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

AMD should be able to have a clean sweep with their low to mid range mobile GPU's then

 

I agree with you guys. Seems like Nvidia is taking advantage of their market dominance to the extent of "milking" the architecture. Volta will be around for a while.

Bye bye, Pascal. Good riddens, in my opinion.

 

Precisely my thoughts, AMD has a lot of room to dominate the budget range, and it would surprise me if they didn't focus their effort there.

 

It's a shame they're hedging their bets on expanding pascal's lifetime. I don't expect to see any laptop volta chips until Q1 '18

 

That's likely the case, at least if you believe SK Hynix about them ramping up GDDR6 production for an early 2018 release.

 

GDDR6 is most certainly a process of "milking" an architecture's memory. Don't get me wrong, it's a decent improvement but it's far from the capability of HBM.

Volta will be around for a while.

 

I'd love to see a switch industry-wide to HBM, seems like in this case cost and availability was the determining factor though.

 

Ah, here's this thread, I'd wondered where it'd gone - found it! This might be interesting: https://www.notebookcheck.net/Nvidia-reveals-more-details-on-the-upcoming-Volta-GPUs.244332.0.html

Ha, they're wrong about the 32 square inch, what are these notebookcheck.net guys on nowadays! ;-)

EDIT: Looks like Volta might give us in the region of 1.5x the performance of Pascal then - fp64/fp32 is showing as 1.5x Pascal in that table in that notebookcheck.net article above.

 

J.Dre started this thread, and he's not been around recently, praps that's why this thread hasn't been in discussion - J.Dre's not been around since June I think - I checked his profile just now to read his latest posts & sounds like his life was going off the rails, hopefully he's ok or getting back on track.

 

Very interesting link. My only concern is about how hot Volta will be. If it ends up running substantially hotter than Pascal, I think a lot of the current laptops will have to be majorly redesigned. I'll tag @D2 Ultima since he is very knowledgeable about these things.

Sent from my SM-G935F using Tapatalk

 

32 inches is about 812.8mm, and the die is something like 814mm squared, so 32 square inches is roughly correct. I blame lack of food and sleep for this atrocity I typed, clearly my brain was not mathing

That is GV100 compared to GP100, FP64 doesn't really exist in any capacity on GP102 and downward, so I don't know how GV102 and below is going to be.

Why? It's not much faster than GDDR5X on a 384-bit mem bus and GDDR6 is supposedly going to destroy it. If they'd kept HBM2 on a 4096-bit mem bus though, it might have been worth some salt (or the cost).

It depends. Remember that GP100 is significantly less hot than GP102 and downward due to a much larger die size despite having the same core count; thermal density is down. The problem is not the dispelled heat but the concentration. It will remain to be seen how consumer volta happens, but we'll find out come GDC 2018.

If chips get hotter than they are now, we're going to be in some serious trouble. Sloppy QC and sloppy design of cooling will be even more detrimental.

 

Don't complain too much on Clevo's HS design, bruh. You find much worse out there. All types of Tripod or even Twopod Heatsink design as this is the new Can it be worse? oh yeah

 

I was thinking of actually everyone when I thought that. Clevo, Alienware, HP, Acer, ASUS, Crapzer, etc.

Then you have trash like the thinkpad with the single pipe heatsink for a 7820HK... or Crapzer's entire lineup when it comes to garbage design.

 

This maths hurts, yo...

One inch is exactly 25.4 mm long. One square inch is an area of one inch by one inch, i.e. 1 in² = (25.4 mm)² = 645.16 mm².

Hence, 814 mm² = (814 ÷ 645.16) in² ≈ 1.2617 in². 32 square inches is about as big as a CD. That would be a huge piece of silicon.

 

Yeah apparently my brain is poop or something. I didn't even think about how 32 inches would look.

 

Nah, it happens from time to time to everyone. I done goofed, the mods done goofed, Trump has done goofed too much, and some people haven't done goofed enough.

 

One thing I'm thinking after reading the available information on Volta is that to fit so many more shader cores on to a chip that is using what is basically a refined 16nm (yeah, they say 12nm, but the stuff I've read says it is just a highly refined 16nm) means that the dies will take up a lot more space, which means it might be the same or lower thermal density as Pascal. However this is on a physically larger chip, and that means more area to put LM pastes on and cool down easier. Please correct me if I'm wrong.

Sent from my SM-G935F using Tapatalk

 

If the total power draw is the same, then it will be cooler on a larger die.

It's the thermal output to core to size ratio that should matter.

Take the 980 versus the 1080 for example. The 980 had fewer cores at a lower frequency on a larger die space and was rated for less power. So it was very cool. Maxwell lost power efficiency fast when OC'd, but the 980s still ran pretty cool as far as previous architectures were concerned.

The 1080 increased core count, frequency, power draw. But die size was 314mm squared versus 398mm squared (from what my short google search just turned up). That's a pretty thermally dense thing right there. If GV104 is something like 3500 cores in a 512mm squared die and still draws 180-200W, then it should be cooler. More cores/frequency sure, but a larger space it's spread out over is the key. Of course it probably won't be that big; GV102 is probably going to top out just over 600mm squared rather than being over 800mm squared, so who knows. It might be slightly cooler if the TDP used is less and the overall die size is larger.

Or, I could be talking crap due to lack of sleep, see my math above for reference. #Disclaimer

 

I certainly hope that it's cooler. Then we will have no need for the Max-Q crap that they pulled recently. But then again, this is nVidia, they will probably pull Max-Q crap even if it is cooler so that the 2080 Max-Q fits in something tiny like the Razer Blade and costs $3999. #Speculation @Papusan @hmscott

Sent from my SM-G935F using Tapatalk

 

Mini-Q could have been done a lot better, for sure. But that would have apparently required basic Pascal to have been done better too, and that apparently is too big a problem to fix.

Who knows what Volta will do

Sent from my OnePlus 1 using a coconut

 

Max-Q aka renamed M graphics will continue. But with higher prices than when we had 980M vs. the new 980 desktop version. Nvidia know how to milk!!

 

Regular pascal in good chassis is a really good option, why do you say is a problem to fix?

Most of the skus stay well above the base MHz. The only thing you could say is that it can't sustain the same turbo as desktops. Is that the big problem?

Enviado do meu MHA-L29 através de Tapatalk

 

Yes, that's exactly the problem. They run slowly. The only reason a 1070N even compares to a 1070 is because it has more cores so when it runs slower it sort of matches. 1060Ns are wildly underpowered down to them even failing VR qualifications (putting them under a 970) in many chassis.

The fix is either raising allowed TDP, or adjusting the voltage frequency curve in the vBIOS. Many people have the same 115W limit but are able to get significantly higher speeds out of their notebooks in varying performance scenarios due to adjusting the voltage frequency curves on their cards. These cards don't even need 1v to hit the upper 1800MHz range, far less the 1.05-1.063v they usually sit at in those ranges. It's wasting power, clean and simple.

If they had used such a stringent curve, Mini-Q chips would have been able to get existing 1070N power in a 100W limit using a 1070MQ, not a 1080MQ, while giving overclocking potential and actual high speeds to people who have normal 1070Ns, and 1080MQ would've been able to suffice with a solid 115-125W limit and get something resembling a 1080N's performance, while actual 1080N users would hold pretty high boost clocks all the time unless the chips got seriously hot.

Pascal is unoptimized out the gate, and they could have given us a lot more power with it if they'd configured the vBIOSes properly. If they do it properly with Volta, we might have some things going for us. But who knows? I don't have the ability to suggest something like that, far less test it and give feedback.

 

Don't have too big hopes. As long AMD can't compete, forget it!!

 

Yeah, I think on the whole, the Pascal architecture is pretty darn good, some good performance improvements over Maxwell, and the notebook variants aren't a million miles away from the desktop variants in terms of stock performance.

One thing I think you can do if your GPU is operating at too high a voltage for any given frequency - you just go ahead and overclock it to the edge of stability, that way you're using the most optimised frequency for all of the voltage points. Then I'm guessing you can use the Power Slider in Afterburner to control the max power you want to see so it doesn't overheat - you've got your most efficient Pascal GPU right there at this point I reckon.

 

No, I think it's pretty poorly designed considering past models to be honest. It's doing what GCN did, running hot and drawing lots of power at a huge inefficiency curve, while also having very varied performance not only based on power draw but on heat. Since day 1 of its launch I said it, it's very underwhelming for a new architecture. But it's still stronger than what we had before so not too many complaints there I suppose.

As for the overclock, I disagree. It won't do much, and just overclocking it won't really do anything worthwhile if you're already TDP limited. I've run some 4K titles that have INSTANTLY, at sub-60c, put me under 1800MHz consistently. Just power limits alone. The best possible thing is to adjust the frequency curve. People with overheating notebooks like ASUS' terrible GL502VS and GL702VS have found that instead of the mid-1500MHz range and overheating, they lock their maximum frequency to about 1750MHz and reduce the voltage and they get constant speeds (and thus performance) AND reduced temperatures.

There's people on this forum who've managed to literally pull 2000MHz under 1v, where mid 1800MHz can use 1.05v constantly in normal circumstances. I've found my cards absolutely baffling. I have seen my primary drop to 1810MHz and sit at 0.9v, but at 1860MHz it's using 1.05v? Like, are you absolutely off your rocker? These things are inefficient as a helicopter rotor on a jetski!

If they had adjusted the curves and tried their best using the worst cards as a baseline and extrapolated the curves for overclocking, I'm pretty sure 2000MHz would be easy as hell for most everyone using a Pascal notebook with something even resembling cooling. Even @Mr. Fox and @bloodhawk who have unlocked vBIOSes say that at most it's about a 5c increase when the card is taking whatever the heck it needs as far as gaming, and I don't believe they're undervolting it either.

Also, as for this MSI Afterburner power limit...



Where is it?

 

Notebook cards don't have access to the power limit slider then? Well, if that's the case there's not as much point overclocking if you're already temperature limited, but even if you're Power limited overclocking should give you more bang for your Watt either way. But, theoretically if you overclock your GPU you're allowing it to run at an optimised frequency for all of the voltage points - and then IF you did have access to the Power Slider you'd just reduce it to fit within your temperature constraints - that way it wouldn't overheat and the frequencies would be optimised for each of the voltage points. I'll give you that Pascal is kind of pushed to it's limits in terms of operating at an inefficient place on the Voltage/Frequency curve, but really that only applies to aftermarket desktop cards that run in the 2000Mhz region - there's hardly any space for overclocking those & adding voltage often doesn't help, so they're already pushed to their limits & thus NOT operating at an efficient place on the Voltage/Frequency Curve, but I still think Pascal is a good architecture for it's improvements over Maxwell. Notebook GPUs are operating on a much more efficient place on the Voltage Frequency curve than their desktop counterparts, particularly in comparison to the aftermarket desktop cards. I don't really agree with your viewpoint that Pascal is inefficient for the performance we see, especially on notebook cards.

 

So, @D2 Ultima— my TL;DR from your comment: Pascal has theoretically good hardware, but lousy notebook vBIOSes keep them from performing as efficiently as possible?

Isn't this what Max-Q is meant to address?

 

I'm too tired to address every point here. But I'll say desktop cards are inefficient too. If you undervolted the reference cards you'd get much better frequencies AND lower temperatures too, even on desktop cards. All of them. Unless you're pushing for 2100MHz+ where you'd actually need the voltage that is.

As for the notebook cards, the only real benefit is that they're similar to desktop cards in performance and design. That's it. Overclocking the cards on units like 1060s and 1070s that get power limited from the get-go doesn't really do anything, it just throttles about as fast to where it normally is, because voltage is still high and it's downclocking because it's at a power limit. You're misunderstanding Pascal here. If your card throttles because it needs more than X watts, overclocking it isn't going to make it faster when it's limited to X watts. You need it to make better use of X watts, which is attained by either cooling it better (cooler = less power drawn) and/or undervolting it. That way it makes the most efficient usage of its power. The desktop cards just often have much higher power limits, with the XOC vBIOS being official and flashable in windows providing no power draw limits if one really wants their cards to run full bore.

Yes and no. Mini-Q attempts to use the default voltage curve with much reduced clockspeeds and a reduced power limit to squish the cards down. Think a GTX 680 4GB (1.2v, 1006 base 1110 boost clocks, 195W TDP) to 780M (1v, 721 base 823 boost clocks, ~106W TDP)

It's just still got the same memory, it's just that with existing cooling solutions able to handle the cards, such reductions in heat output can be handled in supethin units if they try. According to Prema, the guidelines are in place so that OEMs actually DID something, or it would've been far worse.

As for full explanations on my statements about the voltage frequency curve and what each card can do at what power gradients, it's all in that article I did.

 

Oh D2, D2...I've just done some testing to support my prior theory that overclocking Pascal when power limited will increase efficiencies & performance! Here it is:

Methodology

Test A:
1) limit my desktop GTX 1070 to 50% TDP
2) Operate core at Founders Edition simulated base clock of 1506Mhz.
3) Test Firestrike Extreme Graphics Test 1 on loop for 20 mins & observe average core frequency over run.

Test B:
1) limit my desktop GTX 1070 to the same 50% TDP (as with Test A above)
2) Operate core at max stable overclock without moving voltage slider in Afterburner: 1682Mhz base clock.
3) Text Firestrike Extreme Graphics Test 1 on loop for 20 mins & observe average core frequency over run.
4) Compare difference in average core frequency between Test A & Test B.

Results:

Test A (showing 1525Mhz average frequency):


Test B (showing 1641Mhz average frequency):


As you can see by comparing average core clock frequencies above in the screenshots, overclocking in a TDP limited scenario has increased actual core clock frequency by 7.6% (1641Mhz/1525Mhz*100). This is clear proof that overclocking in TDP limited scenarios increases performance. My initial suggestion to you was that to increase performance efficiency of Pascal you would lower Power Slider to a value that your system could cool to a temperature that you're happy with & then you would apply your max stable overclock, thereby effectively maximising & optimising frequency at each voltage point on the curve.

Additional Footnote: During Firestike Graphics Test 1 loop testing above, card was sitting at 50% TDP actual power consumption as shown by GPUz for the entire runs in both Test A & B, so strictly running at the 50% TDP maximum limit I had set - TDP limited the whole time).

One further footnote: you made me aware that notebooks don't have access to the Power Slider used in the tests above, but perhaps you can mod the vBIOS to a specific max power that is in line with the max temperatures you want to see.

 

As usual, NVIDIA fails at firmware. I am confident the trend will continue with Volta. I mean, c'mon, they've sucked at it for years, so there is reason to believe it will ever change. They are abject failures when it comes to firmware. Stock Pascal vBIOS blows like Maxwell did, like Kepler did, like Fermi did. @Prema and @svl7 where the only reasons mobile NVIDIA GPUs have been worth a damn for years, and now they've got the desktop GPU firmware screwed up about as bad as notebooks.

I don't undervolt my GPUs. I don't have any reason to, and the voltage technically isn't adjustable. That half-assed excuse for a voltage curve tool in MSI Afterburner is a piece of trash. I cannot really see that is does much of anything other than waste my time trying to get any kind of meaningful use out of it.

 

Well, however you do it, it could run cooler if you're not overclocking. I should honestly man up and learn how to do it myself. Cooler system is nothing to sneeze at I think as long as performance isn't sacrificed.

But you have 300W power limits so you aren't hungry for that.

 

Fair enough testing, though I must say that vBIOS modding requires a programmer to flash. If you're going to mod the vBIOS anyway you could just raise the power limits if you have a good notebook *shrug*

 

Well, you yourself just said in your post above to Fox that a "Cooler system is nothing to sneeze at I think as long as performance isn't sacrificed." - so my method is one way to achieve that, and more reliable & flexible than using the Ctrl-F curve in Afterburner. But, if you don't have any cooling issues then all you'd do is overclock to the max without bothering to reduce TDP.

 

Your method is impossible for notebook users due to a lack of power limit access. Besides, the point of the voltage tweaking is that you can both get more performance AND less heat by using it. This is the key. Your method is to just overclock as far as it'll go so it'll clock as high as it can at that voltage, and then lower TDP limit. I can't see how it'd be anything but counter-productive for a notebook.

If afterburner's voltage graph is bad, then doesn't Precision X have a similar tool for some? Though I've heard multiple people say they can't stand one or the other. I only really use Afterburner for performance monitoring to use with RTSS, so I would not know. I usually used Nvidia Inspector for all overclocking.

 

I hate to say it, but to coin your phrase that you used on me earlier: "You're misunderstanding Pascal here"! The method I've shown has 2 parts and 2 effects with 1 conclusion, these are:

1) Part 1: Reduce TDP to a max level that corresponds to the temperatures you want to see. Effect 1: prevents higher voltage points from being used.
2) Part 2: Apply your max stable overclock. Effect 2: raises frequency for each existing voltage point to the optimal frequency for that voltage point.
Conclusion: you "both get more performance AND less heat by using it." (that's quoted from your post above by the way).

Yes, you don't have the Power Slider that lets you adjust TDP in a notebook, so you have to edit the vBIOS and flash it using a programmer - but it's not impossible.

 

Yes, Precison X has one that works better (easier and several modes) but I really don't see any benefit. I think it and the one with MSI AB are both worthless for overclocking. My GPU overclocks are less stable and benchmarks are lower if I mess with the voltage curve.

It works fine, but it does't have the feature mentioned above (which is probably irrelevent).

 

1070N has 115W TDP

They get limited to mid 1500MHz ranges

You want me to LOWER TDP

Then OVERCLOCK

And this is going to give more performance?

Seems I need somebody with a 1070N to test. I'm gonna wait until @Coolane gets back his, or maybe @Scerate will test it?

I never mentioned for overclocking, 2100MHz+ like you're aiming for is going to need more voltage. I'm talking for when you normally game and leave your GPUs at stock. Probably could shave a little 5c off each card on say your video with Crysis 3 from a while ago.

 

I've shown in my testing how it can be done, my earlier post with screenshots while my GPU was limited to 50% TDP: the overclocked version was running at a lower voltage & a higher frequency as a result - that's effectively undervolting & overclocking. I really don't think I can explain it any clearer than I already have over the last few posts that we've had since last night & into this morning, so I'm not gonna discuss it any further with you. I suggest you swallow your pride and take a closer look at the testing I did, have an open mind & see if you can work it out - I personally think you've got yourself in an ego fluster ever since your Math blunder a little earlier in the thread, but no matter. I wouldn't have been quite so vehement in my posts to you had your earlier posts to me been not quite so disparaging.

 

I don't let a stupid thing like ego stop me from learning. I indeed didn't see your edit of the original post, because you just gave text before, and I'd been on another page after replying. I had opened the page and forgot to reply and slept, hence why it was so late, despite me not seeing the pictures.

However, that's why I just asked two people who have 1070Ns who can adjust power and voltage limits to do a test between both. I want to see it work on these notebook cards. As far as I've seen in the past trying to overclock 1060Ns didn't really do anything. Firestrike scores didn't change for graphics etc.

Seeing as how your graphics card clocked up (but temperature remained the same), are you willing to finish those firestrike tests so I can see if the GPU score changed up? Your screenshots show that you canceled the run.

 

6% increase in Graphics Score in Firestike Extreme when overclocking in TDP limited scenario:
50% TDP limit, Founders Edition Simulated Base Clock (1505Mhz Base Clock)


50% TDP limit, Max Stable Overclock (1682Mhz Base Clock)



P.S. You say in your previous posts that I cancelled the run & didn't let the tests finish, it was a 20min Firestrike Graphics Test 1 loop I was doing for those tests in that earlier post, so running all tests was not the plan - it was to best illustrate the difference in sustained clock frequencies over long gaming runs - to show the full advantage of overclocking while TDP limited.

 

interesting. I really want to see how this will work on the notebook cards. How about firestrike extreme? If that's not too much trouble.

 

(That was Firestrike Extreme)

 

Welp, I have some extreme tunnel vision. I only saw the scores and clockspeeds and didn't look at the name of the benchmark.

Thanks then. Only need to wait for @Scerate or @Coolane now to echo your findings, see how that works.

I guess the information I had was false, I thought OCing 1060Ns didn't do much of anything; they all score mid 11000s in Firestrike OC'd or not. But then I just checked the benchmark results and in the top scores there's mid 13.5k runs which is what the desktop cards get.

*brain explode*

Unfortunately if modding the vBIOS is needed one could just allow more power, but I guess some of the notebooks that really can't draw that power could benefit.

 

Ha, it happens!

 

Well glad we resolved that. Still interested to see how the notebooks do it. Since their max power is only 115W reducing that would be sending it to about 100W. I already linked to testing at different power limits in the NBC article.

 

Yep. The only reason to reduce TDP though is if the GPU is getting too hot for your liking, otherwise no point & you'd just overclock rather than limiting power too. Limiting Power and Overclocking is effectively like undervolting & overclocking though in as much as it prevents the higher voltage points from being used while simultaneously optimising frequency for each used voltage point, so it has those benefits when your GPU is getting too hot for your liking.

EDIT: the best results in efficiency increases (performance per Watt) will be seen in individual GPUs that can overclock by the largest percentage, as well as when reducing Power Limits by a greater percentage. So when you rightly say that notebook users might only reduce their Power Limit from say 115W down to 100W, that small reduction in power would diminish some of the potential efficiency savings.

 

@Falkentyne can help you testing as he use the TDP mod on his 1070. Or @thegh0sts

 

I did a similar test before. If I draw the default GPU scores together with the OCed scores on the same graph, I can see that a lower TDP OCed score can come close to a higher TDP score with default clock.
For example, overclocking with a tdp lowered to 90W can get closed to the default GPU score at 100W (or higher than 95W default from the graph);
the same applied to 125W OC vs 150W default. A 150W OCed score later surpassed the default score at 170W.
The default scores have the GPU undervolted to its lowest stable limit, if there's no under-voltage was done, the gap will be even bigger.

 

Well what I was considering was a full four-way test, kind of.

Stock run (no core clock changes, no voltage changes, no TDP changes), then reduced-TDP OC'd run (no voltage changes, 100W limit down from 115W), then undervolted run (no core clock or TDP changes, minimum voltage curve you can use), then undervolted + OC'd run (no TDP changes). All with temperature readouts for the tests. Would like a longish test like Unigine Heaven or Valley for this, they still give scores.