Nvidia Tesla V100 Spotted in GeekBench With Staggering Numbers

guru3d.com/news-story/nvidia-dgx-1-with-tesla-v100-spotted-in-geekbench-with-staggering-numbers.html

Other urls found in this thread:

en.wikipedia.org/wiki/Tensor_processing_unit
browser.geekbench.com/v4/compute/compare/485469?baseline=1098943
semiwiki.com/forum/content/6936-ai-asics-exposed.html
wired.com/2017/04/building-ai-chip-saved-google-building-dozen-new-data-centers/
cloud.google.com/blog/big-data/2017/05/images/149454602921110/tpu-3.png
images.anandtech.com/doci/11367/volta_sm.png
hotchips.org/
hpcwire.com/2017/05/25/google-debuts-tpu-v2-will-add-google-cloud/
techrepublic.com/article/azure-how-microsoft-plans-to-boost-cloud-speeds-with-an-fpga-injection/
nvidia.com/en-us/data-center/dgx-1/
devblogs.nvidia.com/parallelforall/inside-volta/
twitter.com/NSFWRedditVideo

Cool. Can I get a $350 1070 now

>gookbench
Oh boy, here we go again.

btw are you even 18 OP?

THANK YOU BASED NVIDIA

>Volta poor navi

?

the staggering numbers are due to the tensor core compute path that is basically tons of dedicated matrix math calculators. These are used for Weak AI applications. Hilariously, it's also a place where tons of new entrants are entering into the game and likely where GPU manufacturers are going to get blown the fuck out in the AI market. Why? because its nothing more than a basic math circuit : en.wikipedia.org/wiki/Tensor_processing_unit

So, in a sense, this will be pure pottery as lots of new board manuf's will rise into this space and likely cross over into the 'GPU' space. Which is good... these fucks need a shakeup w.r.t to these insanely price cards/packages.

>likely cross over into the 'GPU' space
No you fucking retard.
Shader implementations are patented.

> new entrants are entering into the game and likely where GPU manufacturers are going to get blown the fuck out

Just like AMD with tons of experience in making GPUs is BTFOing Nvidia, amirite.

V100 is performing over 2X as fast as P100 in a range of compute cases and workloads, the tensor cores can't be responsible for this.

browser.geekbench.com/v4/compute/compare/485469?baseline=1098943

V100 has moar L1.
That's what responsible for bonkers GB numbers.

I seriously doubt that's what it is, really I doubt it can be attributed to just one thing.

Well Volta has a lot of thing related to compute.

>Literally can't even make a shitpost without mentioning AMD
You fuckers are so pathetic.

This kills the $20,000 big die bullshit. semiwiki.com/forum/content/6936-ai-asics-exposed.html

THE DIE IS REALLY BIG
PLS BUY

This is the exact same shit with Shilltel. They just keep pretending AMD is their only competitio, and not any number of other companies muscling in on their territory. "Intel is da best! NVidia is da best!", bleats the fanboy. "Nobody can possibly beat them at anything ever!". Retards like this need to be castrated for the sake of the gene pool.

>Getting buttblasted because someone insults or makes fun of AMD
Did yo uget lost on your way to reddit, kid? we make fun of everything in here, specially if you make shit products like AMD tends to do.

>l-lebbit!
(You)

lmao all this mad.
I seriously hope you at least get pay for defending AMD

>s-shill!
(You)

Stay mad hahaha.

>m-mad!
(You)

>pretends to not be mad
>goes on a (you) tirade just because someone made fun of amd.


super mad

>t-totally m-mad i s-swear!
(You)

What's even more funny is that it's not like that insult was directed at you, it was directed to AMD a company which you took personally and felt like defending agasin't

hahahaha

>m-m-m-mad!
(You)

still mad?
You've already wasted almost 20 minutes defending AMD.

(You)

>y-you are still m-mad i swear!
(You)

I want Titan XV for 4k@144hz

That would be $3k, gweilo.
The dies are big, you gotta pay!

>implying youre not mad
Explain this then

>Boost clock is 1455Mhz for volta so clock per clock volta is 15-20% improved from previous generation approximately

This, and
>the staggering numbers are due to the tensor core compute path that is basically tons of dedicated matrix math calculators.

So basically, it's fucking nothing but a core count bump.

And moar bandwith and larger L1$.
Still very decent when it comes to HPC.
And it's earlier than Vega20, so there's that.

Geekbench doesn't use the Tensor cores at all.

GB is cache-sensetive.

A 4k144 monitor would be at that price too. Lol

Some of the workloads saw ridiculous improvements, for instance the "Gaussian Blur" test went from 9.43 Gpixels/sec for P100 to 33.2 Gpixels/sec for V100.
This is solely due to the cache improvements?

Ye.
L1$ is also a part of unified memory now.
Much easier to manage.

volta is pascal which was maxwell 2.0 which was maxwell 1.0
the only NEW thing is the tensor core and that is only activated on deep learning since the ISA is being built only for that on tensor cores

the compute path is being done on async path also but emulated on the cpu which probably says a lot about the current state of novidia

1.7 Woodscrews Porkshoulders fuck off Huang.

THANK YOU BASED NVIDIA

>Still spamming this worthless link

No ASIC can compete with Volta that can do FP32, FP64, INT8 and Tensor operations in parallel

You only need to do meme matrix FMA shit for DL.
And you can do that with ASIC without paying the retarded GPU tax.

I want one :c

I want a dead one for keychain.
The die is really-really (really, i swear) big!

>n-nvidia is still better ..

Not with that $20k pricetag and orders of magnitude more power consumption and rack space it isn't.

wired.com/2017/04/building-ai-chip-saved-google-building-dozen-new-data-centers/

The question is where is AMD?

Too busy throwing money at visualisation to bother with meme learning.
And the market is about to be nuked by ASICs anyway.

Applying Infinity Fabric to GPUs

Geekbench doesn't use the Tensor cores, not a single test will run on them.

Volta is getting ~3X as much performance as Pascal with a less than 50% increase in shaders and only around a 30% increase in bandwidth. The changes between Volta and Pascal seem to be quite significant.

It's 2.3X, not 3X

The OpenCL score is only 1.7X, which make it seem likely that the CUDA path is utilizing the Tensor Cores.

Has Geekbench included support for CUDA 9 yet?

Which GB tests would even benefit from the Tansor cores?

literally all the compute tests would benefit from more flops

Tensor cores can only do half-precision FLOPS, the tests would have to be designed for that.

A 72% increase over Pascal in the OpenCL bench is still crazy since the V100 only has a 40% increase in raw performance over the P100.

A good, knowledgeable comment. Of course Sup Forums retards miss the point you where making. I'm sorry.

He's kinda wrong through.
Both Nvidia and AMD have a fuck ton of patents that do not allow anyone to enter the GPU market.

Probably the GV102

GPUs crossed over into compute.
I am speaking about the compute pipeline not the traditional GPU pipeline. You fucking retard.

So, think for a moment what happens if they keep their insane pricing up for this and someone offers the same computer w/ far more distribution features for a tenth of the cost...

RTG/Nvidia will be put on notice soon.

The big margins are in compute. Not in traditional GPU pipelines. There is already notable attention as of TPU (tensor core).

It's a combination of things centered on FP16/Tensor cores. it's a compute pipeline not a GPU pipeline. There are already custom asics that outperform the V100/P100 on power and compute :
cloud.google.com/blog/big-data/2017/05/images/149454602921110/tpu-3.png

images.anandtech.com/doci/11367/volta_sm.png


Both RTG and Nvidia milked compute for too long and thus are going to face real competition soon from new entrants. Other hardware companies did the same only to get blown the fuck away by Big Data companies spinning their own white box hardware with open and transparent software stacks.


Thank you and its fine. The timeline engages soon.
Just making my rounds.

>It's a combination of things centered on FP16/Tensor cores.

The Geekbench tests don't use FP16, the Tensor cores would be useless for those tests.

>semiwiki.com/forum/content/6936-ai-asics-exposed.html
Pretty much. So many players entering the game as the operations are basic matrix math. No need for a $20,000 die, insane power utilization or a traditional GPU pipeline when you have a better more efficient solution. You should take a look at custom hardware and compute network microsoft has cooked up using CPU/custom asics/FPGAs. In all honesty, the GPU guys got greedy and tried to milk it to hard.

They're still holding back tech that's decades old because they want to continue insane margins for arbitrary pro features that will soon be open sourced.

> Pro level card for AI/Compute that utilizes FP16/Tensor cores
> Geek bench
Kys

Were not talking about the consumer dies that are going to have the majority of this shit cut out of them. Stop harrassing me w/ this shithead nonsense and go read about what came out of :
hotchips.org/

>I am speaking about the compute pipeline not the traditional GPU pipeline. You fucking retard.
Not every compute pipeline is some combination of strictly specific matrix math you retard.

You responded to my comment ABOUT Volta's Geekbench performance, the whole fucking thread was started about Volta's Geekbench performance for fuck's sake.

How about you actually pay attention to the comments you respond to and shove that condescending attitude right back up your ass.

I think were done user. I don't think you know a single fucking thing about the compute pipeline in GPUs that's being used in HPC and AI nor do you know fuck all about the alternative architectures that are being used as alternative that a multiple folds more efficient at compute :

Here dumbass :
> hpcwire.com/2017/05/25/google-debuts-tpu-v2-will-add-google-cloud/

> techrepublic.com/article/azure-how-microsoft-plans-to-boost-cloud-speeds-with-an-fpga-injection/

> MUH VOLTA and shit
GPU manufacturers got lucky and stumbled into the compute market. It's not their domain. They went for high margin rape. As a result, the big data corporations spun their own hardware solutions that are far more power efficient and have far more compute performance.

Tensor cores are replacing FP16 because that's all the current weak AI algorithms need. Stop acting as if you know wtf you're talking about w/ your silly ass remarks and go read about what I've been talking about.

- Exit thread

DGX-1
nvidia.com/en-us/data-center/dgx-1/
Wtf does the headline of that page say?
ESSENTIAL INSTRUMENT OF AI RESEARCH
How much does a DGX-1 cost?
$129,000 - $149,000

What is the OP about? A DGX-1. What's in a DGX-1?
> from OP source link :
Now then, the Nvidia DGX-1 unit with Tesla V100 Spotted in GeekBench has 5,120 shader processors. For the record, a DGX-1 setup currently costs roughlt 129K and houses eight Tesla V100 cards, two Intel Xeon E5-2698 v4 processors, 512GB DDR4, four 1.92TB SSDs in RAID 0 and a power supply of more than three kilowatt. So let's call it what it is, a super computer in a box.
> eight Tesla V100 cards
> two Intel Xeon E5-2698 v4 processors
> 512GB DDR4
> four 1.92TB SSDs in RAID 0
and a power supply of more than three kilowatt.

> a super computer in a box.
> a super computer in a box.
> a super computer in a box.

> eight Tesla V100 cards
> eight Tesla V100 cards

What are you going to get as a consumer? A stripped the fuck down micro architecture 10 standard deviations from this card.

So, someone decided to run geekbench for shits and giggles on an super computer.
> oh muhhhh gawd Voltas performance
> 8 voltas in a super computer
> MUH GPU pipeline

How about you actually use your brain so I don't have to for you. It's a goddamn super computer that someone ran shitbench on so brainlets could salivate over a Flagship GPU meant for HPC market thinking somehow that's what they're going to get in a consumer vega thats cut the fuck down 50 ways to sunday.

I'm not being condescending. You decided to try to correct me and get in my ass and I'm calling you out and the other poster as dumb fucks who don't know wtf you're talking about.

...and we'd have it now if Vega wasn't such a fucking failure.

Holy fucking fuck AI memes is not the only usecase for GPGPU you retarded avatarfaggot.
You mean Vega20 is such a fucking failure that it's not out yet?
We're talking HPC dickmeasuring contest, child.

The fuck you're not being condescending and you are fucking wrong anyway, you literally said "the staggering numbers are due to the tensor core compute path that is basically tons of dedicated matrix math calculators" but the fucking "numbers" are from GEEKBENCH you idiot, the Tensor cores aren't the reason for the staggering numbers at all.

Now your going on about the consumer version for... who the fuck knows.

> eight Tesla V100 cards
> eight Tesla V100 cards
> eight Tesla V100 cards
> eight Tesla V100 cards

8888
8
8
8

> wew, staggering geekbench

Lets see how fucking retarded you can be and how long you'll try to save face.

My apologies, I'm sure people are buying it based on its Geekbench scores....

stop replying to bait dumb frogposter

love how theres nothing to reference those scores with like a 1080 or something so you just have to assume theyre good scores

Oh my fucking GOD you think that score is from 8 GPU's together. Jesus Fucking Christ. GB only tests one card, the score is from a single V100 you dumbshit.

it's 300k more than pascal tesla
frogposter is right though, it's only due to tensor cores which are worthless outside narrow tasks

consumer volta going to be boring as hell, we are back to 2009 nvidia times, nvidia is amd now
vega at least was entertaining

AMD spreading it's curse
intel is housefire company, nvidia returning to it's rebranding ways

There's some Titan X entries on the list of scores in the original article.

Reading around on Beyond3D, users there are saying the GeekBench scores have nothing to do with the Tensor cores but Sup Forums seems certain it does. So seriously, which Geekbench tests use any kind of operation the Tensor cores can run.

This.

They are waiting for 7nm as we can see their 14nm Vega suck down too much power

>Vega sucks down too much power
Stop buying consumer cards.

Except Torch, Tensor Flow and caffe are all officially only CUDA accelerated.

Nah I got a 1080 oc that uses 200 watt instead of 300-400+

Vega monumental fuckup is unexcusable

This is coming from a ex 390x owner and that card sucked down as much as Vega if not more

Yeah... They added 50% more cores for FP32, extracted some extra ILP, and/or bumped the clock rates up like 50-100MHz; nothing groundbreaking. You don't magically get more instructions processed just because muh generational leap.

There's no reason why computer vision (which is part of the GB suite) can't be performed via tensor cores though, but the results make sense for a direct scaling of the P100 architecture with or without the Tensor Cores. It's more of a matter of how nV will recruit it into the CUDA compiler, not whether or not it can do it, because of course it can. Laplace Sobel for example would benefit hugely from faster GEMM.

Google literally runs TensorFlow on TPU2.

V100 has 42% more shaders but only 40% more raw power (FLOPS) yet even in the OpenCL branch it's still getting a 72% higher score. The performance gain on GB most likely has nothing at all to do with the Tensor cores.

Someone at Beyond3D pointed to the improved L1 cache, improved memory controller and more flexible SIMD execution method as some of the likely culprits for the substantial performance boost.

damn, its actually been a decent amount of time since the last pure rebrand from nvidia right? at least on good cards

That, and 50% more memory bandwith.

4 years.

daily reminder that nvidia is getting away with selling a $600 midrange gpu because amd are such pathetic faggots

680->770
it's more of a refresh compared to how they used to do it before, at least 770 was cheaper
on side note: 770 killed any will to buy nvidia for me, 1080 looks very appealing right now, but scars are too deep

Wait for Nvidia Volta GV102 GTX 2080 Ti

devblogs.nvidia.com/parallelforall/inside-volta/

>The new Volta SM is 50% more energy efficient than the previous generation Pascal design, enabling major boosts in FP32 and FP64 performance in the same power envelope.

>Volta GV100 is the first GPU to support independent thread scheduling, which enables finer-grain synchronization and cooperation between parallel threads in a program.

THANK YOU BASED NVIDIA

At least they don't sit on their ass doing nothing like Intel

they kind of do nothing, but not because of lack of trying
honestly I just think there is not much can be done for GPUs at current node, math has it's limits
GPUs aren't very clever devices compared to CPU
AMD couldn't do same thing as nvidia while keeping all enterprise capability in one chip, they have no resources to design 3 lines like nvidia
still prim shaders might still surprise us all fingers crossed

They are doing exactly what nVidia does in Vega20.

vega 20 is fp64 right? is it even useful?

...at low nm processes?

>FP64
Yes.
>is it even useful?
Depends on how much money will AMD throw at software.

...

NO

Hello friend!
Stop evading the ban.