GTX 2080/2080Ti RTX for Deep Learning?


#123

Hi @init_27,

Exciting build!

Definitely go for the NVMe M.2 as the main disk and use a secondary hard disk for storing datasets. If you don’t currently own more than 400+GB worth of data, consider purchasing the 512GB Samsung 970 M.2 first and delay the purchase of your secondary hard disk until it starts to fill up so you can take advantage of better technology down the road and possibly falling SSD prices to get a faster or bigger SSD later, considering how easy it is to add a HDD to an existing setup.

Just go with the AIO cooler for the CPU instead of looking for an air cooler if the price difference is not great, as you’ve said. It gives you the option of overclocking the CPU or keeping your case temp lower (use the AIO fan as exhaust) so your GPU doesn’t throttle as easily.

As for Intel vs AMD CPU, I used to consider it an advantage to be able to take advantage of Intel MKL which comes with Anaconda’s Python, which is why I went with an Intel CPU, but I’m not sure whether that speed advantage still holds with the latest AMD processors.


(Sanyam Bhutani) #124

Thanks for the advice @howkhang

I think I’ll implement your advice-I don’t have any dataset of huge size, I’m just trying to give some extra leg-room.

I’ll use a 128GB SWAP+ 50-100GB OS+Misc on the 512GB NVMe M.2 which leaves me with 200-300GB of free storage.

I checked and the motherboard indeed supports 2xM.2 so I’ll get one down the road, just do an OS re-install-that shouldn’t as hurtful as the price for a 1TB one.

Gotcha, these were the reasons I had in mind as well. I’ll keep the Liquid cooler then.

Although I’m still not sure about @lesscomfortable’s comment that it’s slightly unsafe.
Francisco, could you kindly elaborate?

Makes sense. I think I’ll stick to Intel in that case, just wait for CES incase AMD decides to release an amazing processor (The chances of which making to the immediate market are very thin)

Thanks & Regards.


(adrian) #125

Just my two cents: I had a kraken liquid CPU cooler but found the fans pretty noisy (even when throttled), swapped it for a noctua cooler - hardly any temperature difference, less noisy and much easier to manage when you need to swap parts in/out -having that extra rad in the box can make things a chore (rad was on top)


(Andrea de Luca) #126

Mates, we are slightly OT in talking of general hardware. This should be about RTX cards. However:

There are many z370 boards with 8+8. Eg: asrock z370 sli.

No speed losses at all for deep learning: see Dettmer’s blog. No difference in training times between the fastest and the slowest mems on the market.

I agree, but realize you’ll lose Intel MKL support. It’s often useful.

Search for Samsung 970.

Indeed. And less (or no) maintenance.

Use a Cooler Master TX-something. Cheap and efficient.

OS, swap and current datasets will fit on the M.2. Use the SATA as archive.


(Francisco Ingham) #127

There generally is no problem but liquid coolers might leak and that might damage other computer parts. You can google for this but it ultimately depends on how conservative you are.


(Sanyam Bhutani) #128

Sorry for going OT-it’s an RTX build which is why I posted here instead of the build threads.

I’m going for the MSI z370 PC PRO

Good point, Thanks.
I’ll check out the regular Ripjaws ones if the price difference is minor (I think it’ll be huge)-I’ll stick to RGB ones (for the love of RGB)

Okay, sticking to the 8th gen i7 (9th gen has quite a bit of a price jump)

I found Hyper 212 evo Its 20$ cheaper than the liquid one. The TX one is out of stock apparently.

Thanks for the suggestions.


(Sanyam Bhutani) #129

Sorry for not searching first- I thought of seeking your opinion.

My original thought was: If I add 1 more GPU in the future, this might help.
I learnt that I might need the liquid cooler only if I want to overclock the CPU-I won’t be doing that.
Plus I really don’t want it leaking all over the the other expensive components. I’ll switch to the Air one. If I ever need to install the liquid one-I’ll just upgrade in the future.

Thanks.


(Francisco Ingham) #130

Yeah, same for me. I think that is a wise choice.


(Andrea de Luca) #131

I strongly advise for a Xeon equivalent with ECC mem (possibly registered).

Without ECC, data corruption on disk is unavoidable in the long run.

Go for a used Xeon e5-v3 & a X99 board on ebay. You can buy them for less than new z370/8700k-ish rigs.

Check the X99 board support for ecc (for example, all the asrock and the majority of asus do support it).

As a bonus, you will get 40 lanes, support for 3 cards, and support for a ton of memory (often >128Gb) as well as for cores up to 14.
Much more future-proofing despite the older hardware. Consider that with z370/90 you are stuck to 64Gb, 2 cards and a maximum of 8 cores (z390) or 6 cores (z370).

Personally I went for a Xeon e5-2680v2 with 10C/20T, c602 chipset and 96Gb of ECC rdimms. Crap cheap, great rig for DL. NVMe drive on a 4x slot with adapter.
Advised for e5-v3 since they got cheaper recently.


(Sanyam Bhutani) #132

Thanks @balnazzar
I think the issue for me would be it’s hard to get a hold of these in India. Shipping from ebay to India would be too costly.

Tbh, The build for me is with future proofing in mind-Right now I don’t need that huge amount of compute but I’m sure it might help w time, I’m hoping this would be good enough to last me for 2-3 years. I can add another 2080Ti or just replace the existing one with the 24GB Titan if I ever need that much compute.

For RAM-If I need more, I intend to use an Optane as SWAP (Thanks for the suggestion @Ekami)

TL;DR:
Great suggestion but its hard to get the second hand comp in India or have them shipped here.

Regards,
Sanyam.


(Alex) #133

Curious how much it costs nowadays?


(Andrea de Luca) #134

Got the mobo (brand new in OEM package, Fujitsu D3128-B2) for 140 euros, the cpu for 80 euros, the mem (six 16Gb modues in triple channel) for 220 euros.


(Robert Salita) #135

Where do I source these components at your prices? Not ebay, not amazon, so where?


(Andrea de Luca) #136

I bought all of them on ebay (Europe). You’ll have troubles to find these prices at your first try. Have patience and check ebay briefly every day.

Note that as of today, Xeons E5-v3/v4 are starting to have very good prices on ebay.


(Sanyam Bhutani) #137

Hi All,

I’m trying a bit of MP Training myself however I’m pretty sure that I’m doing something very wrong.

Here is the code snippet that I’m using, I’m training on CIFAR-10. I’m at CUDA 10 w the latest drivers and CuDNN

np.random.seed(42)
data = ImageDataBunch.from_folder(path, train=train, valid_pct=0.2,
        ds_tfms=get_transforms(), size=224, num_workers=4).normalize(cifar_stats)

learn = create_cnn(data, models.resnet50 , metrics=accuracy)
learn.fit_one_cycle(30)

learn = to_fp16(create_cnn(data, models.resnet50 , metrics=accuracy))
learn.fit_one_cycle(30)

I don’t see any speedup when doing this.

Even

learn = create_cnn(data, models.resnet50 , metrics=accuracy).to_fp16()
learn.fit_one_cycle(30)

gives the same results.

Am I missing anything here?


(Ajay Arasanipalai) #138

Does anyone know how the new RTX 2060 compares to the other RTX-series cards for deep learning?


(Robert Salita) #139

I’ve been looking for RTX 2060 ML benchmarks but haven’t seen any. My guess is they’ll be predictable, faster than 1060 … I’m hoping the GTX 1160, 1170, 1180 will be targeted at the ML market.


(Ajay Arasanipalai) #140

I have a GTX 1060 now. Will there be a significant difference in performance (for deep learning)? More importantly, how would the RTX 2060 compare to the RTX 2070 in terms of price: performance?


(Andrea de Luca) #141

driver version?


(Peter Kelly (OZ)) #142

Hey Ilia, PLEASE forgive my tardy responce; I have FINALLY got my stuff together. I had to “upgrade” the motherboard to get UEFI bios. The machine is an old Asus Z170-P D3 with 11G ram, i7-7700k at 4.2, Sam. 970EVO 2TB NVMe (WOW-Fast), Gigabyte RTX 2080Ti and a clean install of Win10Pro, Visual Studio Pro, Anaconda Navigator, and FastAI - Which all worked, following Professor Howard’s instructions, “OUT OF THE BOX” - NO EXTRA FIDDLING. It got through "catsanddogs in about 15 mins! However, I noticed that, using “Task Manager”, the CPU was at around 99%, and the GPU was at about 0%, unless I scrolled the screen! All of this has let me to BEGIN reading the docko; yep, at almost 71, I am getting smarter?
If it wasn’t for a friend (Pi, aka TT) putteng on to our AI Rock Star Professor Howard a year ago, I would be ageing gracefully with a happy wife. Well, that was the plan.
I am also looking forward to reading EricPB’s stuff (he does look to be a cool dude) on how to get the GPU going (yep, it IS in the docko. but I can’t find it).
I’ll let you know how it ticks with the 2080, although it sounds (reads) as though you have long since passed me. Now it is up to me to catch up.
Cheers Ilia,
pk.