megatron-uk wrote on 2023-11-14, 08:42:

Just bought a pair of Nvidial Tesla P4 accelerators to go in my Dell T430 virtualization box.

They are roughly equivalent to GTX 1080, but lower power (50-75w) and no video out.

Plan on using them in my HPC-cluster-in-a-box setup, specifically to test out OpenOndemand - i.e. for bookable GPU desktops on our HPC facility, which is normally only batch mode compute.

Not really.

Half the performance by *underclocking* of GTX 1080 with CUDA cores of 2560. True equivalent is GTX 1060 in performance and wastes the 256bits in bandwidth. And hellish of cheaper just to buy two 1060 or 1070 and go with it.

The correct one would be equivalent is P40, comparable to GTX 1080 in performance and P40 comes with 384 bits of 24GB vram. Price reflects this around 200-230 CDN and similar to GTX 1080 in price and watts. Cheaper than Titan XP.

pentiumspeed wrote on 2023-11-16, 00:26:

Not really. […]
Show full quote

megatron-uk wrote on 2023-11-14, 08:42:

Just bought a pair of Nvidial Tesla P4 accelerators to go in my Dell T430 virtualization box.

They are roughly equivalent to GTX 1080, but lower power (50-75w) and no video out.

Plan on using them in my HPC-cluster-in-a-box setup, specifically to test out OpenOndemand - i.e. for bookable GPU desktops on our HPC facility, which is normally only batch mode compute.

Not really.

Half the performance by *underclocking* of GTX 1080 with CUDA cores of 2560. True equivalent is GTX 1060 in performance and wastes the 256bits in bandwidth. And hellish of cheaper just to buy two 1060 or 1070 and go with it.

The correct one would be equivalent is P40, comparable to GTX 1080 in performance and P40 comes with 384 bits of 24GB vram. Price reflects this around 200-230 CDN and similar to GTX 1080 in price and watts. Cheaper than Titan XP.

Unfortunately I don't have the space - the x16 slots in the T430 are right next to each other, so there's room for one double width card, or two single slot cards (since the x8 slot for the disk controller is also next to one of the x16 slots), and I'm not interested in using vGPU to virtualise/slice up a single card. Also there's power consumption to factor in; anything 1050Ti and upwards is going to have additional power draw over and above the supply from the motherboard. The only other real alternative is something like the single slot Quadro P600/620/1000 ... but compared to the Tesla P4, or even a basic GTX 1050, that would be more money for even less performance.

Ultimately, this isn't for actual compute performance - but to replicate what I hope to do on a new multi-million £ HPC facility that my institution is bringing in - bookable GPU desktops, use of HPC facilities for non-command line users (e.g. Matlab or Ansys backed by HPC compute nodes, but accessed simply via browser using tooling like https://openondemand.org/), Jupyter Notebook farm, etc... and to keep my hands up to date on more recent tooling for Slurm (our resource scheduler of choice) to give me more options in the future.

We're likely to actually go for L40S cards in the production facility (A100 is dead and H100 has insane lead-times of up to a year, according to our big suppliers - as well as insane pricing of ~£30K per card). Looking at the actual use cases of our research community (combined medical imaging, genetic sequencing, a little CFD and some AI/ML), it appears most won't be too compromised by the lack of fp64 and nvlink/hbm.

Sorry, wrong thread - I've deleted my post here and moved them to DOSBox (DOS-Box general items)

megatron-uk wrote on 2023-11-16, 09:41:

Unfortunately I don't have the space - the x16 slots in the T430 are right next to each other, so there's room for one double wi […]
Show full quote

pentiumspeed wrote on 2023-11-16, 00:26:

Not really. […]
Show full quote

megatron-uk wrote on 2023-11-14, 08:42:

Just bought a pair of Nvidial Tesla P4 accelerators to go in my Dell T430 virtualization box.

They are roughly equivalent to GTX 1080, but lower power (50-75w) and no video out.

Plan on using them in my HPC-cluster-in-a-box setup, specifically to test out OpenOndemand - i.e. for bookable GPU desktops on our HPC facility, which is normally only batch mode compute.

Not really.

Half the performance by *underclocking* of GTX 1080 with CUDA cores of 2560. True equivalent is GTX 1060 in performance and wastes the 256bits in bandwidth. And hellish of cheaper just to buy two 1060 or 1070 and go with it.

The correct one would be equivalent is P40, comparable to GTX 1080 in performance and P40 comes with 384 bits of 24GB vram. Price reflects this around 200-230 CDN and similar to GTX 1080 in price and watts. Cheaper than Titan XP.

Unfortunately I don't have the space - the x16 slots in the T430 are right next to each other, so there's room for one double width card, or two single slot cards (since the x8 slot for the disk controller is also next to one of the x16 slots), and I'm not interested in using vGPU to virtualise/slice up a single card. Also there's power consumption to factor in; anything 1050Ti and upwards is going to have additional power draw over and above the supply from the motherboard. The only other real alternative is something like the single slot Quadro P600/620/1000 ... but compared to the Tesla P4, or even a basic GTX 1050, that would be more money for even less performance.

t430.jpg

Ultimately, this isn't for actual compute performance - but to replicate what I hope to do on a new multi-million £ HPC facility that my institution is bringing in - bookable GPU desktops, use of HPC facilities for non-command line users (e.g. Matlab or Ansys backed by HPC compute nodes, but accessed simply via browser using tooling like https://openondemand.org/), Jupyter Notebook farm, etc... and to keep my hands up to date on more recent tooling for Slurm (our resource scheduler of choice) to give me more options in the future.

We're likely to actually go for L40S cards in the production facility (A100 is dead and H100 has insane lead-times of up to a year, according to our big suppliers - as well as insane pricing of ~£30K per card). Looking at the actual use cases of our research community (combined medical imaging, genetic sequencing, a little CFD and some AI/ML), it appears most won't be too compromised by the lack of fp64 and nvlink/hbm.

Wanted to know the background:
Is T430 what you got from dept or bought by yourself? This machine is made for storage duty with some processing back-end but poor on slots or PSU capacity.

Actually the appropriate workstation is Dell Precision T7910. Plenty of x16 slots to do dual P40 or dual P100 and dual sockets for V3 processors.

Cheers,

Bought by myself. My employer, or at least the senior management of the IT unit within my institution keep going on about staff development... But it's all just hot air, they are more than happy for everyone to stay in the little boxes that they recently put everyone in.

If I want to keep myself up to date (and I've been a Unix administrator, middleware developer, and infrastructure architect for some 25 years), then it's up to me to do it.

Main purpose of the T430 is as a smallish VMWare install - most of the VM images are to replicate the supporting infrastructure of an HPC environment (DNS and pxe imaging, LDAP users/groups/sudoers, nfs homes, slurm accounting etc, all automated and managed by Puppet and/or Ubuntu's new cloud-init autoinstall mechanism).

I just swapped a RX 580 8gb to a 6gb Nvidia 1060. From one FX 8350 machine to another. I remember why I hate Nvidia software.

Pulled batteries from my Tandy 3000HL's

I'm buying batteries from Digikey because grocery stores want $6.99 for a single coin/button cell of any size (and $11.49 for a 2-pack). I don't know how much food is supposed to cost, but I wouldn't be surprised if they were also putting a 1500% markup on essential stuff. Of course, I have to pay for shipping, but that's $6.99 (ground shipping only). I can buy two batteries and be ahead, but I'll get more than that.

I see that Digikey sells Omron D2FC mouse switches now (even a 60 million cycle version). The Omron datasheet lists both the minimum and maximum current as 1mA, which is weird. I don't need any (I have some Omron D2F-01F switches that I haven't even tested yet), but I thought I would let someone know that there's a reputable source for those low-current microswitches now.

I am mostly just grabbing dollar store CR-2032s at present .. mostly they are okay, about 1 in 5 is too low to use. The problem with these is, they age in the pack too, so unless you get them from somewhere with decent turnover, the "good" brands can be stale and the grocery stores just look at you blank if you say the battery you just bought is dead, "Yeah dude, you used it up, that's how batteries work." so yeah..... hard to find a place where it's worth buying the "good" ones. Some independent computer stores keep a supply at a fair price.

Been a grumpy old man; last night after navy duty I decided to mess around with one of my computers; the ITX build that is supposed to be my Linux PC but currently also function as my Windows PC as I am kind og lacking the funds for the build I wand to do (purely astetics).

Two main issues with this; firstly it is build in a case with a few proprietary cables and pinouts, and mainly the front panel board uses a floppy-style connector which is different wired thus require the original power supply unit, else USB 2, LED and switches wont work, this I had originally made a workaround for as I had taken what I had (a broken Intel fan, stripped one guard from the PWM-socket, and soldered the cables to a power extender cable to avoid risking anything with the motherboard socket or the current PSU) and then USB 2 worked (without the cable it simply always said "overcurrent - system will shutdown in 15 seconds") - one day it started failing again: the connector snapped during transport. I've soldered it back on but the computer won't turn on with it attached. Here is the funny thing: somehow the front panel works fine now (not Power LED - but that may be a faulty board or a wrong pinout due to proprietary/motherboard missmatch) - I can boot even with something in the USB 2 ports (USB 3 always worked, and power switch worked fine too because that was swapped (badly) by previous owner due to some fault). I still have to figure out how to get the power working for the board - but maybe in the end I don't really care and just leave it as it is.

Now the real issue is that my MSI GeForce GTX 1650 D6 AERO ITX OCV1 (short and precise!) suddenly is not recognized by the computer - at all! Not in Windows and not in post. If I used the HDMI cable in the MSI H110I PRO motherboard I can post and loginto Windows where the GTX 1650 is not even shown on devices. The fan do spin up, and I read that a lot of other users of various cards have this issue too.
So apart from having to tinker some stupid cable, I also need to mess around with a "lost" graphics card...
I have a little project PC so I might try to plug it into my Gigabyte B450M DS3H which I still have no idea what to use for at the moment. This is currently housing a Gigabyte GeForce GTX 1650 OC 4G - so more or less same card just a full size and more power hungry. But first I will try to install some Windows updates (I haven't had that computer on for a few months until a few days ago when I then discovered this), and if that is not enough, I will see if I need to update the UEFI BIOS.
Nvidia drivers can't be updates as there is no card installed (according to the driver software and GeForce Expirence) - so let's see what will happen... I kinda enjoyed my little ITX build so I would be a bit pissed if I need to get a new graphics card since it was working absolutely perfectly last time I used it 2½ months ago.

This have postponed my button box project a little bit; I dismantled my Saitek Cyborg Evo to reporpose the boards to make a button box: The stick had gotten very sticky in it's movement and had no precision left, one key cap had fallen off and it had that sticky feeling. I have bought a cheap but very well preserved Mad Cats Cyborg v.1 - it's older but works really smoothly, so until I finally can get my hands (on the money to get my hands on) a Warthog flight stick, I will use the Cyborg v.1 - and why waste perfectly good switches on e-waste when I could much rather use them in a small box on my desk for flight simulators and a dashboard panel for Euro Truck Simulator 2? Although I usually change the keyboard layout a bit when playing ETS2/ATS so Q and E are my left and right turn indicators, a little panel with a real car engine button and some switches would make it a real treat - especally if I can find a wheel that I like, then it will be a blast to haul some goods! If I manage to also make an exchangable mask for the box then I can easily have a fixed setup for flight sims so so all buttons are the same location because the box says so with pictograms, and one for driving simulators - because the buttons says so. And an "Engine Start" can be used both on a plane/helo and a car/truck/tractor!
Still need to figure out what to do with the potentiometers...
Trigger will probably by rewired to my Engine switch, and the hat switch will be kept as it is and moved into the box and being used as dedicated cruise control in ETS2/ATS - Up/Down are Reset/Set and Left/Right -/+ like on my Camry (makes it easy for me to remember).

But first the tiny dragon, then I can start my puzzle - and my Deskpro, and my Gigabyte B450M DS3H, and my weird old Asus, and my Intel-project, and my old laptop project... and so far.

I bought a Minisforum EM680 a few weeks ago for £375. It has a ATI 6800U APU, 16GB DDR5 6400MHz dual channel RAM and 512GB SSD. it's about the size of folded wallet.

It's both incredible and disappointing at the same time. It's completely fast enough for any day-to-day computing whilst being very small and surprisingly quiet. It came with Windows 11 pre-installed (which I re-installed anyway) and doesn't go over 65c. It can do some modern gaming too at 30fps and plays games prior to 2013 max settings kinda thing.

On the downside the gaming performance is a bit lacking for a modern "high end" mobile APU, I think I had too high expectations. The PC has a 30watt limit, comes with a 65watt power brick and doesn't go over 65c. The highest watts Ive seen is 28watts (using MSI afterburner) I wish there was a bios option to increase the TDP of the APU to 40watts as I think it can do more than its set up to do. The HDMI cable/port output doesn't seem to support my TV and sometimes does ghosting on a 5:4 monitor when connected via a HDMI to DVI cable (no other graphics card does this for me). I wish it had an ethernet port but then it is small.

It's great for player older games with a small footprint, quietly.

Got some life in the small MSI computer again!

First I tried to let Windows Update do it's thing, since it had a failed update - after this there were still issues, so I updated the BIOS to the newest version, this also didn't solve the issue. I noticed that my RAM was halved, and I saw that one stick was actually not seated properly, not sure how that happened, but installed it again, checked the graphics card again just to be sure that it was not loose (still not loose, so all ok). Then I removed it, and plugged it into my Gigabyte build which shown me that the graphics card worked fine.
Then I installed the (similar) Gigabyte card in the MSI which also gave screen like it should.

I was puzzled! Putting everthing back together again - and black screen on MSI/MSI combo. This time it was my fault: After reinstalling the card I forgot to plug in the cable. But after this it started up, loaded Windows and began to install drivers.

What the actual freak?! I even before had tried to just reseat the card without any luck - but a full swap and a swap back again did the trick... seriously?

After this it was just hanging out working on something little older before bedtime.

ODwilly wrote on 2023-11-18, 09:48:

I just swapped a RX 580 8gb to a 6gb Nvidia 1060. From one FX 8350 machine to another. I remember why I hate Nvidia software.

That's backwards, downgraded. RX 580 is high end 8GB video card approximately GTX 1080Ti.

Cheers,

Sis's gaming build needed a storage upgrade for the quite old Steam Games set (as she started running out of space on it), so the old Team 1TB SSD (Vulcan Z T253TZ001T0C101) was changed to a Micron 2TB SSD (MTFDDAK2T0TDL), and the separate Downloads drive was changed from a 1TB Seagate HDD (ST1000LM035-1RK172) to the original Team 1TB SSD. As it stands, the build no longer uses any mechanical storage.

I now claim the Seagate drive for additional storage.

The upcoming high-end AMD APUs are absolute beasts for 1080P gaming and media use, because they finally incorporate AMD's latest GPU and CPU tech. On older titles from the early 2000's to early 2010's, you will be able to max out the graphics settings with no performance issues, except when driver/software bugs pop up. I am very much looking forward to the mini-PC products that will come with those and have already started to plan my builds. The only thing I'm really not looking forward to is paying a ton of $$$$$$$$ for high-capacity M.2 SSD storage, as the mini-PC market - at the smaller size segment and the gaming segment as well - seems to be moving away from 2.5-inch SATA in the current and upcoming designs and 8TB SATA SSDs are easy and inexpensive to come by, at around 1/3rd the price of an M.2 equivalent.

Been claiming codes from Boing Boing store, I saw their discounts and weakened, think they are still on until tonight if you wanna look.

Anyhoo, office 2019 for wifey installed and registered okay. Other stuff went through fine also.

Thing I'm wondering about though is the Windows 11 Pro for 2 machines one I got, is there any way to "park" it in your Microsoft account, it seems registered windoze pops up as devices in your account, but pre-install is there somewhere to save it in there, like you seem to be able to save other products for windows codes?

I don't see on the boing boing page that it says it expires like some of the other codes unless registered in 30 days, the claiming the code from the purchase seemed to be the time limited bit and I've got that. I just don't want to put it on a machine quite yet, I have machines to mess with that should theoretically manage to get 11 codes from their 10 codes, but the hardware with no license, which might require workarounds for install, is sitting unbuilt and may not get to it for a while. So yah, just wanted to make sure it validated with MS without installing or whatever you can do.

pentiumspeed wrote on 2023-11-22, 21:35:

ODwilly wrote on 2023-11-18, 09:48:

I just swapped a RX 580 8gb to a 6gb Nvidia 1060. From one FX 8350 machine to another. I remember why I hate Nvidia software.

That's backwards, downgraded. RX 580 is high end 8GB video card approximately GTX 1080Ti.

Cheers,

Well they are similar systems ( FX 8350, 990fx, SSD, 16gb of ram) and the RX 580 is nowhere near a 1080ti. Maybe a 1070 or 1660 super. The 1080ti beats it by 30% in Metro Exodus for example in 1080p.

My dad just hated the Nvidia software so I swapped him cards. Since this is just a backup PC. Oddly enough HDR wasn't working on his 4k TV on the 1060 while on the 580 it works perfectly. Honestly it seems like the Nvidia cards and drivers just don't use the CPU as well as the AMD cards. CPU core usage + GPU usage with the Nvidia card is all over the place while with the AMD card the FX is usually hovering around 90% load while gaming.

Updated my Windows VMs and upgrading Ubuntu VM to 23.10

Writing more pages for a novel I'm writing for fun.

Finding bulbs 3000K-350oK with CRI of 90-95 between 15W to 20W is very frustrating. Either hard to find, expensive from brands or dubious chinese and some solutions have to put together one. Or wrong one, like 4000K with cri of 80.

I have a strips of loose 5730 SMD LED with CRI of 95 at 3500K by Osram were so good, I can't wait for chinese made LED on PCBs to arrive and mod them with better LEDs.

Philips only sells 2700K bulbs warm white that my mom calls them yellow light, yeah, I know with a smile, and it's okay but for me needs better than that but these are usually CRI of 80 or less. And the mom's sight is not that great needs so bright light at daylight white is too stark too white, so cold and color rendering is terrible in our kitchen. Also Philips only have them available in stores that is bit under wattage than reality. Not quite similar to 100W incandescent, more like 80W because phililps has them in 14.5W and the other 60W equivalent is way less than actual.

Chinese sellers of LEDs for 5050 and 5730/5630, 2835 package are designed to source 60-150mA large die LEDs but as expected, these types you can get from chinese are fakes. 20mA LED dies in a larger packages being over-driven to 60mA at most because the die size is tiny as grain of sand. Osram I had got has real 150mA, larger rectangular LED die. Color LEDs in large dies is nearly impossible especially with chinese sellers. I have not seen any SMD LEDs with two or three LEDs in parallel especially from chinese sources either.

Safest to get in larger LED dies, easiest as well, is ones you can get is star LEDs at 3W-5W and does have bit of copper slug under them on a star PCB, but these are very old technology that was developed 20 years ago or more, was originally Lumileds's for luxeon star design, and lens attached is fake on chinese copies, it is plastic hollow dome, real luxeon stars are flat or true optical dome. And designing PCB to use them is difficult to do due to star large overall package requirements, SMD LED packages had developed further in last decade that you can get high watts LED in compact SMD LEDs and cheaper to manufacture, easier to choose any optics, can be high density by packing these more tightly.

Another is safest way to get large die LEDs from chinese sellers is 10W using real copper *plate* back in 3x3 die chips format. Yes you can under drive them too. Available in 20, 30, 50 and 100W. For white LED, the rendering is poor that limits your choices there too but fantastic choice in red, green, blue, yellow, orange and amber. Also RGB too.
Problem is you have to package it, finding optics substitutes is hard to do and making the housing to hold loose bits and sink the heat at same time.

Related to power stuff, first time bought a chinese 300W 60V 5A bench power supply. Needs one for my future electronics experiments and making circuit boards. Couldn't find one that is within my budget and voltage needed even used. Some of my LEDs is 100W and needs at least 36V for evaluating them. No, not at full brightness, I'm going to under drive them once I know how bright they can be in under driven mode.

Majority of light bulbs uses 2835 package.

Cheers,