Nvidia launches Titan V (Volta GV100) graphics card

[HEXUS]

£2700/$2999 card uses same GPU as the Tesla V100 HPC accelerator.

Read more.

[Wozza63]

250w? That's incredibly promising for mainstream Volta GPUs in the future.

[CAT-THE-FIFTH]

From the AT article:

https://www.anandtech.com/show/12135...r-3000-dollars

Moving on and diving into the numbers, Titan V features 80 streaming multiprocessors (SMs) and 5120 CUDA cores, the same amount as its Tesla V100 siblings. The differences come with the memory and ROPs. In what's clearly a salvage part for NVIDIA, one of the card's 4 memory partitions has been cut, leaving Titan V with 12GB of HBM2 attached via a 3072-bit memory bus. As each memory controller is associated with a ROP partition and 768 KB of L2 cache, this in turn brings L2 down to 4.5 MB, as well as cutting down the ROP count.

In terms of clockspeeds, the HBM2 has been downclocked slightly to 1.7GHz, while the 1455MHz boost clock actually matches the 300W SXM2 variant of the Tesla V100, though that accelerator is passively cooled. Notably, the number of tensor cores have not been touched, though the official 110 DL TFLOPS rating is lower than the 1370MHz PCIe Tesla V100, as it would appear that NVIDIA is using a clockspeed lower than their boost clock in these calculations.

As mentioned earlier, NVIDIA is unsurprisingly pushing this as a compute accelerator card, especially considering that Titan V features tensor cores and keeps the TITAN branding as opposed to GeForce TITAN. But there are those of us who know better than to assume people won’t drop $3000 to use the latest Titan card for gaming, and while gaming is not the primary (or even secondary) focus of the card, you also won't see NVIDIA denying it. In that sense the Titan V is going to be treated as a jack-of-all-trades card by the company.

To that end, no gaming performance information has been disclosed, but NVIDIA has confirmed that the card uses the standard GeForce driver stack. Now whether those drivers have actually been optimized for the GV100 is another matter entirely; Volta is a new architecture, markedly so at times. Speaklng solely off the cuff here, for graphics workloads the card has more resources than the Titan Xp in almost every meaningful metric, but it's also a smaller difference on paper than you might think.

As for NVIDIA's intended market of compute and AI users, the Titan V will be supported by NVIDIA GPU Cloud, which includes a number of deep learning frameworks and HPC-related tools.

Its made on an improved TSMC 16NM node known as 12NM which has lower leakage:

https://www.electronicsweekly.com/ne...-node-2016-11/

The GP102 is 471MM2 and the GV100 is 815MM2 in comparison. Now look at the relatively small increase in FP32 numbers(which is most relavant for gaming).

This is more a FP64 and deep learning based computation GPU. It should be in theory be quicker than the GP102 going by the higher FP32 numbers and more bandwidth,but I think Nvidia Ampere is what most gamers might be looking at next year:

https://www.tweaktown.com/news/59816...018/index.html

I would expect far less transistors dedicated towards FP64 performance,etc and for it to have a much higher FP32 ratio compared to FP64 performance.

[Xlucine]

250w? That's incredibly promising for mainstream Volta GPUs in the future.

Not really - it's wider than the old titan, but also slower. This isn't looking like the same step-change in performance we got from pascal compared to maxwell so far, I'd want to see something bigger and faster for the same power before it suggests any great improvement

[LSG501]

If I had a use for the tensor cores which is basically the main selling point here then it's a bargain but for anything else it's a 'lol how much' response.

[QuorTek]

It is a professional card, price range is good for people that need this.. I wonder if AMD has something up their sleeve for this move, but one can only speculate.

[lumireleon]

so its single precision performance is comparable to VEGA 64? why would you buy this for rendering workloads?

[CAPTAIN_ALLCAPS]

It is a professional card, price range is good for people that need this.. I wonder if AMD has something up their sleeve for this move, but one can only speculate.

AMD still doesn't have an answer for Pascal yet so I don't see them responding to this colossus any time soon.

[blokeinkent]

Sorry? For a minute there I thought somebody said £2,7..... OMG! They did <sh*ts the bed>

[Troopa]

Hexus posting a story after 6pm on a Friday and after QOTW?

[CAPTAIN_ALLCAPS]

Sorry? For a minute there I thought somebody said £2,7..... OMG! They did <sh*ts the bed>

Expensive, but it is an actual pro card with a good amount of FP64 and compute - unlike the last few Titans, which merely hid nVidia's best gaming chips from the market so they could overcharge for them and pocket all the cash.

[CAT-THE-FIFTH]

Here are some synthetic benchmarks with the Titan V:

https://videocardz.com/74382/overclo...chmarks-emerge

So this is an overclocked card against an overclocked GTX1080TI which is not a fully enabled GP102.

Around 15% extra performance,and it seems the card runs at a much lower base and boost clockspeed than the consumer cards too.

Remember this has the following advantages over the 471MM2 GP102:
1.)A lower leakage version of TSMC 16NM.
2.)New uarch.
3.)HBM2 and significantly more bandwidth,and probably better compression.
4.)73% more transistors,and 76% more surface area at 815MM2
5.)Much lower clockspeeds than consumer Pascal cards,meaning you can drop power. The GTX1080TI base clockspeeds on many models is the same as the Titan V BOOST clockspeeds when it is overclocked.

This shows you adding extra stuff like increase FP16 and FP64 throughput,etc not only increases die size,but reduces gaming/mm2 significantly.

Even if this card were 40% faster in games it would have worse performance/mm2 than Pascal. Its not suprising as Nvidia made two large Pascal lines,ie,GP100 and GP102,which had exactly the same amount of shaders but the GP102 was a leaner GPU made for FP32 and consumer workloads.

This is why Nvidia has started specialising its highest end lines from Maxwell onwards:
https://www.anandtech.com/show/10222...-power-for-hpc

Meanwhile GK210 will be in an odd place as it will likely be the first NVIDIA GPU not to end up in a consumer card; prior to this generation every GPU has pulled double duty as both a compute powerhouse and a graphics king. But with GM204 clearly ahead of GK110/GK210 in graphics, GK210 seems destined to Tesla cards and at most a Titan card for the budget compute market.

The follow up GP100 never made it to consumer cards either:

https://www.anandtech.com/show/11102...s-quadro-gp100

The last time we checked in on NVIDIA’s Quadro workstation video card group, they had just launched the Quadro P6000. Based on a fully enabled version of NVIDIA’s GP102 GPU, the P6000 was the first high-end Quadro card to be released based on the Pascal generation. This is a notable distinction, as NVIDIA’s GPU production strategy has changed since the days of Kepler and Maxwell. No longer does NVIDIA’s biggest GPU pull triple-duty across consumer, workstations, and servers. Instead the server (and broader compute market) is large enough to justify going all-in on a compute-centric GPU. This resulted in Big Pascal coming to life as the unique GP100, while NVIDIA’s graphical workhorse was the smaller and more conventional (but still very powerful) GP102.

I would certainly expect a more gaming optimised card,made on the same process node,to be significantly smaller,and will be able to boost much higher,as I expect the clockspeed "sweetspot" to be significantly higher,since it is not held back by all the other non-useful stuff on the chip.

So the important thing is to consider,how much faster,a leaner FP32 focussed card will be.

Now imagine a Volta based GPU with stripped down FP16/FP64,no Tensor cores,etc and a higher base clockspeed?? Having said that rumours say the next consumer generation is called Ampere though.

Its also why I am worried about AMD,since they continue to try and use one "largish" GPU to tackle multiple segments,ie,a jack of all trades,master of none.

ATM,484MM2 Vega 10 is trying to fight:
1.)471MM2 GP102 which is primarily FP32 focussed
2.)610MM2 GP100 which has enhanced FP16 and FP64 support
3.)815MM2 GV100 which has enhanced FP16 and FP64 support

Vega 10 has greater FP16 and FP64 throughput AFAIK when compared to the GP102,but this seems to come at the expense,of a chip which is worse in gaming,and probably why performance/watt is poor on Vega64 as its clocked outside its comfort zone,and the Nvidia GPUs tend not to be. FFS,Polaris actually has higher performance/mm2 and performance/watt of a fully enabled Vega 10 is really not better than a RX480 or RX580,and if we want to get more accurate representation of where Vega should be clocked at,Vega56 looks more like the sweetspot clockspeed for the GPU.

I think if AMD has any chance of competing for the high end crown sucessfully ,they really need to consider having two top tier lines,ie,one FP32/consumer focussed and another more focussed towards commercial stuff. However whether they can afford to do this is another question(and CPUs are probably more important for them as a company).

Edit!!

Not saying they might be able to throw a curve ball but we will need to wait and see on that though!!

[CAT-THE-FIFTH]

Someone has started benchmarking their one:

https://www.reddit.com/r/nvidia/comm...ing_ask_me_to/

Deus Ex Mankind Divided
88 avg 66 min 112 max @ DX11, 1440p, ultra
86 avg 64 min 108 max @ DX12, 1440p, ultra

The game uses a standard benchmark anyone can run and these are the scores with my GTX1080FE and RX470 in March when I ran it in a system with an IB Core i7:

http://i.imgur.com/VTrgkpU.jpg



They have a Core i7 8700K. This is the Hexus review of the GTX1080TI:

http://hexus.net/media/uploaded/2017...bbbd59475c.png



Its at Ultra settings too but with a Core i7 7700k.

So basically a GTX1080 will get around 50~55FPS,and a GTX1080TI around 70~75FPS.

Overwatch, all epic settings

67 fps (my GTX 1080 was 40 fps) @ 5k while standing at the training level starting point
Between 55 and 67 fps @ 5k in King's Row during a QP match
Between 95 and 130 fps @ 4k in King's Row during a QP match

I'm not certain but I think a 1080ti get around 60fps on that setting? I guess this is only slightly better in term of video game. I understand this is not a gaming card, of course.

If you're referring to Overwatch, yes it gets about 60 fps with 1080 ti in that area and settings.

Thank you so much hellotanjent for your posts today! First solid benchmarks I've found! Really excited to see more over the next week from everyone!

So it is technically the faster gaming card in the world but not massively faster than a GTX1080TI it seems though.

[Xlucine]

Here are some synthetic benchmarks with the Titan V:

https://videocardz.com/74382/overclo...chmarks-emerge

So this is an overclocked card against an overclocked GTX1080TI which is not a fully enabled GP102.

Around 15% extra performance,and it seems the card runs at a much lower base and boost clockspeed than the consumer cards too.

Either this bodes terribly for ampere, or it's a sign of nvidia being a dick. The second one isn't unlikely...

Would nvidia bother releasing this if ampere was about to offer massive improvements over it? Nvidia pushing out faster cards right after charging an arm and a leg for a titan is common, but for 3 grand? It does beg the question of why this wasn't released sooner (as the card it's based off of is hardly new) - maybe nivida had a shedload of dies with defects in the memory controller that they needed to get rid of?

[CAT-THE-FIFTH]

It seems AT got the FP32 figure wrong and it should be 14.9 TFLOPs for the Titan V.

Either this bodes terribly for ampere, or it's a sign of nvidia being a dick. The second one isn't unlikely...

Would nvidia bother releasing this if ampere was about to offer massive improvements over it? Nvidia pushing out faster cards right after charging an arm and a leg for a titan is common, but for 3 grand? It does beg the question of why this wasn't released sooner (as the card it's based off of is hardly new) - maybe nivida had a shedload of dies with defects in the memory controller that they needed to get rid of?

This is because it has a huge amount of the chip dedicated towards FP16 and FP64 compute,and is still drawing power,which also means more heat produced. Hence the card is probably not boosting as high as consumer Pascal ones when it comes to gaming loads. For non-gaming stuff this card might make much more sense especially in AI research for example.

Even if newer drivers boost performance,I think a more consumer orientated gaming card,will either match or beat this card with a smaller chip and probably better performance/watt.

The thing is HBM2 does integrate most of the memory controller onto the actual HBM chip itself AFAIK(could be wrong) so it could be a combination of die defects and cost?? It might even be down to trying to keep within the TDP for more consumer related workloads,as it should save some power consumption reducing the number of HBM2 chips used,as well as the cost. The clockspeed of the HBM2 used also seems slightly lower too.

[CAT-THE-FIFTH]

In titles which can use Async compute,the performance uplift is much higher:

https://www.gamersnexus.net/guides/3...r-volta/page-2

It seems DX11 titles show usually upto a 20% improvement,but anything which can use DX12/Vulkan properly,its more like 20% to 40% it seems.

We’re entering territory of informed speculation. Please be aware that, from this point forward, we’re using our data to fuel conjecture on possible outcomes for Volta.

Purely observationally, based on the data we have presently collected, it would appear that the Titan V has two primary behaviors: (1) Applications which are built atop low-level APIs and asynchronous computational pipelines appear to process more efficiently on the Titan V; (2) the Titan V appears to host more cores than some of these applications (namely D3D11 titles) can meaningfully use, and that is demonstrated fully upon overclocking.

Given that overclocks in D3D11 applications produce performance uplift of ~20% (in some instances), it would appear that the high core count becomes more of a burden than a benefit. The GPU needs the faster clocks, and can’t access or leverage its high core count in a meaningful way. The result is that the Titan V begins to tie with the Titan Xp, and that the 1080 Ti closes-in on the Titan V. In lower-level API games, however, the Titan V pulls away by large margins – 27% to 40%, in some cases. The gains are big enough that we retested numerous times on numerous cards, but they remained. Our present analysis is that these applications are better able to spin-off multiple, simultaneous, in-flight render jobs across the high core count, whereas the tested Dx11 titles may function more synchronously.

As for the Titan V specifically, it can certainly be used for games -- but only in the context of, "I bought this thing for work, and sometimes I play games." If you're just gaming, clearly, this isn't the right purchase. Even for those users who have non-scientific uses for their scientific cards, the Titan V does appear to have some frame pacing problems that need to be worked out. We are not yet informed enough on the Volta architecture to root-cause these behaviors, and would suggest that it's either drivers or related specifically to the Titan V.

That’s what we think right now, anyway, and that may change. This is still early in Volta.