Available instances: 0 Device API: vfio-pci Name: GRID P4-2B4 Description: num_heads=4, frl_config=45, framebuffer=2048M, max_resolution=5120x2880, max_instance=4 nvidia-243 Available instances: 0 Device API: vfio-pci Name: GRID P4-1B4 Description: num_heads=4, frl_config=45, framebuffer=1024M, max_resolution=5120x2880, max_instance=8 nvidia-63 Available instances: 0 Device API: vfio-pci Name: GRID P4-1Q Description: num_heads=4, frl_config=60, framebuffer=1024M, max_resolution=5120x2880, max_instance=8 nvidia-64 Available instances: 0 Device API: vfio-pci Name: GRID P4-2Q Description: num_heads=4, frl_config=60, framebuffer=2048M, max_resolution=7680x4320, max_instance=4 nvidia-65 Available instances: 0 Device API: vfio-pci Name: GRID P4-4Q Description: num_heads=4, frl_config=60, framebuffer=4096M, max_resolution=7680x4320, max_instance=2 nvidia-66 Available instances: 0 Device API: vfio-pci Name: GRID P4-8Q Description: num_heads=4, frl_config=60, framebuffer=8192M, max_resolution=7680x4320, max_instance=1 nvidia-67 Available instances: 0 Device API: vfio-pci Name: GRID P4-1A Description: num_heads=1, frl_config=60, framebuffer=1024M, max_resolution=1280x1024, max_instance=8 nvidia-68 Available instances: 0 Device API: vfio-pci Name: GRID P4-2A Description: num_heads=1, frl_config=60, framebuffer=2048M, max_resolution=1280x1024, max_instance=4 nvidia-69 Available instances: 0 Device API: vfio-pci Name: GRID P4-4A Description: num_heads=1, frl_config=60, framebuffer=4096M, max_resolution=1280x1024, max_instance=2 nvidia-70 Available instances: 0 Device API: vfio-pci Name: GRID P4-8A Description: num_heads=1, frl_config=60, framebuffer=8192M, max_resolution=1280x1024, max_instance=1 nvidia-71 Available instances: 0 Device API: vfio-pci Name: GRID P4-1B Description: num_heads=4, frl_config=45, framebuffer=1024M, max_resolution=5120x2880, max_instance=8 We can assign a vGPU to a VM, start the VM, install the drivers, use OpenGL, and PARSEC/RDP into the VM. Fantastic. Here's where the problem starts. I am not understanding the licensing server setup at all. He wrote fucking several cliff hanger moments then jumps ahead with no further explanation as to how he got to the next step or what the next step is cause he wrote it to work with like 5 different platforms and just just did a bad job of organization and presentation. Think between the two of us (my hardware, your brain) we can figure it out? Cause right now the Windows VM isn't happy that it has no vGPU license. I'm also looking around to see if the Author offers any kind of support but outside of a Discord for the parent article I'm not seeing anything promising.