PC build for gaming, programming, machine learning

[Blasqen]

Approximate Purchase Date: Within 2 weeks
Budget Range: Around 1800
System Usage from Most to Least Important:
- Programming
- Machine Learning (eg training neural nets)
- Gaming/streaming
Parts Not Required: -
Preferred Website(s) for Parts: -
Country: Netherlands
Parts Preferences:
Overclocking: No
SLI or Crossfire: No
Monitor Resolution: Currently 1080p, will upgrade in the future

Additional Comments:
- I find it difficult to choose between the intel i7 8700 and the Ryzen 7 2700x. I am not sure yet which one to go for so I am proposing 2 builds. I think the Ryzen 7 2700x might be a better pick, because I often run more than one programming environment at the same time as well as virtual machines, so the good multi core ability might be handy here. If anyone has any advice on this and/or suggestions on either of the proposed builds I would be most grateful.

- For the Ryzen build in particular I would like to know whether or not the current motherboard will be good enough as I am reading that the VRM isn't very good. I don't want to overclock, but will I be able to get maximum performance with this motherboard (without overclocking). Will precision boost work properly?

- I already own the Samsung SM951 SSD

- I am going to use the gtx1080ti not only for gaming, but also to train machine learning models such as neural nets.


AMD build:

PCPartPicker part list / Price breakdown by merchant

CPU: AMD - Ryzen 7 2700X 3.7GHz 8-Core Processor ($318.99 @ SuperBiiz)
CPU Cooler: be quiet! - PURE ROCK 51.7 CFM Sleeve Bearing CPU Cooler ($34.69 @ OutletPC)
Motherboard: Asus - ROG STRIX B450-F GAMING ATX AM4 Motherboard ($123.91 @ OutletPC)
Memory: Corsair - Vengeance LPX 16GB (2 x 8GB) DDR4-3200 Memory ($159.99 @ Amazon)
Storage: Samsung - 960 EVO 250GB M.2-2280 Solid State Drive ($94.99 @ B&H)
Storage: Samsung - SM951 256GB M.2-2280 Solid State Drive ($228.90 @ Amazon)
Video Card: Gigabyte - GeForce GTX 1080 Ti 11GB AORUS Video Card ($654.99 @ Newegg)
Case: Fractal Design - Define S ATX Mid Tower Case ($89.98 @ Newegg)
Power Supply: Corsair - RMx 650W 80+ Gold Certified Fully-Modular ATX Power Supply ($79.00 @ Newegg)
Total: $1785.44
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2018-09-17 15:18 EDT-0400

Intel build:

PCPartPicker part list / Price breakdown by merchant

CPU: Intel - Core i7-8700 3.2GHz 6-Core Processor ($303.75 @ Amazon)
CPU Cooler: be quiet! - PURE ROCK 51.7 CFM Sleeve Bearing CPU Cooler ($34.69 @ OutletPC)
Motherboard: MSI - B360 GAMING PRO CARBON ATX LGA1151 Motherboard
Memory: G.Skill - Aegis 16GB (2 x 8GB) DDR4-3000 Memory ($135.98 @ Newegg)
Storage: Samsung - 960 EVO 250GB M.2-2280 Solid State Drive ($94.99 @ B&H)
Storage: Samsung - SM951 256GB M.2-2280 Solid State Drive ($228.90 @ Amazon)
Video Card: Gigabyte - GeForce GTX 1080 Ti 11GB AORUS Video Card ($654.99 @ Newegg)
Case: Fractal Design - Define S ATX Mid Tower Case ($89.98 @ Newegg)
Power Supply: Corsair - RMx 650W 80+ Gold Certified Fully-Modular ATX Power Supply ($79.00 @ Newegg)
Total: $1622.28
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2018-09-17 15:57 EDT-0400

Thanks in advance!

 

[vapour]

1800 in usd?
To me, depending on what software to train your neural network. If it is something like tensor flow, 2 pieces of 1080ti is more important. For streaming, 2700X is really good. I will vote for a 2700X build as X470 has 8 more PCIE lanes, crucial for multiple gpus. Threadripper has even more pcie lanes but much more expensive.

Can you explain more on the machine learning work you need to do?

 

[Blasqen]

Oops, I forgot to turn this into a discussion topic, instead of a question, can this still be changed?

1800 in usd?
To me, depending on what software to train your neural network. If it is something like tensor flow, 2 pieces of 1080ti is more important. For streaming, 2700X is really good. I will vote for a 2700X build as X470 has 8 more PCIE lanes, crucial for multiple gpus. Threadripper has even more pcie lanes but much more expensive.

Can you explain more on the machine learning work you need to do?

Mostly for university/hobby purposes. I am mainly using tensorflow. I don't do kaggle competitions or anything like that and usually don't train very large models. My laptop lacks a decent GPU so I currently do everything on an intel I7 CPU, which is way too slow, so a gtx1080ti is going to be a huge upgrade. With a second 1080ti I would be way over my budget, if I ever need more power I will probably opt for the cloud instead.

 

[vapour]

I see, expensive hobby!
I will just change to X470 MB
drop the 3rd party cooler, stock is more than enough
get a better psu, 850 is better for potential gpu expansion.

 

[Blasqen]

I see, expensive hobby!
I will just change to X470 MB
drop the 3rd party cooler, stock is more than enough
get a better psu, 850 is better for potential gpu expansion.

BTW, budget range is ~1800 dollars, Obviously I will buy the parts in euros at different webshops, but I wanted to keep it in line with the Part picker thingy.

Thanks for the sugestion! The reason I chose for this particular MB is the fact that it has 2 M.2 PCIe 3.0 x4 slots for both of my SSDs. Many of the X470 boards in the same price range have one M.2 PCIe3.0 x4 slot and then another slot that is e.g. M.2 PCIe 3.0 x2 or M.2 PCIe 2.0 x4, which will limit the speed of one of the SSD's slightly. Although, I will probably not notice the difference in real world usage. Just to be sure, my chosen SSD's both are compatible with either of these slots, right? They will just cap the speed?

Say, if I would be absolutely sure I don't want GPU expansion, would you still advice to take an x470 MB over a B450 MB considering the Ryzen 7 2700x CPU, or should I be fine as long as I don't OC?

I could get the RMx850 PSU, it is only 5 euros more expensive, would this be a good pick? Will this be better/worse/equal when it comes to power consumption?

Thanks!

 

[vapour]

Yes, 850 RMx is very good psu
you can still oc on B450, but will be lower than you can achieve on X470
I will get X470 to keep potential for SLI.
For M2, I don't think it is quite an issue, it will cap speed but if you want to have high speed on both, good X470 is preferred.

 

[Lucky_SLS]

For machine learning and neural networks, Nvidia released the rtx cards. There are no benchmarks as of yet, but it's advertised for those workflows.

You might want to wait and decide on getting a rtx 2070/2080.

If your cpu utilization is over 90% during those workloads and on all cores, a better cooler is recommended. Something like noctua NH D15 or cryorig R1 or be quiet dark rock pro 4.

You can wait and decide on the cooler. Ryzen master application let's you monitor cpu core Temps and usage. But make sure to get a case with enough clearance to house those big cpu coolers.

 

[Blasqen]

For machine learning and neural networks, Nvidia released the rtx cards. There are no benchmarks as of yet, but it's advertised for those workflows.

You might want to wait and decide on getting a rtx 2070/2080.

If your cpu utilization is over 90% during those workloads and on all cores, a better cooler is recommended. Something like noctua NH D15 or cryorig R1 or be quiet dark rock pro 4.

You can wait and decide on the cooler. Ryzen master application let's you monitor cpu core Temps and usage. But make sure to get a case with enough clearance to house those big cpu coolers.

Well, I already bought the gtx1080ti, since one particular vendor had it on sale and it was 650 euros, which is about a 120 euros cheaper than normal here, quite a big discount. So I took my chance and bought it. I haven't unpacked it yet, so I could still return it and get an RTX2080. The problem is that the cheapest RTX2080 card sells for 850 euros; 200 euros more than what I currently paid for the gtx1080ti. When it comes to the RTX2070 I highly doubt that it will be better than the GTX1080ti.

Early benchmarks have shown that the RTX2080 is only marginally better than the GTX1080ti, while it would cost me an additional 200 euros (+30% more expensive). Though I haven't seen anything about neural network training time benchmarks so it would be hard to tell what would be better. I am inclined to stick to this gtx1080ti, considering the price I got it for.

keep in mind that I am also somewhat tight on my budget. I can't spend much more than 1800$. So if I take the RTX2080 I would have to cut on something else. Let me know whether or not sticking with this GTX1080ti would be the best option in this case.

About the cooler: So I should not by a CPU cooler at first? Check how hot it gets, and then decide on a cooler later? Will the current case (Fractal Design Define S) be big enough to house those big coolers?

What do you say about the motherboard? I probably won't overclock as I hear many people say that it is rather pointless with a Ryzen 7 2700X as it has this thing called precision boost. Would I get better stock performance with a X470 motherboard than with a B450 motherboard, or won't it matter?

Thanks a lot for your input! I appreciate it a lot

Yes, 850 RMx is very good psu
you can still oc on B450, but will be lower than you can achieve on X470
I will get X470 to keep potential for SLI.
For M2, I don't think it is quite an issue, it will cap speed but if you want to have high speed on both, good X470 is preferred.

Alright, I will swap the 650RMx for the 850RMx, considering the marginal price difference. But still, would this consume more/equal/less power overall? I am reading that when it comes to PSU, they are maximally efficient when they are under 50-75 % load.

What do you mean "high speed on both"? both SSD's or the SSD's and the processor? Many of the X470 boards in the same price range do have 2 M.2 PCIe ports, but usually one of them will be PCIe 2.0 x4 or PCIe 3.0 x2. I know I probably won't notice a speed difference of 200 MB/s, but I probably won't do dual GPU's, and I don't really want to overclock as many say that it is a rather pointless thing to do on the Ryzen 7 2700X.

The problem I have with dual GPU's is this: I probably won't be able to buy a GTX1080ti for at least 1 year, probably more. By that time, there is a good chance that the GTX1080ti won't be available anymore, and I would have to look for a used one, which has probably worn off a lot. Are my concerns valid?

Come to think of it: if I were to buy a second GPU down the road, does it have to be the same GPU? If not, than my argument is pretty much invalid, and I might opt to add an RTX GPU next to the GTX1080ti. As long as setting it up, won't be too much of a hassle.

Thanks for all the help btw, really appreciate it!

 

[Lucky_SLS]

Those early performances are for games. The rtx cards have rt cores including Cuda cores. Those rt cores are what designed to do the machine learning and neural network and stuff.

BTW, I dunno which software you use or how that software is programmed to use the gpu resources. Rt cores are all new and requires separate implementation to use it. Although the dx12 and latest Nvidia drivers will have the necessary drivers, I don't know if such tasks will be automatically offloaded to the rt cores in them.

That's why I said wait for the reviews. This is still bleeding edge technology, but can have huge performance gains if used properly.

Define S can house those beefy coolers, no issues there.

 

[vapour]

For machine learning and neural networks, Nvidia released the rtx cards. There are no benchmarks as of yet, but it's advertised for those workflows.

You might want to wait and decide on getting a rtx 2070/2080.

If your cpu utilization is over 90% during those workloads and on all cores, a better cooler is recommended. Something like noctua NH D15 or cryorig R1 or be quiet dark rock pro 4.

You can wait and decide on the cooler. Ryzen master application let's you monitor cpu core Temps and usage. But make sure to get a case with enough clearance to house those big cpu coolers.

He says he is using tensorflow, speed is sub-linear to # of 1080ti's, but you see the scaling.

 

[Lucky_SLS]

Solving your tensor equations is basically matrix multiplication operation. Correct me if I am wrong op.

And about rtx and its tensor cores : https://www.tomshardware.com/reviews/nvidia-turing-gpu-architecture-explored,5801-5.html

 

[vapour]

Solving your tensor equations is basically matrix multiplication operation. Correct me if I am wrong op.

And about rtx and its tensor cores : https://www.tomshardware.com/reviews/nvidia-turing-gpu-architecture-explored,5801-5.html

https://www.tensorflow.org/

 

[Blasqen]

Those early performances are for games. The rtx cards have rt cores including Cuda cores. Those rt cores are what designed to do the machine learning and neural network and stuff.

BTW, I dunno which software you use or how that software is programmed to use the gpu resources. Rt cores are all new and requires separate implementation to use it. Although the dx12 and latest Nvidia drivers will have the necessary drivers, I don't know if such tasks will be automatically offloaded to the rt cores in them.

That's why I said wait for the reviews. This is still bleeding edge technology, but can have huge performance gains if used properly.

Define S can house those beefy coolers, no issues there.

Yeah I know. The only thing I have read is that the RT cores are for used for model inference when doing things like DLSS to speed up game rendering. The RT cores can, and will be used for to accelerate deep learning, but how much? Model inference is "cheap" in comparison to training, so I don't expect any spectacular results to justify spending an additional 200 euros. I could still return the GTX1080ti, but it is really risky, because I don't think I will be able to get a GTX1080ti at this price point anytime soon . All I know right now is that the GTX1080ti is still a really good card for these purposes.

And as you said, it is still bleeding edge tech. There is no guarantee that Tensorflow will make optimal use out of it anytime soon.

btw, I am going to be buying the an x470 board instead of the b450 I selected. It just seems a little bit safer when looking at the VRM and the read speed. And I probably won't notice the 150 MB/s speed cap on the SM951 SSD. a max. read speed of 2000 MB/s is still very, very fast anyway.

If your cpu utilization is over 90% during those workloads and on all cores, a better cooler is recommended. Something like noctua NH D15 or cryorig R1 or be quiet dark rock pro 4.

You can wait and decide on the cooler. Ryzen master application let's you monitor cpu core Temps and usage. But make sure to get a case with enough clearance to house those big cpu coolers.

Does this apply to the current motherboard? Or am I also going to need better cooling with an x470 motherboard (no overclock)?

 

[Lucky_SLS]

Go for the asrock X470 taichi. Good vrm and features.

 

[Blasqen]

That MB is way above my budget. I am thinking about the Gigabyte Aorus X470 Ultra Gaming, would that be a decent choice?

 

[Lucky_SLS]

Msi gaming pro carbon? Asrock fatality k6? Gigabyte aorus gaming 5?

 

[Blasqen]

Msi gaming pro carbon? Asrock fatality k6? Gigabyte aorus gaming 5?

Aorus X470 Ultra Gaming isn't good enough? it is 140 euros here.

All the MB's you mentioned seem to be boards with intel chipsets?

 

[Lucky_SLS]

Asrock fatality and gaming pro carbon have their X470 variants. Maybe wrong on the aorus.

The ultra gaming is their starting lineup X470. Not so good.

If you can get atleast the X470 gaming pro from msi, go for it. X470 prime a from Asus is also good.

 

[Blasqen]

Are you sure? Because the MSI X470 gaming pro is actually 13 euros cheaper than the Gigabyte Aorus X470 Ultra Gaming. Of course higher price doesn't necessarily mean better performance.

 

[Lucky_SLS]

Yup, msi gaming pro is good. Only the gaming plus and lower are meh.

If you can get the gaming pro carbon, even better.

 

[Blasqen]

Yup, msi gaming pro is good. Only the gaming plus and lower are meh.

If you can get the gaming pro carbon, even better.

Thanks! I am also now looking at the MSI B450 Gaming Pro Carbon. VRM wise it seems to just fine. It packs a better audio chip too.