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

Very true and AMD APU Zen s are supposed to have it during first half 2017

Sent from my SM-G900P using Tapatalk

 

Noice. That gives me some time to save up for a 21:9 screen and a new GPU

 

Here are some new VRAM possibilities to throw in the mix for Volta and beyond - GDDR6 (2018) and HBM3 (2019-2020)

http://videocardz.com/63391/hot-chips-26-first-details-about-hbm3-and-gddr6
https://www.computerbase.de/2016-08/samsung-gddr6-14-16-gbps-2018/
https://www.computerbase.de/2016-08/sk-hynix-samsung-hbm2-hbm3/

 

Hmm, do we really think we'll see DDR5 in 2018? Seems ambitious with how slow the the 3-4 transition is (Not even mentioning how long 3 was on the market before that)

 

They could do like they did with gddr3 or 4 and pretty much have it for a year or two then jump to the next standard that will hold the market for awhile. Gddr now is competing for bandwidth dominance against hbm. As such, you could see quick advancements and a bifurcation of high end and low end markets for type of ram. Gddr is cheaper to produce and on a more perfected platform for production than hbm, so they want to get the race to higher bandwidth to maintain as choice over low yield, New hbm.

Sent from my SM-G900P using Tapatalk

 

I think you mean for GPU specs, I think (and could be wrong) that they meant DDR5 RAM memory standard, not Video cards. I just don't see them moving that fast.

 

Hbm is coming to cpus soon. I used that as an example of why they could move quicker. I realize they referenced both ddr5 coming quickly, as well as gddr5x being supplanted by gddr6 quickly and hbm3 coming in 3 years, approximately. But, the industry is also looking at integrating optical interposers into motherboards starting in 2018, so a jump in memory tech on this timeline makes sense. Just wanted a historical event to show the rare occurrence of them not milking a technology greatly before jumping to the next iteration.

Sent from my SM-G900P using Tapatalk

 

I wasn't aware they were pushing for that for a RAM spec. Thanks!

 

http://wccftech.com/nvidia-roadmap-2017-volta-gpu/

"Word around the grapevine is, the Volta GPU is going to be landing one year early in May next year at the GTC event held annually by Nvidia."

NVIDIA just loves to give us enthusiasts the middle-finger when it comes to upgradeability. I feel sorry for all of you who purchased a Pascal laptop, particularly those whom spent north of $4000. They probably won't be upgradeable as the Maxwell 980M is basically "end of life" now.

Volta will feature HBM2, so that is why I believe it won't be upgradeable. It completely changes the memory layout on the PCB and may also require a faster connection than MXM can provide.

This time, I hate being right.

 

I'm sitting on a 2.5 year old Sager with a 860m and while idling my temps are 90C+ CPU and 85C+ GPU. I have virtually no knowledge about repasting and stuff like that. And the battery has conked. Tried everything, but no luck.

I live in India, so selling stuff is not going to net me much. Worth upgrading to Pascal or wait for Volta?

Sent from my SM-G935F using Tapatalk

 

Sounds like you need to read a few tutorials or watch some videos on how to repaste and maintain your laptop, so you can extend its life and wait for Volta. But if not, then you may as well grab a Pascal laptop.

 

These Pascal laptops are mighty impressive with their GPU power (when compared to Maxwell too), I don't think there's much point in waiting for a supposedly early launch of Volta.

 

There's nothing early about Volta's launch... 2017 has been the target after everything got delayed by the 20nm fiasco. Everyone seems to forget that... 2017 isn't an early unveiling when Nvidia has two contracted super computers using Volta that are due before the end of the year. But yes, Pascal will be short lived.

 

I've always known I'd be skipping Pascal.

This GTX 980M is still too strong for me to upgrade, in good conscience. It hasn't even been forced below a mix of high and ultra settings at 1080p. I may ride it until 2018.

 

Wise decision! I'd also recommend those with any Maxwell GPU's to wait (960M or greater).

Worst case scenario, you save some money and Volta doesn't come as early. Best case scenario, you purchase the better technology and reap the benefits.

 

J.Dre, didn't you show an NVidia Roadmap a while back that showed Performance on y-axis vs launch date of architecture (Kepler/Maxwell/Pascal/Volta) on the x-axis, and it looked like Volta wasn't a large performance increase vs Pascal - whereas Pascal showed a bigger jump over Maxwell - this is what makes me think that it's not worth waiting for Volta when you can get Pascal now.
EDIT: Yeah, it's in your first post in this thread (post #1 of this thread) - although it looks like it's just memory bandwidth on the y-axis, not general performance. Have you got any roadmaps showing general performance of Volta architecture vs the other architectures - another roadmap like the one you showed but with General Performance on the y-axis?

 

This article is the closest thing to performance I could find on my cell in a 3-5 minute search. http://wccftech.com/nvidia-volta-gp...erformance-2017-summit-sierra-supercomputers/ It's a very old article!

It will depend largely on the die size and process. Some rumors suggest the memory arrangement allows them to resize the die and put as much as 6000 CUDA cores on the chip. But at this point in time, we don't know performance. Each process change will increase performance significantly.

Very early articles suggest an 80% performance increase with NVLink and Multi-GPU performance over Pascal, and a 300% increase in performance over PCIe gen 3 x16. However, if Volta remains 16nm, we probably won't see more than a 50% increase over Pascal. But upgradeability should remain through 2019.

 

After reading that article I couldn't really find anything in it that would say how much more performance we're likely to see out of Volta in comparison to Maxwell - the article was all based on supercomputers and improved efficiencies of many many GPUs working together, there didn't seem to be anything there to reliably extrapolate what single consumer card performance would be. The fact that it's on the same 16nm node, then I don't think we'll see a massive gain over Pascal, not like how we saw Maxwell vs Kepler. Maxwell managed a massive increase in performance over Kepler on the same node, so I think they've blown their load for a while on that front, I just can't see Volta achieving those type of gains just from architecture alone. For those reasons I think it makes sense to get a Pascal based notebook now rather than hanging on for Volta. (Haha, I'm not in the market for a notebook at the moment, but if I was I would buy a Pascal now).

 

Seems there are indications Volta may come out already around mid-2017 (based on IBM's plans for their Power9 servers which will use NVLink 2.0, which needs Volta):





http://images.nvidia.com/content/gtc-kr/part_5_ibm.pdf

 

It should be fairly significant, enough to justify waiting, especially for the sake of upgradeability. From what I've seen, in regards to TFLops, it's 5x faster than Maxwell.

My point in suggesting waiting was never for performance, but rather longevity and upgradeabilty. Pascal is a one-year product. It's like spending thousands on a prepaid cell phone.

My guess is on the previous pages (page 3) of how performance should be. I don't want to repeat myself a dozen times, haha.

 

Abandon all hope and go to monastery...

Or do what J.Dre suggested.

 

Those of you who chose, "I don't care to answer or speculate" realize you answered, right?

 

@J.Dre, are you planning to get a notebook with a Pascal GPU, or are you waiting until Volta?

 

I'm waiting to see what AMD does with Vega. If they launch Vega with HBM2, then I know NVIDIA has HBM2 (and probably a new socket type) planned for 2017.

Maybe it's Pascal w/HBM2, maybe it's Volta. But I know it's coming, and fast.

 

http://www.thecountrycaller.com/81149-nvidia-corporations-nvda-volta-coming-in-2017-with-nvlink-20/

Not much new information but there is still pertinent information. I'm excited for this - real change for once.

 

An MSI rep told someone who called in, that the trade in program will have 2 years of eligibility.

If true, Volta here I come, assuming it's out by the 31st of October, 2018.

 

If true I may skip Pascal and just stay with my old Sager, but swap in the 980M. It's not like I mind playing the game I like on medium settings.

 

if thats the case @2 years, u can be sure that msi will introduce their new machines on november 1st 2018

Sent from my Huawei Mate 8 NXT-AL10

 

I bet you can get 980M's for a very good deal now.

 

Already got one!

I have 4 laptops that are game capable. Three Alienware machines and my Sager. Going to be doing some swapping around before I sell.

Sent from my overpriced Galaxy S6 Edge +

 

The nvlink on both the mb and each card is what I'm waiting for. They have the high bandwidth connectors, but having nvlink to connect sli and directly to the cpu will be quite a bit faster... I also hope amd is in a better financial position for the 2018 debt obligation to be able to compete... More games will be optimized for it on the basis of getting the die shrinks, neo and Scorpio (whatever xbox is calling its next console)... But time will tell...

Sent from my SM-G900P using Tapatalk

 

Micro$oft is supposed to release a new Xbox next year, too. Apparently it will feature the new Vega GPU's from AMD (by rumor). But yeah, I'd save up and wait for an announcement of some kind.

Volta will definitely be around longer than Pascal. Pascal is just a placeholder because HBM wasn't ready and the market couldn't last three full years without new GPU's. Remember, the 980M was released in 2014. A lot of people I speak with didn't know that.

It will be interesting to see if they jump to 10nm (ahead of AMD) or stick with 16nm and introduce Volta, then refresh it with 10nm or 7nm.

My guess is they'll stick with 16nm for 2017 and skip a second generation Pascal, going straight to Volta.

 

Well, it's "I don't care to answer or speculate", so I want to answer but not speculate.
Seems logical to me

 

You still answered. Volta, anyone?

 

I'll have whatever comes in 2H 2018. Hopefully we hit 10 nm on both CPUs and GPUs by then. And if I have to hack Windows 8 into working with the CPUs then, I'll do it.

I abhor the new OSes (Windows 10, any OS X including and after Yosemite). I think it's high time I moved to Linux, using Windows only as a gaming OS. For now, though, it's a little too much of a hassle, and Windows 8.1 will have to suffice.

Any recommendations for a Linux distro, guys? Ubuntu is nice but it seems a little bloated.

 

At least for now it seems to be Polaris, larger than RX 480 or higher clocked. Most likely larger and down-clocked.

I think that Volta's release is tied with Vega's release and so far we don't know when Vega would happen. So 1Q Vega = 2Q Volta, or 2Q Vega = 3Q Volta (basically ~3 months between the two) and that's my answer to the poll as well.

If the process is good and the fabs are not hogged by phone chips (yeah, dream on ).

Or Pascal is third gen Maxwell, which is essentially it, so no need to skip anything.

 

Pascal is not third-generation Maxwell. I want some of what you're smoking.

 

well, in essence pascal is just a shrink of maxwell. aside from color compression and support for gddrx vram all performance gains came from higher clocks due to smaller process node. thats it.

Sent from my Huawei Mate 8 NXT-AL10

 

^^ What he said.

To fix your predictions? Naah It's better than yours, that's for sure, I'm still waiting about the M chips

 

If you call a new architecture "a shrink of previous generation," the you may as well say Volta is just "fourth-generation Maxwell." It's wrong either way you put it. Only in CPU technology does this fit more accurately.

@triturbo,
You act as though the year is already over! Must be some good stuff you got there. September and October is when they'll apparently be releasing more cards. Plenty of time.

 

Yeah, pulling a leg, but I still think that whatever they release would be 1050(Ti), 1040 and etc. Time will tell indeed. And the stuff's so good that sometimes I can't find my crystal ball

 

Yeah, but nVidia is likely to screw with the lower end.

I'd be really happy to be wrong on this, but it's fairly likely that the lower-end chips will be re-branded Maxwell (GM206 and GM107). Meaning the 1050 Ti might be a seriously crappy deal, i.e. less powerful than a 970 or even 980M.

 

Oh so NOW you believe me.

Anyway there will undoubtedly be Maxwell chips in the line just because Nvidia has so many of them to use and they still have headroom on every chip they released to make a refresh but I'm not totally sold on the 1050 being Maxwell. It needs to fit the TDP of those 960M machines with a decent performance boost which would be really difficult to do.

 

Huh? I've been saying this for a long while - low-end to mid-range Pascal would possibly be re-branded Maxwell.

Why not, though? The GM107 and GM206 chips are very efficient; a couple of tweaks here and there, a binned, over-volted chip could get the job done. Or else they'd probably just stick a GM206, what was the GTX 965M in place of the GM107 (GTX 850-960M) and call it the GTX 1050 Ti. The 1040 would possibly be the 860M, chip-for-chip and clock-for-clock and the 1050 a 960M, likewise.

And then nVidia would win, because the performance gap from the 1050 Ti to the 1060 would be on the order of the latter being 2x as powerful.

Mid-range GPUs for notebooks tend to get a crappy deal - not powerful enough for high, too powerful for medium and low settings.

 

When I suggested that the 1050 Ti would be Maxwell you were quick to point out how much of a performance disparity there would be and how unlikely it would be that Nvidia would do that.

If Nvidia pushes the 960M replacement as rebranded Maxwell, it opens a door for AMD to try to recapture that vital 1k price point which is why I don't see Nvidia being that stupid. Anything under that though will most likely be Maxwell simply because Nvidia needs to start putting the fabs on Volta ASAP.

 

The Aorus X3 page already shows a re-branded 1060M with a few more cores, higher clocks, and 6GB GDDR5. Definitely something incoming... What exactly is the question.

Apparently they'll be presenting at Pax this week, along with Alienware.

 

I only found this, how do you know it's a rebrand? They state Pascal in the description.

As for the Ti/nonTi - Ti for Pascal, nonTi for a Maxwell rebrand (something like the already discussed 1050 = 960m~965m), unless they plan to skip the mid/lower end entirely and fill it with rebrands. Makes sense?

 

Check XoticPC's site for the Aorus. Must be a typo somewhere.

 

Will Volta for laptops launch in Q3 2017? Or will it take longer? And will Volta be a cooler architecture than Pascal, which seems to be hot hot hot? And what % performance increase will we see?

Sent from my SM-G935F using Tapatalk