Video: What does a $3000 GPU look like? NVIDIA TITAN V Unboxing and Teardown!

Subject: Graphics Cards | December 12, 2017 - 07:51 PM |
Tagged: nvidia, titan, titan v, Volta, video, teardown, unboxing

NVIDIA launched the new Titan V graphics card last week, a $2999 part targeted not at gamers (thankfully) but instead at developers of machine learning applications. Based on the GV100 GPU and 12GB of HBM2 memory, the Titan V is an incredibly powerful graphics card. We have every intention of looking at the gaming performance of this card as a "preview" of potential consumer Volta cards that may come out next year. (This is identical to our stance of testing the Vega Frontier Edition cards.)

But for now, enjoy this unboxing and teardown video that takes apart the card to get a good glimpse of that GV100 GPU.

A couple of quick interesting notes:

  • This implementation has 25% of the memory and ROPs disabled, giving us 12GB of HBM2, a 3072-bit bus, and 96 ROPs.
  • Clock speeds in our testing look to be much higher than the base AND boost ratings.
  • So far, even though the price takes this out of the gaming segment completely, we are impressed with some of the gaming results we have found.
  • The cooler might LOOK the same, but it definitely is heavier than the cooler and build for the Titan Xp.
  • Champagne. It's champagne colored.
  • Double precision performance is insanely good, spanking the Titan Xp and Vega so far in many tests.
  • More soon!

View Full Size

Source: NVIDIA

December 12, 2017 | 08:27 PM - Posted by extide

Very cool, surprised you guys got that thing so quickly!

December 12, 2017 | 10:02 PM - Posted by Nice3000ComparedTo10000orMoreForQuadro (not verified)

"•This implementation has 25% of the memory and ROPs disabled, giving us 12GB of HBM2, a 3072-bit bus, and 96 ROPs."

So the GV100 base die has more than 96 ROPs available and this SKU has the same number of ROPs as some of the GP102 based Quardo/SKUs had available and it's time to look at what the total numbers of ROPs may be available on GV100. The GTX 1080ti is based on a GP102 die with 88 ROPs enabled while the Titan X and XP also based on GP102 has 96 ROPs available. It's interesting that this Titan is GV100 based and not a GV102 variant based like the previous generations of Titan. I Guess Nvidia is trying to get Volta Tensor Cores into the hands of more developers that way and who knose what features GV102, GV104, GV106/108 may retain compared to GV100.

Not only machine learing can be done on any Tensor/Matrix math cores as GPU/Graphics applications can make use of matrix math for effects filtering. AI edge detection may come in handy on the Volta for graphics applications and AI can be used for acceleration of some other graphics processing functionality like converting a bitmap image to an SVG(Edge detection is needed there for sure). So Volta's tensor cores do not know that they are for AI they just do the matrix math that they are told by the software to do. And I'm sure that just as Shader cores can be made to work on general math acceleration that the Tensor Cores on Volta can be made to do some Image Filtering workloads along with even AI accelerated Image filtering tasks.

Nvidia needs to develop more graphics processing related functionality into its libraries that are available for Volta's Tensor Cores and for sure the same AI accelerated edge dection that is needed for self driving cars can be used for edge dection on Images for Graphics art or even gaming uses.

December 12, 2017 | 10:05 PM - Posted by Nice3000ComparedTo10000orMoreForQuadro (not verified)

Edit: knose
To: knows

December 12, 2017 | 10:31 PM - Posted by Anony mouse (not verified)

The drivers are under GeForce

Product Type: GeForce

The Titan Xp is under 10 Series

The only thing different is Series. I suspect thats because its Volta where the others are Pascal

December 12, 2017 | 10:58 PM - Posted by throAU (not verified)

77Mh/sec on ethereum at 230 watts or something.

this is a mining monster.

December 13, 2017 | 06:56 AM - Posted by booman (not verified)

good job Nvidia, you could have charged for 6000$ for this !
these mindless trash will keep following the marketting plan anyway. Oh and super idea to charge them for a 1HBM chip that is there and they will never going use...

December 13, 2017 | 11:32 AM - Posted by AltruisticReasonsNotSoMuchFromCUDAman (not verified)

That still gives them 3 JEDEC HBM2 stacks each with that 1024 bit(3072 bits total) wide interface. The JEDEC standard interface of 1024 bits offered for each HBM2 stack so each of those 3 HBM2 die stack's 1024 bit interface is, according to the JEDEC HBM2 standard, subdivided into 8, 128 bit independent channels. And according to the JEDEC HBM2 standard each of those 128 bit channels on the newer HBM2 standard only can be further subdivided into 2, 64 bit pseudo channels for finer grained memory access.

And Nvidia has not charged for the one HBM2 stack not enabled as that's most likely a defective part somehow so has been binned for sale at a lower price point. This Volta SKU is actually based on a GP100 base die design and not a GV102 base die as was the previous Titan X/XP variants that where based on the GP102 base die design. So this is full GV100 silicon and that's a very large amount of shader cores to offer for $3000 price point and those Tensor Cores are there also. So this Vega SKUs is no different than AMD's Radeon Pro Duo SKU in that the Radeon Pro Duo was also marketed as a lower cost software development GPU for software that would be run in a production environment on the more expensive Professional products Radeon Pro WX(formally branded FirePro) or Quadro SKUs that cost many times more.

The Fact that Nvidia actually offered up its full GV100 dies at such a low price(Low by Nvidia's standards) is the most interesting development and that points to Nvidia's desire to get its GV100/With Tensor Cores product into the hands of as many developers as possible just like AMD/RTG did with its Radeon Pro Duo SKUs that allowed developers to get even more shader cores 8192 shader cores and 128 ROPs and double everything else for a very low price point.

There are some Production workloads in the Real World and not gamer fantacy land that have to be run on the expensive production certified Radeon Pro WX/Quadro variants because the Professional variants have all the costly certified drivers and the Error Correction hardware enabled and the driver/firmware cost millions of extra dollars to have certified for the Radeon Pro WX and the Quadro variants that run structural engineering workloads and airplane design workloads where mistakes can result in lives being lost.

So these lower cost Volta SKUs and Radeon Pro Duo SKUs can be of great and more affordable use by software developers targeting the software that will be run on the required professional Production GPU variants those Quadro and Radeon Pro WX variants that have all the error correction bells and whistles turned on. Oh and software/firmware support over the life of a computing product costs many time over what the hardware costs and the professional GPU SKUs cost so much to cover the Software/Firmware certification/validation costs and the extended support that comes with the professional Radeon Pro WX and Quadro branding! So those professional market products are costly to support and costly for driver development certification/validation costs and 5+ year product support cycles that come with the $10000+ SKUs.

For Nvidia this is a good job in offering a full GV100 based product for $3000 and Nvidia is not doing this purely for altruistic motives and neither did AMD with its Radeon Pro Duo SKU. Then again IBM did not force Intel to cross license the x86 16 bit ISA with AMD and others for any altruistic reasons either even though that lead to the x86 PC market that we have today.

Your hardware is worthless without developers, developers, developers, as the ranting monkey would often scream. So Nvidia needs the academic folks to get their hand on thes lower cost Volta variants and develop, develop, develop those CUDA applications that Nvidia needs for these things to work doing lots of things and CUDA is Nvidia's lock-in over the OpenCL/Open TensorFlow ecosystem that AMD/Others will be using!

December 13, 2017 | 02:28 PM - Posted by Anoan (not verified)

Nvidia shillbot?

You know that TMSC make these dies for 42.98$ to be exact.

now get back to your desk! #nvidaPR.

December 13, 2017 | 05:47 PM - Posted by NoShillzBotsButGPUsAreForGraphicsAndComputeNotGaming (not verified)

I'm no shill for Nvidia or AMD and I sure as hell do not care a rats A$$ about gamers of the Green or Red variety. The Radeon Pro Duo is as nice for what it offers software developers as this Volta offers for developers. And this Volta offers Plenty for potential graphics(Non Gaming) usage that does not include any redneck entertainment FPS/Racing game dung! Those Tensor cores have Graphics processing potental using stright up matrix math or TensorFlow(Matrix Math) AI libraries for AI based Image processing via Gimp/Adobe/Other graphics software.

GPUs are being taken back for productive usage from the Gaming Hicks and both Nvidia and AMD are moving in that direction far away from any damn dirty gaming inbreds and their crappy FPS(Frames Per Second) ePeen obsession.

Let's look at the post that you replied too Mr. Pecker S. Wood! The world does not need gaming to function and Nvidia and AMD Only use that gaming market to get rid of those poop dies fed to the gaming pigs!

December 13, 2017 | 02:31 PM - Posted by bobb (not verified)

Dude just go away please... enough PR we know that your are an Nvidia PR.

December 13, 2017 | 05:52 PM - Posted by NoShillzBotsButGPUsAreForGraphicsAndComputeNotGaming (not verified)

Bugger off AMD Gaming Redneck and take the Nvidia Green Toothed Gaming Peckerwoods with you whe you go!

Volta is not about any NASCAR round and round inbred entertainment.

No one is trying to sell you a GPU it's jut that gamers exist to eat the poop dies but the professionla markets are the ones paying the bills so take you Red and Green Team nonsense over to ESPN!

December 13, 2017 | 02:34 PM - Posted by Dnoa (not verified)

omg... enough with the PR already ! just milk the cow as long as you got the market for yourself ! #Nvidia

December 13, 2017 | 06:01 PM - Posted by NoShillzBotsButGPUsAreForGraphicsAndComputeNotGaming (not verified)

GPUs are for compute and non gaming graphics and gaming Gits get the poop dies from AMD and Nvidia!

This Volta GPU is not about gamers and their usless gaming habit, you Gaming Peckerwoods do not have a single brain cell counted among the entire lot of you gaming fanboy hayseeds.

Poor gaming scrubs you get the poop dies that do not make the grade for proper professional GPU usage. And you'll fight wuth the miners for those poop dies like a bunch of dung beetles!

GPUs are for compute/non gaming graphics usage and those professional markets can and do pay a proper mark-up as opposed to some dirty gaming scrubs who whine about paying AMD enough to cover its Vega costs, and AMD needs to price it's gaming poop dies higher so Nvidia can price its gaming poop dies even higher!

December 13, 2017 | 02:37 PM - Posted by Moona (not verified)

wow.. so much work gone into this article of yours.
Thats right ARTICLE not a comment.. your employer must be proud of your work !

December 13, 2017 | 06:17 PM - Posted by NoShillzBotsButGPUsAreForGraphicsAndComputeNotGaming (not verified)

Volta GV100 for 3 Grand is a bargain for development just as AMD Radeon Pro Duo is a bargian for what it offers for developing software targeting the professionl GPU SKUs that cost 3 times as much if not more.

I'd say hat the Radeon Pro Duo is better at DP FP than even this Volta SKU but Volta has Radeon bettered in the Tensor core area currently.

But I'm not employed by any GPU makers and GPUs are not really about gaming or gaming scofflaws and their wasteful gaming habits. You gaming peckerwoods need to all be inducted into the military as you are mostly wortless hicks that only add to the crime rate. So let's get you as government issue and going out vertical and coming back horizontal. Let's get you overseas and killing other rednecks and getting you killed by those other rednecks and earing the investors of Amalgamated Body Bags a good return on their investment.

The professional markets are getting GPUs out of the hands of the useless gamers and back into the productive fold of the scientists, artists, and the other productive folks and the GPU market away from of the wasteful hands of the mindless gaming gits!

December 19, 2017 | 08:12 AM - Posted by NascarFan1996 (not verified)

"productive fold of artists"
LOL now im sure you are retarded.

December 13, 2017 | 06:59 AM - Posted by ted (not verified)

Seriously... why is that HBM there if I cant use it ? wtf !
BTW this gpu wasnt supposed to be the new GTX 2080 ? and the full die version the GTX 2080Ti ??

December 13, 2017 | 07:02 AM - Posted by noF (not verified)

full Volta : GTX 2080Ti
this one was supposed to be : GTX 2080

why the change of plans ?
Well if you got all the market for you.. why not fuck the shit out of these people just because you can !

December 13, 2017 | 07:03 AM - Posted by beebs (not verified)

sadly ... that is so ture.

December 13, 2017 | 07:05 AM - Posted by beebss (not verified)

*TRUE

December 13, 2017 | 07:08 PM - Posted by CNote

Calm down Sybil

December 13, 2017 | 01:49 PM - Posted by Anonymous2 (not verified)

1180 Ti.

December 13, 2017 | 08:37 AM - Posted by psuedonymous

Like with Pascal, I doubt we will see a GV100 based gaming card. Too much die area wasted on Tensor Cores & DP & ECC, and HBM2 is too expensive for consumer cards (you get more GDRR5x/6 at the same speed for less money).

December 13, 2017 | 01:45 PM - Posted by GamingPoopFerGamerz (not verified)

Gaming scum will get the GV104 and the GV106/GV108 based variants with much less compute and maybe no Tensor cores. But as always there will be more ROPs to fling out plenty of frames per second(FPS) so Bubba Monobrow gamer can have his bargging rights. You See those bigger FPS numbers mean more to Bubba knuckle dragger than actually playing the games on that GPU. So Bubba likes that GPU/Rig and them flashing idiot lights more for show than actual play.

A Gamer's real function is to eat those poop dies that do not make the grade to be sold in the professional SKUs. Those poop dies are what fodder the gaming pigs and they love to wollow in the feces(Poop Dies) filled muck and eat that slop, that sometimes makes more than a few trips through old Bubba's intestinal tract.

Now Nvidia will after GV104 and GV106/108 have all been released to market will invariably be building up a stock of high quality GV102 based binned dung to tack on that "Ti" branding and get even more ROPs from that high quality droppings to feed to them gaming pork bellies at top dollar feed(poop die) prices.

December 14, 2017 | 12:27 AM - Posted by throAU (not verified)

So uh, explain non-top of the range workstation GPUs with "poop dies"?

December 13, 2017 | 07:20 AM - Posted by Anony (not verified)

"full Volta : GTX 2080Ti
this one was supposed to be : GTX 2080

why the change of plans ?
Well if you got all the market for you.. why not fuck the shit out of these people just because you can !
"

100% true

December 13, 2017 | 09:52 AM - Posted by Jabbadap

Could you run Beyond3d forums DimitryKo's Direct3D 12 feature checker with it?

Other thing I would like to know for it is the fp64 performance, were there DGEMM test for gpu?

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

By submitting this form, you accept the Mollom privacy policy.