|
I don't know too much about virtualization but how transparently virtualizable are multi core multi GPU setups these days? Is there a big performance hit? Put another way, would it be possible to create a multi seat gaming setup with Threadripper and multiple Nividia cards?
|
# ¿ Jun 5, 2017 08:02 |
|
|
# ¿ Apr 25, 2024 12:25 |
|
how cleanly does (python) scientific code parallelize? i'm tempted to throw money at a ryzen r7 1700 as an upgrade over my current i5 4590; would I get close to 2x performance moving from an i5 4 core to ryzen 8 core for hobby machine learning workloads (e.g., sklearn, xboost)?
|
# ¿ Jun 14, 2017 06:56 |
|
yah i'm aware of the GIL - the code and frameworks I use spin up new processes. my question is whether i'll get a straight 2x performance boost moving from an i5 4 core to ryzen 8 cores or do the architectural differences blunt that substantially.
|
# ¿ Jun 14, 2017 07:28 |
|
Paul MaudDib posted:Matrix math tends to parallelize real well thanks to time-tested primitives like BLAS/LAPACK (which Numpy hooks as native C). I'm guessing if you're getting good scaling (4x speedup) on a 4-core already then you would get 8x out of Ryzen, yes. awesome thanks. and is there a sweet spot between the 1700, 1700x, and 1800x models?
|
# ¿ Jun 14, 2017 07:37 |
|
Machine and deep learning are pretty resilient to imprecision (and inaccuracy) if you set up the models right. Anyway it's hobby stuff shrike82 fucked around with this message at 08:34 on Jun 14, 2017 |
# ¿ Jun 14, 2017 08:31 |
|
Klyith posted:XoxOX0XXX I might have missed the conversation but what the hell is that
|
# ¿ Jun 17, 2017 04:24 |
|
I just set up a Ryzen 7 1700 with a 1080 Ti. I've installed the latest chipset drivers. Is there any other setup I need to do to optimise my system?
|
# ¿ Jun 22, 2017 10:20 |
|
Thanks, I've set XMP for my ram. I'm wary about overclocking the CPU, I understand that doing so disables idling?
|
# ¿ Jun 22, 2017 16:51 |
|
SamDabbers posted:Use P-State overclocking if your BIOS supports it. It's probably in a menu under AMD CBS. thanks man, i've OCed to 3.7GHZ. Dumb question but how does overclocking here differ from the max turbo core speed of 3.7GHZ that the 1700 advertises? Does that also go up proportionately?
|
# ¿ Jun 22, 2017 23:09 |
|
I built a R7 1700 system to replace my i5 -4590 (4C/4T) desktop as a home workstation/gaming PC and am pretty chuffed with the performance. Gaming flies and more impressively, my hobby machine learning code cleanly parallelizes from 4 to 8 cores for close to 2x the performance. Scikit python code that used to take 2h22m per run now takes slightly over an hour. One downside is I can't overclock my PC when running full on all 8 cores using the Wraith Spire cooler. I hit 80+C and the system shuts down for overtemp protection. I've noticed that the CPU idle/load temps don't scale up linearly with the OC rates - I tried overclocking from 3.7GHZ down to base and there's a sharp rise in temps once you overclock at all. I'm fine with stock for now (hitting >100FPS at 1440P with a 1080Ti for almost all games) but am thinking of moving to a custom cooling solution down the road. What's a good/quiet solution - I was looking at the Cryorig H7 Quad or an entry watercooling solution, the Coolermaster Masterliquid Lite 120.
|
# ¿ Jun 24, 2017 22:49 |
|
Paul MaudDib posted:Turning OC mode on at all disables all of the power-throttling. This is another reason to go with the 1700 over the 1700X, you can't underclock the 1700X very easily... Unless I'm mistaken about pstate overclocking, that's not true?
|
# ¿ Jun 25, 2017 03:57 |
|
A casual Google search didn't bring up anything conclusive - would populating all 4 of my ram slots (from my current 2) mean I can't run them at native 2933mhz? The manual for my asrock b350 mobo seems to point at lower speeds for 4 sticks versus 2 but I'm not sure.
|
# ¿ Aug 2, 2017 10:44 |
|
Bizarre that he's never across this issue given he's spent a lot of time benchmarking Intel CPUs in the past.
|
# ¿ Aug 6, 2017 00:37 |
|
lol when people do poo poo like this
|
# ¿ Aug 7, 2017 16:24 |
|
#fakenews that image is apparently from a Thai bitcoin facility from Nov 2014 https://coinjournal.net/cowboy-miners-facility-burns/
|
# ¿ Aug 10, 2017 03:06 |
|
there's probably a big overlap between white libertarian coinbros and sex tourists that go to Thailand. IIRC, the Dutch owner of a darknet bitcoin-based market was recently arrested in Bangkok
|
# ¿ Aug 10, 2017 03:30 |
|
lol
|
# ¿ Aug 11, 2017 07:46 |
|
if u're using linux, could you try running the phoronix test suite? pretty simple to use - package install and you can run the entire test suite from a single CLI command. curious to see the scikit-learn performance
|
# ¿ Aug 14, 2017 05:36 |
|
Look at the viewership numbers for Linus versus every one else though.
|
# ¿ Aug 15, 2017 02:04 |
|
Speaking of gaming, the new Assassin's Creed game scaling well with cores is a good omen for Ryzen aging nicely as a gaming platform.
|
# ¿ Nov 1, 2017 01:51 |
|
RE the AC:Origins benchmarks, the 4C/8T 7700K being 50% faster than the 4C/4C 7600K is pretty crazy. It's too bad they don't have the 8600K on the list as a point of comparison against the 8700K.
|
# ¿ Nov 1, 2017 12:07 |
|
We should be able to see "clean" benchmarks once someone cracks it.
|
# ¿ Nov 1, 2017 15:52 |
|
lol neither the crack scene nor Ubisoft are trustworthy but the latter just released a statement about thisquote:Ubisoft: AC Origins’ DRM On PC Has No “Perceptible Effect” On Game Performance
|
# ¿ Nov 1, 2017 16:10 |
|
Combat Pretzel posted:Gotta admire their tenacity about this DRM thing. The more complex versions don't delay pirated releases significantly. They might as well just drop in some token copy protection scheme to make casual copying impractical, because that's what the complex ones still only do, and leave it at that. Not to defend DRM that just wastes cycles but AC:O is apparently uncracked a week on while Wolfenstein literally got cracked two days before the game actually launched with their CTO whining about piracy. Watch Dogs 2 was another Ubisoft release which managed to delay a crack release buy a couple months. There was an interesting article on arstechnica a while back that delaying a cracked release by a couple months results in a material impact on sales.
|
# ¿ Nov 2, 2017 00:22 |
|
Game released Nov 15, 2016. First crack was Jan 18 apparently.
|
# ¿ Nov 2, 2017 00:37 |
|
It'd be hilarious if AMD does a "Powered by AMD" technical partnership with game devs to drive multi process utilization by spamming DRM VMs.
|
# ¿ Nov 2, 2017 02:51 |
|
Lol at Paul trying to spin this as a loss for AMD
|
# ¿ Jan 4, 2018 04:25 |
|
SwissArmyDruid posted:Okay, so I have watched the video now, and what he says is very convincing.... but it still leaves the question: I have trouble saturating my 4x 1080 Ti with my 16C/32T 1950X on certain specific ML workloads
|
# ¿ Jan 26, 2018 16:31 |
|
Volguus posted:What's the point of these tests? They cannot provide any meaningful information on anything. Centos 7 has very different libraries (glibc, gcc at least) than fedora 27, than ubuntu , than debian testing, than opensuse Tumbleweed. Not to mention default filesystem and so on. It's comparing apples and asteroids. welcome to phoronix
|
# ¿ Jan 28, 2018 04:45 |
|
I got bitten in the rear end with the ThreadRipper/IOMMU issue for the first time - spent the entire weekend trying to debug why a dockerized tensorflow GPU model that worked on my office machine (7820X) wouldn't work on my home ThreadRipper setup. Spent two days thinking it was a code issue without much success until I tried disabling IOMMU in Grub, worked perfectly.
|
# ¿ Apr 23, 2018 00:45 |
|
Take a look at AMD's revenues compared to Intel's. AMD could dominate benchmarks across their offerings over the next couple years and still fail to make a dent on market share. Speaking as someone with a 1950X at home, cost-to-performance isn't the only metric that's important when it comes to big iron or enterprise deployment.
|
# ¿ Apr 27, 2018 23:27 |
|
Subjunctive posted:500-1100% on DL tasks between Pascal and Volta) LMAO, I'm curious what "DL tasks" these are. Most benchmarks show a 10-30% difference for training/inferencing between the Titan V and the 1080 Ti. shrike82 fucked around with this message at 01:17 on May 6, 2018 |
# ¿ May 6, 2018 01:13 |
|
yeah that's bullshit marketing which isn't borne out in actual benchmarks. We've actually looked at switching from 1080 Tis to the Titan Vs for our Dev ML servers, the latter aren't much faster. A big part is the fact that Nvidia crippled FP16 throughput for consumer cards - but that's not used in actual training/inferencing.
|
# ¿ May 6, 2018 01:28 |
|
Well I'm guessing you're not a technical guy - 1. Don't take marketing "benchmarks" at face value especially when the X-axis is "time to solution" (presumably convergence) and that can be manipulated 2. They compare the V100 to the P100 when the 1080 Ti is faster than the P100 and cheaper from a TCO standpoint if you're doing dev ML work 3. This is an example of a real-world benchmark 4. ResNet-50 is a pretty small (i.e., toy) image architecture by modern standards
|
# ¿ May 6, 2018 01:39 |
|
I'm more surprised about the mention of an SSD tbh
|
# ¿ Apr 16, 2019 14:26 |
|
I'm pretty curious about the performance of AMD hardware for next gen consoles - I'm guessing 4K60 is out of the question?
|
# ¿ May 4, 2019 06:45 |
|
How tricky is it to replace a CPU? I'm thinking of upgrading a B450 board with a 1700 with a 3000 series CPU. I'm wondering whether I should get a new gen mobo as well (GPU is an RTX Titan, 64GB ram). I vaguely remember the (3rd party) CPU fan clamp being a bitch to install. 1. Do people normally remove the parts surrounding the CPU (GPU, ram) to help with the install? 2. Are there any issues with the OS on a CPU change? 3. Can I expect to get the full performance of the CPU with a 1st gen board?
|
# ¿ May 29, 2019 02:16 |
|
I think his point is that Cinebench scales well with cores. Intel CPUs still beat AMD ones core for core.
|
# ¿ May 30, 2019 04:53 |
|
I have a 1950X so you're preaching to the choir
|
# ¿ May 30, 2019 05:47 |
|
|
# ¿ Apr 25, 2024 12:25 |
|
There's a limit to how much Nvidia can price gouge - we saw that with the dismal sales of the 2000 series.
|
# ¿ Jun 16, 2019 04:08 |