The absolute state of computing in the late 2010s:
>Nvidia Titan V has been churning out wrong answers for scientific calculations
>The report suggests that Nvidia Titan V “gave numerical errors about 10 percent of the time”.
>These calculations are supposed to output identical results but that is not the case with Titan V.
segmentnext.com
Other urls found in this thread:
nvidia.com
nvidia.com
anandtech.com
my.mixtape.moe
pcgamer.com
twitter.com
>gaming card
>scientific calculations
The fuck did you think was going to happen?
Get fucked nvidya
Can't wait to pick these up from the bargain bin
>>gaming card
nVidia SAID otherwise famalamadingdong
boring story
it doesn't affect normal day to day use and anyone buying a titan v isn't doing heavy calc, they're playing pubg at baely 60fps
>because nVidia, hence correct
The Titan V is not being marketed as a gaming card. It's listed under the AI products and deep learning products. All the blurbs say nothing about gaymes, it's all marketing speak about 'driving breakthroughs in all industries' bla bla bla.
>muh nvidia said
No ecc, no 10bit opengl, gaming drivers, and gaming tier support.
It's a fucking gaming card.
>nvidiots are unironically defending nvidia's version of FDIV bug
wew
Titan isn't gaming oriented series, it's advertised for small business.
It getting gaming tier support doesn't change the fact it's advertised as professional one.
>nvidiots are unironically defending nvidia's version of FDIV bug
Nah. I'm just not a retard who pretends that it's not a gaming card.
>Titan isn't gaming oriented series, it's advertised for small business.
Uh... yeah it is. It's their meme tier gaming card.
>It getting gaming tier support doesn't change the fact it's advertised as professional one.
The fact that it gets gaming tier support means that it is nothing more than a gaming card. If you're dumb enough to believe the "advertisements" then you literally never needed a workstation card...ever.
>buy CPU
>oops, FDIV bug
>>NOT A BUG GOY, JUST BUY XEON WITH ECC RAM
This is you.
Actually it's more like Nvidia is lying out their asses about the Titan V. Titan V was not supposed to be a gaming card, Titan Xp is the latest gaming Titan. Go look at the Titan V pages vs. the Titan Xp pages
Titan V: nvidia.com
Titan Xp: nvidia.com
Which one of those is marketed more towards gaming? Nvidia did their best to strip all mentions of gaming and 'Geforce' from the Titan V marketing.
fucking libtard millenials looking for bullshit to complain about. Have it perform the calculation 10 times and use the most frequent result, bam no problem. But noo, 'muh evil company against communism'
Fuck me. Last weekend, took too many pills and tried buying one of these from the site (never expecting it to go through, they usually sell them to amazon who add $2k to price).
I wanted it to use with HTC Vive Pro when it comes out next month. Guess now I can either return it or be stuck with a $3k piece of shit.
>10bit opengl
10-bit/HDR output in OpenGL apparently works with Geforce cards now
Only in exclusive fullscreen though
I've seen it somewhere 25 years ago.
>Which one of those is marketed more towards gaming? Nvidia did their best to strip all mentions of gaming and 'Geforce' from the Titan V marketing.
>muh marketing
see
>No ecc, no 10bit opengl, gaming drivers, and gaming tier support.
>It's a fucking gaming card.
Maybe no one is shocked, because everyone who sees the article just shakes their head, and wonder why that researcher tried to use that card for that specific task.
>Why did he try to use a gaming card (is what most people ask themselves with a poised look).
I mean personally I wouldn't even use that card for video editing, or print, but this guy is fucking trying to calculate scientific shit for actual research. What the fuck!
Running a computer and anything on it is literally calculations bruh
Imagine your computer doing something wrong 10% of the time
You took pills and ordered a $3000 piece of hardware?
What kind of job do you work that allows that to happen?
what the fuck, so It was mining my buttcoins at only 90% efficency? impeach trump now
> $3,000 card marketed towards AI developers
> HBM 2.0 supports ECC natively.. Ngreedia niggers the stack so you have to buy a $10k Tesla V100
> Getting ass raped for $3k and realizing nvidia just sold you an expensive Geforce prototype overlclocked to the point of memory instability
>10-bit/HDR output in OpenGL apparently works with Geforce cards now
Only in exclusive fullscreen though
Sauce on that senpai? I knew that 10bit was only enabled for directx fullscreen for geforce, but I haven't read anything stating opengl.
>What kind of job do you work
It was the weekend mate, I made a bunch of shekels by cashing out on crypto by buying in at a ridiculously low price and selling a few das before it crashed. I'm so sick of the coinmining at the this point driving prices up.
Anyone know the difference between the Titan -v and Titan-XP? I'm basically an agorapgobe who doesn't leave the house and just researches investments 18 hours a day. The only thing I enjoy is my fucking VR and feeling like I'm on "vacation" like I used to go with my ex-wife before I got divorced at 27.
it's HBM2, why would you assume that they locked it out of ECC mode like they did? This is shit you can't find out until you install the card. Yes, it turns out the Titan V is more gaming card than workstation card, this however is not how the card was marketed. None of the official materials indicate the Titan V is a gaming card. You don't advertise deep learning software with gaming cards.
>Imagine your computer doing something wrong 10% of the time
Imagine looking at product specs, and then using the available information to choose the right tool to purchase for the job.
UC Bekerly grads at Greenbank do almost all of their number crunching on gaming cards, they even water cooled them, for non-mission critical gaming cards are way better value wise than muh Quadro/FirePro
AI and deep learning data sets being so noisy make those applications prime candidates for cheaper cards
You don't need ECC for AI and machine learning, much less 10 bit output (or any output at all) or OpenGL
anandtech.com
>For Pascal, NVIDIA is opening things up a bit more, but they are still going to keep the most important aspect of that feature differentiation in place. 10bit color is being enabled for fullscreen exclusive OpenGL applications – so your typical OpenGL game would be able to tap into deeper colors and HDR – however 10bit OpenGL windowed support is still limited to the Quadro cards
>potentially linked to the GPUs memory and these errors in calculations might have spawned due to the overclocking of the GPU.
Overclock HBM
Get garbage results.
Holy cows, TITAN V is garbage.
NVIDIOTS BTFO
did you rad the fuckign description on the site mate?"NVIDIA TITAN V is the most powerful graphics card ever created for the PC, driven by the world’s most advanced architecture—NVIDIA Volt..."
What exactly do you think "...graphics card ever made for the PC" means? They can't exactly outright say "We're going to fuck you for $1,500 extra shekels to have it 4-5 months early," so you get bullshit excuses like this.
Also, don't think it's a coincidence that the new HTC-Vive Pro is coming out in a few weeks, just when people will be lookign for improved cards for better results on VR.
I have a Titan X 12 GB,(now going for $1,800 on Amazon)and I got a message the other day that my setup "couldn't run the VR properly."
So people are dropping 3k on a gpu designed for scientific computing specifically and play shitty ftp games on them?
Try again sweetie
>fpsp
shills are waiting in the corner for damage control.
you are implying that gaming cards, which always share gpus with workstation/pro cards, is ok to have faulty FLOP alus.
pffff.. novidiots.
kek
IDK how it works on TITANV, and if it's an internal HBM2 thing or not, but on vega, that's how it behave while mining:
OC memory to 1025Mhz.
Get 41.5MH/s on eth until HBM temp hits 82°C
Then 38.6MH/s.
It seems to autoswitch to shittier timings. GPU-Z still reports 1025Mhz.
I've heard elsewhere that HBM2 is very temperature sensitive. If this timing thing is from AMD's memory controller, and not an internal HBM chip's thing, and NVIDIA didn't implement something similar, you've got your answer.
>people are dropping 3k on a gpu designed for scientific computing specifically and play shitty ftp games on them?
Itcomes to almost $3,300 with taxes and the rest. I bought mine for better VR results and soI don't have to wait until fucking Christmas and the the price either stays the same or goes up. The fucking thieves have done it with ever card since I remember.
>TITAN V(ideogames)
I rest my case.
Its a fucking housefire. I assumed that they would make a gaming chip with tensor cores removed and we would just get a streamlined 5,400 cuda die on a better node as the new flagship.
I assumed that......until Gameworks 2.0 was announced. They are making even worse proprietary gameworks code thats so convoluted and fucked up that it only runs on volta cards with tensor cores which means that the new volta design will incorporates these things WHILE also breaking game engines with their shitworks code that can only run on these cards.
its a nightmare. Tensor cores are worthless for gaming and all they do is as die size (cost) and extra power consumption and heat output which means a limit to boost clocks and OC and a higher TDP.
What the fuck nvidia? what the fucking fuck!?
>I've heard elsewhere that HBM2 is very temperature sensitive.
wider gap between HBM and GPU die combined with seperating the heatsink contacts so that each HBM die has its own freefloating heat pipe running up to the radiator or making an air gap in the water block so that the contacts on the HBM are not connected to the GPU contact.
why the fuck are people defending this by sperging out about "muh gaymer cards"
are you that much of a brainwashed consumerist
Yeah, let's do this geophysical simulation that takes two whole days ten times just to be sure, because the manufacturer introduced a uncertainty factor in the machine.
Better use it as quantum circuit simulator then.
This is why you use a card that supports ECC for that shit.
A calculator that misscalculates shit has no use, nor purpose.
few researchers are able to or need to shell out 10k for say a tesla when they do not need it
Well I've worked up the delta being 62°C gpu temp for it not going over 82°C.
So I just have fans go at whatever speed they need to maintain that. It's hellishly noisy, but it works for now, but I doubt it will in summer.
GPU is running at around 1350Mhz, so it's cooler than running it stock (obviously)
Lower than that, and MH/s start to drop.
Should have gone for the LC edition, in retrospect.
>Titan (Windows) XP
rest a bullet in your skull
how is ml and ai not heavy calcs
seriously i dont get how people don't get this concept
>nvidialets
Machine learning workload, especially training, isn't considered mission critical because getting a good model is hard to compute but easy to verify. As such Titan V does good in gaming and ML but overclocks poorly in "generic" scientific computation.
NO
It might work early on, but it will fuck up your getting from 99% accuracy to 99.99%
Because it will invalidate right models, and validate wrong ones.
99% is unacceptable high risk.
Every problem with Machine Learning as a field finds roots in interpretability--the ability to quantify and interpret actionable insights to improve a model.
In other words, ML is a field that mostly consists of tinkering.
"You added this extra layer in this place and it works better than it did before? Great!"
No explanation necessary. Or rather, the person who could rigorously explain why it works better would be a very rich man.
The reality is that machine learning isn't actually a quest to get the highest accuracy, but a quest to find interesting things to do with it. Which is why Ian Goodfellow, who invented Generative Adversarial Networks is famous, and not the umpteenth researcher who achieved state-of-the-art on CIFAR-10.
Titan V ships with HBM ECC disabled, you don't use it for FP64 crunching proper, unless you're a moron.
Basically be a good goy and pay $10k for PCIe Tesla V100.
>not spending $20k for AGP Edison W200
This is nothing but fake news garbage
Only 1 person whines to TheReg when no one else has problems with Tesla V100 or TITAN V, blatantly trying to cause trouble and fake news
Explain why?
Well I threw 99% out there as a wild guess.
Could be way lower, actually.
What's for sure is, it will take you way more time to get there than you should on such hardware.
Which means a more expensive power bill than you thought.
I'm not even sure if it would actually get any result at all.
The way it works is, 10% of your calculations are incorrect. If your model has a lot of calculations, you're fucked because you're bound to have errors in every fucking run.
>let’s just reduce performance by a factor of 10 to fix the problem
Whew, bet that sounded a lot better in your head
That's not how it works at all
I have four Titan Vs. They do fine in gaming/ML but overclock poorly with integer workload and presumably floating points as well. In several games I tested it can run at up to 1800MHz with no visible issue but incorrect integer results can be expected at above 1450MHz.
>In several games I tested it can run at up to 1800MHz with no visible issue but incorrect integer results can be expected at above 1450MHz.
So the XBAWKSHUEG die has it's downsides, I guess.
They bin everything.
Well, supposing any amount of your GPU calculations are wrong:
I'm not familiar with other machine learning frameworks, but in Tensorflow, the GPU is typically only used to calculate the gradient. Which means that at the very worst, N% of your datapoints will probably become outliers in the resulting gradient calculation.
The important part, which is the aggregation of gradients, happens on the CPU.
Keep in mind that machine learning is essentially applied statistics. We have ways of dealing with outliers, and given enough data we can essentially just steamroll the inaccuracy away.
As they say in chemistry: the method to dealing with pollution is dilution.
Well it depends.
If the cards has rounding errors that are still somewhat accurate, you might get somewhere.
If the cards throws garbage, you're fucked.
Because then, you get calculation done on your garbage data.
Which might throw garbage too, but that's irrelevant at this point.
A perfectly fine model will get invalidated.
While others get invalidated as well.
I can see how it would snowball to 0% accuracy, depending on your use case, where every fucking run gets a very low rating, and you're just stuck there.
Well, you'd think TitanV buyers would have reported it by now, so it's most likely bullshit ffrom someone overclocking it, and discovering it actually throws garbage before dying.
>titan
>pro
>ecc
>pro drivers
The state of Sup Forums
Titan V is literally marketed as prosumer card.
>titan v
>ecc
Idiot
That's the point.
ECC is baked into JEDEC HBM spec yet nVIDIA niggers it from Titan V because reasons.
>2018
I guess this is the year, right?
In this thread there are people DEFENDING A MALFUNCTIONING PRODUCT, and they DO IT FOR FREE.
Has the world gone mad? Did I forgot to take crazy pills?
To everyone saying this is fake news, here's how Titan V fails heavy integer load in under a minute with a mild +120MHz overclock
my.mixtape.moe
Depends on how wrong, doesn't it?
At least one company is already selling it as a game machine, this being "the tiki."
"Nvidia's surprise announcement and launch of the Titan V upped the ante for the fastest graphics card in the world. Former Titan cards cost $1,000, and the Titan X and Xp moved the target a bit further north to $1,199. With the Titan V, Nvidia goes for a completely different class of users, checking in at a breezy $2,999. Who could possibly need such a graphics card? Well, not gamers—or at least, not people who are only interested in playing games.
We didn't make Nvidia's shortlist for an initial Titan V sample—"it's not for gaming" was the curt response—but Falcon Northwest came through with a drool-worthy PC that kicks practicality to the curb and goes all-in on style and performance. The latest version of the company's Tiki naturally includes the Titan V, but the luxurious hardware doesn't stop there. The potent totem includes Intel's fastest CPU for gaming, the Core i7-8700K, paired with 32GB of DDR4-3200 CL14 RAM, and then forget about storage bottlenecks with not one but two Samsung 960 Pro 2TB drives, configured in RAID 0. And just for kicks, there's a slot-load DVDR for good measure."
pcgamer.com
(IF I KEEP THIS FUCKING CRD PROBABLY GONNA HAVE TO SHELL OUT ANOTHER $3K-#4K TO DO IT RIGHT)
The brainlet you responded to didn't grasp this user.
>ttps://www.pcgamer.com/the-3k-titan-v-is-the-fastest-graphics-card-even-though-its-not-for-gaming/
lol... overpriced scrunched up gaming rig.
AMD turbotards are just mad that the titan v is still better than anything amd has crapped out.
>hur durr let's try to ruin nvidia with this lame news
meanwhile people will just buy something else made by nvidia.
Oh god that thing looks so good
Up there with the GTX690
t. Samefag who is unable/pretend to be incapable of understanding marketing
>When you defend the $3000 price tag it's a pro card
>When you defend the lack of features and incorrect calculations it's a gaming card