This is a bit of a long shot and open ended. I hope this fits into the category well enough. I’m looking for some help sourcing desktop hardware on a limited budget. The goal is to get as close to a workstation as possible.
To give some context I’m a PhD student in Canada. My university supplied desktop has a failing motherboard that likes to turn off the system when it’s most inconvenient. Due to limited funding a replacement won’t happen for a year or more. In the mean time I’ve been using my personal laptop. Unfortunately it struggles thermally despite a repaste, some undervolting and plenty of verbal encouragement.
My research revolves around numerical analysis, specifically my area is numerical relativity. So it’s helpful to have a moderate amount of local system resources for developing and testing my code. My personal system at home is fine. I have access to a HPC cluster to run my full simulations. So I don’t strictly need some unholy HEDT monstrosity born of 99 forbidden industry partnerships for my office.
I do need something better than I have however, and a replacement for that desktop is left up to me. I’m not keen on spending my own money on work. I’ll be weighing budget and hardware specs against each other so I’m not approaching this with any specifics for either.
If I can manage a dumpster desktop for free with my own parts and whatever I can scrounge up then that’s what I’ll do. If it stays powered on and is able to out perform a thermal throttling 4700MQ it hits my minimum requirements.
That said I’m open to spending my own money if it gets me a great deal closer to workstation level performance. It really comes down to how much performance my shiny Loonies actually get me.
My hope is for some advice on where to look for used or new desktop hardware beyond the normal consumer/retail spaces. Maybe someone knows of often-missed locations, sellers, or events with hardware deals for a Canadian to check out? Maybe someone can point me towards programs, retailers or companies with a history of donating hardware to academia that I could reach out to?
For workstation use it really depends on what your specific workload asks for.
If it’s CPU Cores and more cores, you can either go with a Ryzen 2700, which you can probably snag used from someone upgrading, maybe even with board, or you buy a used Server/Workstation with Xeons in it. Older Server (In Tower Cases) can be had for sub 200 bucks and deliver 8-24 cores and plenty ram for the price. Your ST performance will tank and it’ll consume some power, but for Cores per dollar, you nearly can’t beat an older dual Xeon Workstation from Dell or such.
If GPU play any role, RX570s are cheapish, if you need Nvidia, well… You’ll have to spend some money.
Other options include the Ryzen APU’s with Vega graphics. Incredibly powerfull for what they cost.
If it’s clockspeed you’re after, i guess an 6th or 7th gen i7 can be found for reasonable money. Those still can be overclocked to close to 5Ghz, and nothing has been past this since.
TL:DR; look at your workload and ask yourself what the number one resource is to get your work done. After that, you can look at options. And used is almost always a great option. Used Professional Hardware often twice so.
There are the Chinese socket 2011 motherboard and an old Xeon CPU and ECC RAM. Not so good for gaming but general computing and workstation workloads not too bad. You will have what 5 plus years ago would have been a propper workstation PC.
There are of course the issues around getting stuff from AliExpress around the time of delivery and honesty of sellers.
Does the work you do in your PhD relate to any particular industry? Would it be worth reaching out? Something I would think that you would want to be doing anyhow for future job prospects. Your relationships you should be cultivating for your future may be open to giving you some support now.
I’ve been in a similar situation (recently defended a PhD in numerical acoustics). Ended up finding two e5-2690 and 64 Gb very cheap second hand and combined it with a Chinese motherboard. However, with the current declining ram prices and ryzen I’m not sure this makes sense unless you can find it very cheap second hand. In fact, I ended up doing quite a lot of code development directly on a cluster…
It’s all tensor calculus that ultimately boils down to geometrical/vector problems. My application is strictly to theoretical models so right now the direct usefulness to industry is limited. That said those basic tools are used all over, fluid flow, machine learning, and so on. So there could be options out there, I just wouldn’t know where to start.
It’s good to know I’m not alone haha. Also congrats!
That’s what I’m doing right now. We used to have a couple smaller servers for testing code but those were all retired in the past few months. I hope in the long term I can avoid having to book cluster resources for the development part, but we do what we have to do…
I was contemplating a basic Ryzen APU system or something along those lines but computer hardware in Canada tends to be on the expensive side. Going for something like a 2400G system still requires around $500-600. Not horrible but enough to make me look for other options!
You really need to be looking at a couple of generations older server hardware. An APU is an OK starting point for someone wanting to do some basic office and web stuff, even some light gaming. but proper maths?!? No you need cores and lots of them. Even if its a five year old second hand Xeon.
I have a build ready to go with full Ryzen. Since all you need is basicly a spreadsheet editor machine (since you could offload heavy calculations), I’d go with:
CPU: Ryzen 3 2200G (3200G is $20 more, 2400G if 8 threads are mandatory)
Motherboard: ASRock Fatal1ty B450 Gaming-ITX/ac Mini ITX (Cheapest mITX motherboard I could find, B450 is a better chipset if you can afford it though)
Memory: 2x8GB 3200 MHz DDR4 RAM
Storage: 250 GB NVMe
Case (PSU included): 250W small case. cheap
All this around 375 US dollars. Of course, if you got a spare case and PSU, that would be $150 saved. Hope that helps.
If you are doing geometrical vector/tensor problems wouldn’t it make sense to look into cuda and offload the workload to gpus? Or perhaps you are already doing that on the cluster…
If you are going the old xeon route I think it only makes sense if you can find the cpu very cheap second hand in Canada, even getting them on eBay seems quite expensive…
One thing I forgot to mention - the kit I specified can have an nVidia GPU added for CUDA. You could save $200 in total (roughly PSU $75, Case $75, storage $50) by reusing some parts, so if you have an old kit lying around then your best bet is to use that and only do the $200 - $250 motherboard + APU + memory combo. Your IT support section should have spare cases with power supply and harddrives available.
if you were to encounter errors in the calculations from your machine, would this be scientifically disastrous? or would you be checking results against a server/workstation grade box?
do you know if your workload can be gpu accelerated?
if #1 is true, then you’ll be wanting ECC memory if possible
if #2 is true, then GPU is relevant, otherwise don’t bother.
Ryzen is great bang for buck and can support ECC. The 8 core Ryzens (any generation) really are the modern day poor man’s workstation CPU and bang for buck outperform almost anything intel had in the workstation space only a few years ago.
I kinda agree, perhaps try to find a second hand first gen ryzen 1700 or so. If the machine is just for code testing and large simulation are offloaded to your HPC, I don’t really see the need for ecc ram, eventhough it is nice to have.
@BlueJedi Have a look for Computers for Schools(or Reuse Tech BC if your are in British Columbia) in your Province/Territory. You will need a teacher or counselor to apply on your behalf but you may luck out on a workstation or at the very least an i-series desktop that can handle some upgrades.
Pricing depends on where you live but you can usually get a good setup for fairly cheap. I work at CFSY(Yukon Territory) and our pricing is $10 for a Desktop with a second gen or newer i-series chip and if people request a Xeon workstation we charge the same price. I think some of the Southern shops are a little more expensive but I can’t see it being that much.
This is the issue. If I want to leverage GPUs and keep my sanity, then at the development stage it’s nice to have even a basic CUDA GPU locally. Just so the code can run for testing. Otherwise I have to develop on the cluster which is not ideal. Our clusters are scheduled or use task queues.
For 1, errors won’t matter for local computations. Any data sensitive to hardware errors that’s used for publication will be done on the cluster. It has ECC and related features.
For 2, there is a good number of tasks in my work I could leverage a GPU for. Depends on the VRAM available and the problem. Right now I’d consider it a “nice to have” but not essential. The cluster I currently use doesn’t offer GPU compute options but it’s something I could get access too.
I found the one for my province. I’ll send them an email and I can get my supervisor to make the request if they have options. Thank you for this!
It’s definitely nice to have the ECC and GPU locally for development but it’s hard to justify paying a lot for them with the cluster there for the heavy lifting. That’s why requirements and budget were hard to set for this.
If I can’t find a dirt cheap used 8 core Xeon workstation then it seems like the consensus is Ryzen is the way to go. It’s probably worth paying a bit extra for Ryzen to get the performance uplift of the newer hardware too. So I’ll keep that in mind when looking at the used Xeon gear.
I do like that ECC would be an option down the road for a Ryzen system. Might need to be careful about my motherboard choice? I vaguely remember Wendell looking into that at one point.
If you need to upgrade the GPU, you might need to be aware what CPU you are pairing it up to in order to avoid bottlenecks. Even if you go with the cheapest CPU you can find, you pretty much will need a 3600 or even 3700X to get the full benefit of a 2060 or 2070 card. On the other hand a 3700X will be massive overkill for a 1030 card, the card will simply not be fast enough leading to an excess of clock cycles.
For your build, I’d say that a 1050 Ti + Ryzen 1200 would be sufficient then. It is not a huge deal if your CPU is underpowered or overpowered - just be aware you will be letting some clock cycles slip away if the two are not quite compatible.
To the best of my knowledge a CPU bottleneck shouldn’t cause massive losses for my compute work loads.
My work loads are more like rendering one frame of a very complex scene all loaded at once. As opposed to rendering a scene many times very quickly, each requiring small updates from the CPU.
For my use case I really only need the hardware there for development anyway. Testing, initialization, running sample problems and so on. I won’t actually need to run the solvers all the way through a full problem locally.
I’ll see if I can price the same thing from Canadian retailers. Thank you for taking the time to spec out builds for me, it’s appreciated. This gives me another option to look at!
Hopefully I can scavenge a few things locally too.
Ryzen 1700x (used ebay etc) £100
Cheapest b350 or b450 motherboard with m.2 slot (new) £50
128gb nvme m.2 £30
16gb ddr4 (2x8) as cheap as possible fast stuff (new) £65
240 AIO cooler (new) £45
2x pwm fans £10
500w decent bronze psu (new) £35
Nice case (new or used, taste dependant) £50
GTX 1070 (used ebay etc) £200
Total cost without additional storage £585
RGB, Sleeved Cables etc are all luxuries but the nvme gets you onto windows and gets your programs on. HDD or SSD storage add to your own needs or recycle your old drives.
This will make you a very decent machine to work on plus be able to do other things due to having 8c16t of cpu and plenty of cuda and vram on the GPU
Great bang for the money and will serve you very well
You’ll need an additional drive for storage, which means you’re realistically looking at 2x the cost of that single 128GB m.2; there’s a floor to spinning rust prices…
Which (when you add the rust plus 128 GB SSD) puts you in the ballpark of 256-500 GB M.2 drives which may be “enough” to go 100% SSD (given this is a workstation and not a steam library machine). Certainly if you raise budget slightly to go to a TB of m.2 or more most likely.
I don’t think there is any point buying 128 GB SSDs in 2019 unless you’re using it as a cache for a large hard drive - in the case where you need a huge amount of storage.
Sure you might get fast boot time but in a workstation that is spending most of its time not rebooting windows, but rather doing work - you’re probably even willing to boot from slow media (once a month) and have your actual work running from SSD.