2

I used laptops for a while and I am not up to date with the current available hardware.

I'm looking for a new PC mainly for Neural Network computations and some gaming from time to time.

The rig mainly needs strong Graphic Cards for OpenCL/CUDA computations. The CPU must not bottleneck the GPUs. The cpu needs proper single thread performance.

My target is to have a very upgradable workstation and install a relativly cheap CPU and Graphic Card. I want to upgrade the CPU when needed, so the socket should be somewhat futureproof or at least not already outdated. Maybe the i5-6600K on socket 1151 is what I need?

I want a single Graphic Card for now and add more once my Neural Network research progresses. The Motherboard should have at least 4 PCIe slots, at least two of them PCIe x16. That should be enougth room for Graphic Card expansion.

I need to be able to upgrade to at least 64 GB of fast DDR4 RAM, but will use 16 GB for now. Low lateny ram is favored.

Currently the Software I use for my Neural Network research has no support for AMD cards and requireds nvidias CUDA. So I have to buy nvidia cards. SLI is not needed, since the Cards can run those computations as individual units and Gaming is no priority.

I'm not sure if a consumer mainboard fits my needs well enough. Maybe a Serverboard?

I'm targeting a bugdet of below 1500€ (~$1660) for its first configuration.

1 Answer 1

1

Your question fundamentally breaks down to which motherboard you should go with, because the motherboard will determine the CPU, GPU(s), RAM, and everything else. I'm going to recommend, therefore, that you go with a dual socket 2011 board, like this ASRock EP2C612D16C-4L SSI EEB Server Motherboard, though any other board like it would probably do as well for you.

These boards support a LOT of PCI-E throughput - this particular model supports 3x 16x slots and 2x 8x slots, in a configuration that would allow you to run up to three double-height cards without risers, with the caveat that the top card would need to be an ITX version of whatever card you wanted - these cards are available without a huge price bump for the most of the recent Nvidia GTX lineups. Furthermore, this board supports the socket 2011 HPC Xeon CPUs, which offer extremely performative CPUs with lots of threads for massive parallel work. An example configuration would be something like this, where CPU slot 2 is left open and can be filled later (the CPUs are cheapest on the used market - vastly cheaper, actually). Finally, unlike some other socket 2011 boards, this one does run on DDR4 spec RAM, so that's in line with your request. As it is, it can support up to a theoretical maximum of 1Tb RAM, but more realistically about 512Gb.

Right now, this CPU sits near the absolute top of the performance charts, unless you start getting into E7 Xeons and POWER8/9, I think, and it certainly enjoys a fantastic price/performance ratio thanks to the flood of parts on the used market.

5
  • Great recommendations! I'm a bit concerned about the single thread performance. Even in parallel workloads there is often a coordinating thread that can bottleneck the rest of the system. Xeons are not really overclockable since the multiplicators are locked. The (relativly) low clock speed might become a problem to me.
    – maklemenz
    Aug 8, 2016 at 20:45
  • All true! If you could give us some kind of data about what kind of single thread performance is needed (e.g. a benchmark score, a known-good CPU model, etc), we could say with a great degree of confidence whether the CPU given in the example could do what you want.
    – Adam Wykes
    Aug 8, 2016 at 20:47
  • Good CPU recommendations regarding neural network computations are hard to find. Most people don't analyse or understand the bottlenecks and their root cause. Apart from that, some people use the i7-5820K to feed four Titan X cards. Sidenote: From what I just read low latency ram plays a relatively big role.
    – maklemenz
    Aug 8, 2016 at 21:38
  • 1
    Well shoot, if low-latency RAM is the way to go then I'm not even sure DDR4 is what you're after. DDR3 (and for that matter DDR2) has much lower latencies inherent to its design. That being said, at some point you're going to run up against a RAM bandwidth bottleneck, so no doubt there's a sweet spot to hit in that regard. The i7-5820K is an interesting choice for feeding four GPUs, since it doesn't physically have the PCI-E lanes needed to fully exploit the bandwidth - so perhaps the PCI-E bandwidth isn't a concern for this job, much like how it doesn't really matter in gaming?
    – Adam Wykes
    Aug 8, 2016 at 21:55
  • 1
    If you can't figure out what kind of CPU performance you need, you probably need to start out by performing experiments to see what CPU loads are like on a reasonably high-threaded system. If someone you know has something like an FX-8xxx series CPU or better, run your stuff on that PC and find out what the program behaves like. That's the only way you're gonna know what you need to do.
    – Adam Wykes
    Aug 8, 2016 at 21:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.