Volta: NVIDIA's Next Generation GPU Architecture (2017-2018)

I'd actually like to see Samsung's Quantum Dot and Supreme Motion Rate technology in gaming laptops one day.

True, though. There will be many new and exciting technologies to come in the next few years. Even the first six-core mobile CPU.

 

Yeah. Maybe new battery technologies that will give them longer lifespans and longer time between charges.

And I know this one is a bit far-out, but maybe e-ink key caps for contextual shortcuts?

Sent from my SM-G935F using Tapatalk

 

Pretty much what I've said three months ago, so it's not unexpected on my side, or the way it turned-out.

The Quantum Dot is not Samsung's firstly and secondly the first (yeah, a word play) QD TV was Sony (XBR X900). The SMR is not entirely different than any other fake Hz rating. Interested to see QD displays, not so much about the later, would rather get 120-REAL-Hz.

I dream for such keybord for the longest time (4 years, maybe more). I think there was an e-ink keyboard project, but not sure what happened. FAR better solution than the power-hungry monster the Lebedev's OLED keyboard was. It had its plus sides - it was made and you can make colorful buttons.

 

OLED keyboards are fine. There's much less energy lost through that than the massive (LCD) screen.

With that said, the reason we have physical keyboards is the keys provide tactile feedback for quick operation without relying on reacting to visual clues. In other words, you're not supposed to look at your keyboard. If you want to operate with visual feedback, touch screens are much more flexible.

 

Google Lebedev's OLED keyboard and you'll find two things - 1. It comes with I think 120W brick (or was it 90, still quite the toll for "just a keyboard") and 2. It IS a keyboard with 104 (I think that's the correct number) separate OLED displays. And to drive the point home, I was talking about the e-ink equivalent. You don't need to tell me about tactile feedback, I hate touch-screen even though my both phones are (maybe because they are). I'm not saying that there are no advantages to touch screen though.

 

Thanks. Didn't know that.

What is going on inside the Lebedev? OLEDs, especially when showing bright logos on mostly dark/black backgrounds (which is exactly what a screen-on-keyboard is expected to do) should have pretty good energy efficiency. Is most of the power really going through the OLED?

 

It's the OLEDs, they might be efficient displaying white letters on black background, but in any other case they became quite the consumer, especially in the inverse case (black letters on white background). Other thing that "helps" for the power consumption is that they refresh constantly, because this is how they work. This is where e-ink is vastly superior - once the desired pattern is displayed they stay so without any extra power. Another thing is that with OLED I'm pretty certain that there would be burn-in with something as static as what the keyboard is. Anyway. let's not derail further

 

I think it's fair to say that 2018 will be the year of real GPU upgrades - a process shrink (possibly to 10/7 nm), not to mention stacked HBM2 memory. Improvements need time, my 860M in 2018 will firmly be in the low-end, and a hypothetical GV206 chip may be as much as ten times as powerful as Maxwell 1.

Patience is key. This is not my year.

 

And as an added bonus to the GPU upgrades, 6-core laptop processors!

Sent from my SM-G935F using Tapatalk

 

http://wccftech.com/nvidia-volta-gv100-gpu-fast-pascal-gp100/

They say that Volta (GV100) is twice as fast as Pascal (GP100). Does that mean that we will get at least a 50% performance increase from the GTX 1080 to the GTX 2080?

 

Twice as fast, as in their own magical vacuum of perfect optimization and such.

So between 50-100%, but in reality 25-30%.

I will GLADLY eat my words if they sh** on Moore's Law and decide to jump to the moon instead.

EDIT: Seems like nVIDIA is embracing Intel's old tick-tock cycle atm.. my god

 

I think it's easier for NVidia to keep up with Moores Law in terms of providing double the performance for their GPUs because smaller transistor size just means more cores per mm squared each generation & when combined with any increases in efficiency of the base architecture (clock for clock efficiency improvments) then they can make great strides with each node shrink. Whereas with Intel & their CPUs they're only sticking to the same number of cores per CPU - the usual 2 or 4 or 6 or 8, so they can't rely on simply doubling the number of cores/transistors with each node shrink - I think it's easier for Moores Law to apply to GPUs - I get the impression it's easier to throw a load more cores at a GPU and efficiencies will increase, but I think CPU design is more restricted, they can't keep chucking cores at CPUs.

 

I agree, Moore's law makes more sense for GPU's these days.

 

Knowing NVIDIA, they won't be doubling performance every generation, even if they could. I'm 90% certain that Pascal has a one-year shelf life and Volta will be around for quite a while - re-branded multiple times with improvements each iteration. The first of Volta should be 16nm with HBM, which is the same process as Pascal. And 2018 should finally bring us an entirely new architecture (i.e. 10nm).

If they move directly to 10nm this year with Volta, I will be surprised. But hey, that would be amazing.

 

theres always that option of combining Volta arch with 16nm process...

 

https://www.mobipicker.com/summit-s...cal-p100-9-5-tflops-fp64-compute-performance/

Seeming more and more likely we will see Volta's official announcement in May. We're only six months away! I can't wait.

Will be interesting to see if they go with GDDR6 or HBM2/HBM3. I imagine the GeForce cards won't have HBM2/HBM3, at least not the low-tier cards. Perhaps the flagship or Titan will. There seems to be an issue with power / cooling.

 

only 6 months? thats HALF A YEAR...

 

@ J.Dre, HBM3? I know it's been announced, but we've hardly even seen HBM in GPUs, let alone HBM2, we can forget about HBM3 yet.

I agree, though, that Volta next year is possible. Pascal was merely a die shrink of Maxwell.

 

I only mentioned it because articles suggest there are TDP issues with HBM2, and they appear to be fixed with HBM3.

Just curious to see what will happen, that's all. GDDR6 seems far more likely for 11-series Volta.

 

I think they'll be another iteration of Pascal for 2017, bigger chips and more cores. I think Volta will come later with the node shrink. This is me just thinking about it, rather than having done any researching for leaks or the like, so don't shoot me if it's blatently wrong! ;-) I cast my vote in the poll for another iteration of Pascal.

 

Nvidia has already filled up the slots for power usage, how are they going to fit more onto a chip without increasing the power usage?

With the Titan (Pascal) and the unannounced 1080ti at 250w, 300w+ under OC load, where is the headroom for more performance?

I am starting to lean into the idea that Volta might release sooner, next year between April and September.

The only thing I think that could get in the way is allocation of chips for the GV100, with the Summit Supercomputer one the most visible customers, which is supposed to get Volta silicon in 2017, and go online in 2017-2018:

SUMMIT - Oak Ridge National Laboratory's next High Performance Supercomputer.
https://www.olcf.ornl.gov/summit/

NVIDIA Volta GV100 GPU Chip For Summit Supercomputer Twice as Fast as Pascal P100 – Speculated To Hit 9.5 TFLOPs FP64 Compute
http://wccftech.com/nvidia-volta-gv100-gpu-fast-pascal-gp100/

Currently I voted for 2018 for Volta, but if AMD releases new silicon to compete with Intel / Nvidia, the timetable could be moved up.

 

Honestly, that'd be a whole new arch. Maxwell = 28nm. Pascal = maxwell on 16nm (basically).

Them dropping Volta to 10nm is going to be a new arch if it launches on 16nm.

Edit: They even cheaped out on the cards for mobile using Pascal. Mobile 1070 is a shrunk mobile 980. Mobile 1060 is a shrunk 970M. Desktop 1060 is 960 OEM (970M base specs) too. Don't underestimate them lol.

 

Well I reckon they can do it, because the 1070 and 1080 cards sip power really, see the following two links - GTX 1080 only sipping 184W, GTX 1070 with 161W.
http://www.guru3d.com/articles_pages/nvidia_geforce_gtx_1080_review,8.html
http://www.guru3d.com/articles_pages/nvidia_geforce_gtx_1070_review,8.html
GTX 780 and GTX 780ti were a 'huge' 231W and 262W seen at the same links above! So, looks like they've got another 50W to spare for a sucessor to the GTX 1080, and sure some room to spare for a GTX 1070 sucessor as well. This is not even considering any kind of performance efficiencies they can increase as the process is maturing - I reckon they can do it! They've cheaped out on number of cores & die sizes with the Pascal cards, there's room to grow them for another iteration of Pascal before Volta comes along in 2018 on the next die shrink - this is my prediction!

 

Nvidia have already released the 1080 + 50w model, it's called the Titan (Pascal), drawing 250w

Hopefully in January we will also see the Nvidia 1080ti model, also drawing around 250w power.

That's why I suggested Nvidia has already done their 2nd push of Pascal GPU's, and now there is no further room to grow power-wise.

It's going to take an architectural new release or yet another die shrink, or both.

 

Considering Vega is anticipated to be launched Q1 2017 ( Read more...), I'd definitely agree with the above and also add that Volta is highly likely to be announced in May.

 

What?

Volta is not annouced in May for sure......

Nvidia dont even start shipping the high end volta until prob september.

Vega at best is around 1080 ti. Probably around 1070/1080 level. Priced around 450-500.

AMD dont have the money to throw around big dies. In fact, AMD historically did better with small dies.

 

@LunaP you wanted Volta discussion...

Sent from my SM-G900P using Tapatalk

 

Where is everyone getting the idea vega=volta?

 

Tanx! ze fact dat @D2 Ultima didn't link/tag me first hurts lol, he knows this is what I'm living towards!

 

You're just going to accost me in a dark corner of the internet the second Volta launches and demand I spill my guts (of information) anyway.

And I'm going to LIKE IT.

 

Lol! It's hard to keep on top of every thread (or every thread that has been closed)... We had talk of this in the Pascal thread also. So don't be too hard on him (I just found it myself)...

Sent from my SM-G900P using Tapatalk

 

Hell, the Titan Z was a 479W card (albeit dual GPU card), so they could still release bigger dies that consume more than 250W - also shown in the same charts that I linked in my last post. I'm still sticking with my prior prediction. (I'm not saying they're gonna release a 480W Pascal card, that's expensive and somewhat unpractical, but there's still a lot of wiggle room there between a 250W and 480W card). Perhaps just raising Watts by 50 W or so on each card combined with increased efficiencies gained from a more mature 16nm process, this could result in good chunk of performance increase that could satisfy a reiteration of Pascal for 2017, and then leaving Volta for 10nm in 2018.

 

I hope they don't do that. It'll just be an even hotter GPU. They can barely cool Pascal as is.

 

Lotsa fab news, shows change is accelerating just when we thought it was slowing down again

TSMC Ramps Up 10nm Chip Production For Apple’s A11, MediaTek And Huawei’s SoCs
http://wccftech.com/tsmc-apple-10nm-chip-production/

TSMC Roadmap Update: Will Be Ready To Take 7nm Orders By April 2017, Volume Production Scheduled To Begin By 2018
http://wccftech.com/tsmc-roadmap-7nm-orders-2017-2018/

TSMC Betting Big With New 80 Hectare Facility For 5nm And 3nm In Taiwan
http://wccftech.com/tsmc-iphone-8-5nm-3nm-10nm/

But, this new news isn't good, there could be delay's for 10nm production, or it's just the normal early production glitches that will quickly be worked out.

10nm was supposed to be shipping immediately in 2017...

TSMC And Samsung Suffer Poor 10nm Yield Claims Taiwanese Media
http://wccftech.com/tsmc-samsung-10nm-yield/

"With 2016 coming to an end, one of the biggest launches next year will be 10nm. Manufacturers and consumers alike are waiting for the big update in hardware. Samsung and Qualcomm will take the lead, as they take the stage to announce the Snapdragon 835 and the tentatively dubbed Exynos 8895. To that end, TSMC and Samsung have already announced big plans for the process. It was TSMC who took the lead in announcing that sufficient capabilities for 10nm production had been achieved. Today, we’ve got some more news on the matter. Take a look below for more..."

 

Isn't that mainly due to them just putting the desktop version in , and removing wut 7% of the hardware and being like "this is fein"...omg I gotta draw this now gdi... LOL

 

People say it is slowing down only because Intel won't be below 10nm until after 2020-22! They think because Intel raced to the smallest die size before they hit their wall means everyone hit the wall. Instead everyone played catch up while Intel was waiting for EUV. They all need it for below 10nm and 10nm was originally planned on euv. But EUV won't be too manufacturers until late 2017 into 2018. Meanwhile, because Intel is having issues at 10nm (everyone is having yield issues, no news there) and because they milk the market, glofo, Samsung, and tsmc are set to hit 7 likely before Intel does.

Sent from my SM-G900P using Tapatalk

 

http://satprnews.com/2016/11/27/zei...ration-of-euv-lithography-due-in-early-2020s/

“In 2018 the first chips made on current technology EUV scanners are expected to roll off the production lines of our customers. A lot of work is still being done to ensure the introduction of EUV in volume manufacturing, in tight conjunction with our highly successful immersion scanners which we continue to improve. Yet ASML and ZEISS are looking beyond this important milestone. We see a long and successful future for EUV lithography in advanced chip manufacturing and with this agreement we set the right conditions for development of the next generation of EUV by ASML and Carl Zeiss SMT, so that our customers can reap the rewards of their EUV investments up to the end of the next decade,” said Peter Wennink, President and Chief Executive Officer of ASML.

“High-NA is the logical next step for EUV, as it circumvents complex and expensive 0.33 NA EUV multiple patterning. High-NA EUV is a robust way for chips to scale all the way down to the sub-3 nanometer logic node in a single exposure with high productivity and reduced cost per feature. That is several generations from where we are today and underlines our commitment to propel Moore’s law,” said Martin van den Brink, President and Chief Technology Officer at ASML.

http://spectrum.ieee.org/semiconduc...makers-eye-euv-lithography-to-save-moores-law

"Now, though, the technology really does seem to be turning a corner. The brightness of the EUV light source made by Dutch lithography-tool manufacturer ASML Holding seems to be closing in on a figure long targeted for commercial production. ASML, which has emerged as the technology’s standard-bearer, is now shipping EUV scanners that it says should be ready to mass manufacture leading-edge microprocessors and memory starting in 2018. The world’s most advanced chipmakers are working hard to determine when and how these machines will be incorporated into their production lines." - discussing GloFo


http://www.eetimes.com/document.asp?doc_id=1330971

Taiwan Semiconductor Manufacturing Co. (TSMC) said that it plans to build its next fab for chips made at the 5-nm to 3-nm technology node as early as 2022 as it aims for industry leadership.

Sent from my SM-G900P using Tapatalk

 

I think some of the difficulties with cooling Pascal (desktop version keeps cool ok due to big coolers) is the fact that it's a small die outputting the 'normal' amount of heat - the area that the heat transfers over (surface area of the chip) is lower than Maxwell, increased heat density which slows down heat transfer to the heat sink. Also, with Pascal they have basically pushed the Mhz & voltage to the max, which is why there's not much overclocking headroom in these chips. They could still improve performance by increasing number of cores while keeping Mhz the same, because the area of heat transfer would just increase together with the increased die size - so creating bigger chips isn't gonna necessarily increase temperatures.

 

Given the same cooling setup, scaling up the chip with identical density would still increase the temperature, as the overall load is increased. But it may not be as bad as increasing thermal density, which would require both increased heat transfer efficiency at the surface and higher cooling capacity downstream to keep the same temperature.

 

Yeah, that's right, as long as the heat sink can remove that much heat from the cooling fins, then it shouldn't be any worse.

 

10nm working for mobile chips doesnt mean it works for high performance chips.

 

It's unlikely we'll see 10nm GeForce cards by NVIDIA next year. It's very likely to be 16nm/14nm.

http://www.electronicsweekly.com/news/business/samsung-tsmc-hit-poor-10nm-yields-2016-12/
http://www.digitimes.com/news/a20161222PD207.html
http://www.mobilenapps.com/articles...-disrupt-market-debuts-ipads-iphones-2017.htm
https://9to5mac.com/2016/12/23/tsmc-10nm-a10x-chips-2017-ipad-delay-claim/

TSMC, Qualcomm, Samsung, etc. all are having trouble meeting demand as it is, and none of the articles mention NVIDIA.

 

I don't think Pascal outputs a normal amount of heat. The GTX 560Ti generally drew as much or more power than the GTX 1080, and had a slightly larger die size (332mm squared, as opposed to 314mm squared). But the 1080 uses what was essentially a beefier GTX 580 cooler, with the same effective performance as the 560Ti's stock cooler.

In this essence, Pascal is hotter than Fermi, and Fermi was considered one of the hottest architectures we've seen, in addition to being power hungry (which means more TDP to dissipate). Also, the GTX 680 was 295mm squared and rated for higher wattage, but didn't need a cooler as beefy as what they shoved on the 1080 which barely works.

Make no bones about it, Pascal is extremely hot. The fact that it's an overclocked form of maxwell with more cores in 3/4 the die size and a constrained voltage variance band probably doesn't help. Forcing Maxwell's voltage to a constant generally increased power draw and heat outputted, from what I remember. It was, as you say, manage-able with larger coolers on desktop cards, as some Pascal cards are doing now, but Pascal is unnaturally hot. And if Volta ends up hotter... I have no idea what we're going to do.

 

I came up with a unique water cooling solution that may prove beneficial for Volta. Never got around to drawing up designs or filing for a patent, though. But it allows even the smallest of laptops (i.e. 13") to have water-cooled components.

But yeah, I agree. It will be very difficult to cool with standard fan setups they currently use for Pascal. Volta will create more heat. That's an undeniable fact. It may even force them to resort back to an "M" series (or something like it).

 

By heat I mean not the temperature that the chips typically run at, I'm talking about the Watts consumed. I don't dispute what you say there, it's a high heat density chip which makes it more difficult to cool. My desktop 1070 stays cool though, but it's a third party cooler, and for sure more difficult to cool in laptops.

 

Well yes, that's why I was using watts consumed to show that Pascal's problem isn't watts consumed in a smaller die than normal. It's normal watts consumed in a normal sized die... but it's exceedingly hot. Most likely from the density of the package itself, rather than pure die size/power consumed.

nVidia won't watercool. They simply won't. Their ego is too much. And laptop ODMs won't want to spend that kind of money. I don't know what Volta is going to be like, but the signs point to horrible right now.

 

This is also why, in the discussion of 3D ICs, heat dissipation is a hot topic, including the use of heat pipes through the stacks for cooling (or, at least I've seen that a year or two ago)...

Sent from my SM-G900P using Tapatalk

 

It would be for laptops but okay. Shouldn't cost too much.

 

They still sell 5400RPM drives and thermal throttling machines instead of focusing on cooling... I don't think any but the highest end will even consider it, and it'll probably charge a solid $200+ premium.

 

How much they charge and how much it costs is completely different, though. Or am I misunderstanding what you mean?