harshitaneja

Thanks a lot for your help. I think I might be able to add in one Titan RTX for those one of models that require the extra memory. Thanks a lot.

I don't intend to use 8 GPUs. Was asking about using the case with 4GPUs. Would find such a board. Thanks. Buying threadripper in country. 2950X is 1000$ inclusive of all charges so would be very cheap. I wanted to wait for gen3 for pcie4 and 7nm and higher core count but with the unsure launch time frame and my current requirement it sadly wouldn't make sense to wait.

I live in india and newegg delivers but the custom and delivery charges really add up and you can view them until you reach check out. Local players don't deliver server CPUs (at least can't find any online). I checked on newegg and EPYC CPUs max out at 3ghz+ clock speeds. I looked at intel CPUs as well as they provide higher clock count but threadrippers seem the best fit between the number of cores and clock speeds. I am not sure if such clock speeds are sustainable though. Would the 8GPU miner case work with threadripper?

My work requires extensive use of CPU as well. A lot of statistical models target CPUs and don't have GPU support. So higher clock speeds offered by threadripper would really help. A similar EPYC CPU i think would be cost prohibitive I imagine. Would threadrippers with 64 lanes not work here? Thanks for your help.

It is not a pure GPU rendering build. My work involves a lot of statistical models which are targeted for CPUs rather than GPUs and I would really benefit from the 4Ghz+ clock speed of threadrippers. EPYC with similar configuration would be cost prohibitively costly I imagine. At that point it might make more sense to make multiple PCs than a single one. But why would threadripper not work here? They have 64 lanes which should be enough for my build. Sorry about the confusion about portability. By portability I meant that the structure should be transportable as I have heard of issues regarding transporting Liquid cooled builds. Size is not a constraint. Thanks for your help.

I should clarify what I mean with portability. By portability I meant that the structure should be stable enough to transport. I have heard issues with transportation of liquid cooled builds. I have no issue with the size. It can get as large as required. I mentioned my GPU and CPU requirements. 4x 2080Ti and 2950Xwhich is a 16core CPU. I am looking at 15000$ maximum.

??

I am building a PC with following parts- Threadripper 2950X 64GB RAM(16GB x4) RTX 2080Ti X399 Motherboard I wish to upgrade this build to add in 3 more 2080Ti later and increase the RAM to 128GB. This is for my datascience and development work. The machine would be running 24x7. What are the recommendations for this build? Would using 4 RAM modules instead of 8 cause any issue? Which RTX 2080Ti to use? I am thinking of using Asus ROG Strix RTX 2080 Ti OC 11GB. Asus released Matrix recently but I am not sure if price difference is worth it. Also would it be possible to use such huge sized cards for an eventual 4 GPU build. Would GPU risers help to mitigate the issue? What about cooling? I am planning to run this under heavy load for most of the day every day. I would want something a bit portable as I have multiple offices(though if not possible I can always SSH). Should I go for liquid or air cooling? How many of these? Something that would hopefully do all this while being quiet. Which RAM modules to use? Would paying extra for the extra clock speed be worth it? I am thinking of using Corsair Vengeance Lpx 16GB (16GBx1) DDR4 3200MHz. But my knowledge of latency and other things regarding RAM and how they work with Threadripper caches is quite limited to make a decision. I am looking for a motherboard that can handle the eventual upgrade. X570 now support thunderbolt 3. Is it possible to get something similar here? I have a large number of thunderbolt 3 devices. Which case would help to contain the eventual build and still allow noise dampening? Pricing is not a concern. I wish to have the best parts for each that make sense. I wished to make my complete build at once but the complexities of motherboard real estate and cooling made it quite overwhelming especially as I am building a PC after a long time. Thanks for your help. EDIT: By portability I mean that the structure should be transportable as I have heard issues occur with transportation of liquid cooled builds(not sure). Size is not a constraint. Sorry for the confusion.

I had considered that. Titan RTX would be beneficial if dataset sizes were a concern. My models can run well in 11GB RAM and wouldn't benefit much from the added RAM. But having access to more GPUs would allow me to run more number of models simultaneously.

I wouldn't mind getting a server chassis. Can you recommend any place to look for them? I need to run this 24x7 most probably. What would you recommend for cooling?

I need to run multiple different models for various clients at the same time. And even the same model needs to be run with different parameter configurations. With this I need a dev bench for getting started with distributed computing so I plan to spin up multiple VMs.

I need to build a workstation for my data science and statistical work. I plan to use Threadripper 2950X as the CPU with 4x 2080Ti cards. I currently only have the budget for 3 cards so starting there. I intend to use Asus Strix 2080Ti OC for this. 1. Would you recommend this card for such a build or should I go for another one? 2.This card is massive in size- 3 slots(2.7 slots technically). Would any motherboard be available for such a build? Do I need to use GPU risers? 3. What about heating? Would liquid cooling be better or air cooling for this kind of build? 4. I hope I can build a device that is not noisy. I don't know if it is a reasonable expectation. What could I do to ensure less noise?

CUDA is sadly required as tensorflow and certain others only support it which is bad as I really don't want Nvidia being a linux user and their abysmal driver support. I have edited the question to include CUDA support.

Maybe I should have mentioned it in the answer, 2950X is for the extra cores as some of my work involves running statistical libraries which are currently CPU only. Extra PCIe lanes are cherry on top.

Thanks for the response. Titan V is an older card which has similar if not worse performance in certain cases in comparison to 2080Ti in FP16 workloads. Even in FP32 difference isn't large enough to warrant paying around 2.5x for it. I think you are confusing it with Titan RTX. Even though Titan cards have 24GB of memory in comparison to 11GB of a single 2080Ti. But total memory of all three planned GPUs 11+8+8 (GB) should be sufficient to counter that(now I know I am comparing a single memory with sum of multiple memories which is not a straightforward comparison) without much of an increase in price. With this given that my work mostly is parallelizable I can run it across all the three at the same time which should give higher performance and if not, it offers the benefit of running different models on all the three GPUs. I need the 16 cores offered by 2950X(well I wanted the 32 core 2990WX but couldn't justify the cost) as I need to run a lot of virtual machines and some of statistical work is done using libraries which don't yet have GPU options and thus need the cores for that. So PCIe lanes are not an issue. Thanks again for your input.

Building a workstation with threadripper 2950X(need the cores) and 3 GPU- 1x 2080Ti and 2x 2080. Mostly for data science work. NVLink is not needed as work will be parallelized across multiple GPUs. Need help with GPU recommendations(CUDA support required) and motherboard. I need to live in the same room as this device which will run 24x7 so would need to really quiet. Water or air cooling? I live in India where it is tough to find people to help with water cooling(at least that I know of) so would be doing it myself. Is IChill Black 2080Ti a good option? Any suggestions or recommendations or pointers to places I could ask or read at. Thanks a lot for the help.

Alright. Q3 is too late, so would get it now. Already delayed a lot. Thanks a lot again.

Huh. I thought with all the ryzen 3 rumours it would be around the corner

Ya. Thanks. What do you think, should I wait for threadripper 3 or just go ahead?

I use tensorflow a lot and most of the other libraries I leverage use CUDA. Not having access to CUDA would be a real disability for my work. RocM support for tensorflow is shaky at best. Otherwise I would have preferred Raedon especially because I run linux and Nvidia driver support is not realy great for consumer GPU on linux, something where AMD is quite great these days.

Android provides a good UI toolkit and creating good UI in linux is not fun and electron is really heavy. So thinking of getting an android tablet and flashing AOSP and build the system on it. Any android tablet recommendation for which I can get Android 7+ AOSP with some community support. Nexus line back in the day would have been perfect for this. Don't know of any tablet these days. Get a raspberry Pi like SBC(something more powerful) and hooking a touchscreen to it and a USB modem perhaps. Does anyone know of good 720p+ touch displays for this usecase? Or should I give up and just buy an android auto and drive on with my life? Thanks a lot for the help.

Alright. 32GB modules are hard to get though, will try to find. Datasets lie in 50-80GB region generally but not so rarely I find myself working with sets of sizes 200-500GB+. I have worked with these even on my old laptop with 8GB of RAM but it gets really tedious so having more than necessary RAM is the luxury I want. I can max afford(I can go further i guess but only if it really makes sense) 3 2080 but would that be better than the other configurations namely 2 2080Ti or 1 2080Ti with 2 2070 or 1 2080Ti with 3 2060. Thanks a lot for your help

Yes, 8x16GB is what I am looking at. Memory bandwidth issue is a concern but as almost all of the code I would run is my own so I should be able to avoid the issue to some extent. Not being able to keep it outdoors is saddening. I really need to worry about keeping the noise under control then. Thanks a lot

I have read this article many times. Although it discusses multigpu builds I was unsure of the specifics and thus my question.

Datasets tend to be large enough that they saturate 64GB RAM ocassionaly. Currently I have to engineer ways to load data asynchronously, which is fine for production models but while experimenting they take substantial time and effort that I would rather wish to spend on analysing data. 64GB should work but my major worry regarding upgrade was the impact of number of channels used on efficiency of RAM in threadrippers. I assumed I need to use all 8 slots for maximum efficiency. Budget is around 7-8k $. But I am in India so parts tend to be a bit more costlier here.