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

Wow, so those default/stock GPU scores were all with an undervolt applied, so as you say that would make the gap between OC'd vs stock even bigger in terms of performance per watt if your default GPU scores were actually at the stock voltage! Well, it certainly supports my findings that overclocking in a TDP limited scenario will improve performance & efficiencies. (+rep for the testing & presentation in a graph)

 

Right. Without undervoltaging the default clocked GPU, I believe the scores would be 5~15% lower.

I can test that combination out when I get the card, now it's being reinforced and enhanced, .

 

What do you think I'm waiting for? xD

Sent from my OnePlus 1 using a coconut

 

Wow, this is pretty cool, so spurred on by the benefits of overclocking in TDP limited scenarios, I decided to test my card with various TDP limits to work out the most efficient TDP which operates at a 'silent' fan level; not only that but I decided to see (on the off chance!) whether reducing the TDP would allow for a higher stable overclock - crazy right! This is a graph of my TDP vs Superpostion 4K benchmark score (chosen because uses lots of power):

You can see that the efficiencies start to drop off more dramatically after the 70% TDP value, however my card is 'silent' & cool at 80% TDP on Auto Fans, so that's the TDP I'm gonna choose. Note that my GTX 1070 Amp Edition has a high TDP of 220W, so 80% of that is still 176W, which is still quite a lot for a GTX 1070 as the Founders Edition cards are only 150W cards.

I then wanted to experiment to see if I could gain a higher stable overclock now that I have a reduced TDP of 80% - my theory being that the high load / high voltage & high frequency points would be cut out due to the reduced TDP ceiling & therefore perhaps the card is operating in a more efficient & more overclockable/stable zone on the curve. Well, it turned out to be true, I managed to increase core clock by 2 overclocking notches without even increasing voltage - so increased core by an extra 25Mhz. Here is a quick summary of my different stable overclocks, two of which will be used for testing today:

1) +75Mhz (1682Mhz Base Clock), no added voltage, 120% TDP, Aggressive Manual Fan Curve
2) +87Mhz (1694Mhz Base Clock), 50% added voltage, 120% TDP, Aggressive Manual Fan Curve
3) My new overclock found today based on 80% TDP allowing me to overclock higher: +100Mhz (1707Mhz Base Clock), no added voltage, 80% TDP, Auto Fans ('silent').

Below are my results comparing Profile 2 vs Profile 3:
Results:
Profile 2) Previous Max Stable Overclock at 120% TDP with Added Voltage & Noisy Aggressive Fan Curve (1694Mhz Base Clock):


Profle 3) New Max Stable Overclock at 80% TDP / no added voltage / 'Silent' Auto Fans (1707Mhz Base Clock):


From above results: 80% TDP new overclock is same score in Firestrike (slightly faster), but instead drops just 1% performance in 4K Superposition (reason being is that Superposition wants to suck down more power than Firestrike, even though Firestrike is still TDP limited for the most part at 80% TDP). This new higher overclock at 80% TDP is a big win for me over the previous overvolted 120% TDP profile I've been using historically - same performance but a lot quieter and lower power draw by typically 20W!

Who'd have thought that lowering the TDP would allow for a greater overclock! (from previous 1682Mhz Base Clock up to the new 1707Mhz Base Clock)

Footnote: stability of overclocks both new & old tested with looped 2hr+ run of Firestrike Extreme Graphics Test 1. Historically, for me, this finds an unstable core overclock within 45mins if it's gonna fail, otherwise it tends to be stable for 8hrs+. (Use TimeSpy Stress Test to weed out unstable VRAM overclocks - most sensitive.).

 

Looks like you've got yourself a nice discovery; those sub-100 points differences in superposition is going to mean virtually nothing in games.

How about trying no added voltage, higher TDP allowance, and a higher overclock? Think you could make 2100MHz?

 

Yep, that's the only setting I didn't show in my last post, but it is a setting I have tried & also run historically in the past. My highest overclock with zero TDP restrictions (120% TDP) and no added voltage is +75Mhz on the core giving 1682Mhz. (My bad, I mistakenly said in my post above that my added voltage overclock was 1682Mhz, but it was actually 1694Mhz - I will correct post above - adding voltage in that scenario allowed me to add an additional +12Mhz. NOW CORRECTED ABOVE) So, yep, these are my different stable overclock configurations:

1) +75Mhz (1682Mhz Base Clock), no added voltage, 120% TDP, Aggressive Manual Fan Curve
2) +87Mhz (1694Mhz Base Clock), 50% added voltage, 120% TDP, Aggressive Manual Fan Curve
3) My new overclock found today based on 80% TDP allowing me to overclock higher: +100Mhz (1707Mhz Base Clock), no added voltage, 80% TDP, Auto Fans ('silent').

2100Mhz is actually achievable now with my newly discovered overclock (#3 option above):

It happens to be the highest boost clock available to the GPU when running Base Clock of 1707Mhz. It won't stay there during gaming though - depending on temperatures & load - remember the 80% TDP restriction of course. But, for most games for most of the time the 80% limit is not reached, so I'm getting increased performance in those vs my benchmarks I've shown today. I use the Firestrike benchmarks for stability testing, because in my experience the Graphics Test 1 is the most sensitive to unstable core clocks - I tune it to that & then it's stable in all my games as a result.

 

Mmm I'm thinking if you left about 90% TDP and no added voltage and bumped fanspeeds a bit you could probably hold that 2100MHz straight, then. Have you tried tuning the curve to see what the min voltage you can achieve at 2100MHz is? I'm certain cards can handle 2100MHz+ at 1.063v max, because the mobile cards do it with a modded vBIOS.

 

(Ha, I'm still in the process of trying to edit that earlier post of mine (Post #406) to correct the mistake - but it's having ramifications on the wording of that post, making it hard to correct - I will correct it soon!) - NOW CORRECTED!

That's a good idea you have there. My 120% TDP setting is a little misleading though, because my card never gets close to 120% TDP, but it does get to about 93/94% TDP during the Firestike Extreme Graphics Test 1 - so my TDP reduction to 80% is not quite as extreme as it initially sounds. So, I don't think that the small reduction to 90% would make much of a difference - but I think it's worth a try actually - the only problem is that it negates the reason for me choosing the silent 80% TDP profile - would be interesting to try for Science though, I'll try it.

 

If 80% TDP gives you the OC but can't hold it due to power or heat, then 90% TDP and a slightly faster fan profile should make it consistently about 2100MHz for you, is my thought process.

 

Although thinking about it - without a TDP restriction it wasn't stable before at high loads at the highest voltage & boost points, which is how I've found that a TDP restriction allows me a higher overclock - now if I remove that TDP restriction completely then it will cease to be stable, but perhaps I could push it up from 80% to 90% TDP and only have a miniscule TDP restriction so it's as close to 2100Mhz as possible - but I know I'll never get it to stay at 2100Mhz at all loads & for it to remain stable (because with 1707Mhz base clock 2100Mhz is the ABSOLUTE MAXIMUM boost). I'll try though to see how close I can get it to hover near 2100Mhz though.

 

You can only really bump 100MHz? If you bump 150MHz does it just kind of break or something?

Well I guess you've got a lot of checking to do xD

 

Well, it's run 40mins so far at the 90% TDP at the previous +100Mhz - Firestrike Graphics Test 1, so looks stable so far. I've kind of assumed that I can't get a higher overclock than +100Mhz because I couldn't believe reducing TDP could really make my overclock increase massively massively: +100Mhz is already 2 notches above my previous best overclock with no added voltage. I suppose I could try adding another notch or two to see if it's stable - I'm a bit tired right now & can't remember if I had already tried that earlier today or not - ha!! In 20 mins I'll run an obscene +150Mhz at 80% TDP to see if it's stable - letting my current testing run described above finish first.

 

Heh, good luck! Sleep and start fresh what can I say?

 

Ha, yep, I won't let it disrupt my sleep, I have done in the past! I'll be asleep in about an hour & a half, just time for testing one further overclock.

EDIT:
+150Mhz @80% TDP not stable, failed within 1 minute
+125Mhz @80% TDP not stable, failed within 5 minutes
+113Mhz @80% TDP....testing, but to be honest that's only 1 notch above my +100Mhz stable overclock I achieved earlier today, so praps not worth the testing!

 

Interesting. But then the question is how does 90% and 100% work, and how's dropping max voltage to 1.063v work =O

 

Ok, replying to my own post here, because after further testing I've found that my previous conclusion that "reducing TDP allowed for a higher stable overclock" is untrue! Instead, my further testing has shown that my GPU now magically has an additional 2 notches of overclocking stability on the core regardless of TDP setting - this ain't no bad thing for me, but I apologise for my incorrect conclusion previously. I'm trying to work out some of the possible reasons for this sudden magical extra overclocking potential:
1) Temperature of GPU has decreased about 4-5 degC since I have done some case mods to improve GPU airflow, combined with liquid metal on the GPU core (although liquid metal aspect only gave me 1 degC drop).
2) Has Firestrike had updates in the last few months that somehow allow for it to be less punishing to overclock stability? (This was my method of overclock stability testing).
3) Any changes in NVidia drivers that allow for a higher overclock? (Although historically in every GPU I've ever owned the driver hasn't made any difference to a 100% stable overclock)
4) Windows 10 Creators Update perhaps somehow allowed for higher stable overclock?

Those are just some of the changes that have happened since I last tested my overclock. I'd recommend you guys try testing your max stable overclock again, perhaps something has changed that is common in our ecosystems that allow for a higher overclock?

Here's a screenshot of my now balls-to-the-wall max stable overclock with 50% overvolt at 120% TDP setting in a 45min Dirt Rally benchmark loop (+113Mhz overclock, 1720Mhz Base Clock, 2113Mhz max boost clock, 2075Mhz stable boost clock seen throughout latter portion of run):

Although, one thing my testing has shown is that if I want a quiet system I can just run +100Mhz at zero added voltage at 80% TDP on a 'Silent' Default Auto Fan Profile and I'm only sacrificing 2% performance along with gaining 20W power savings (in situations where TDP limiter does kick in, doesn't on Dirt Rally) - so this is the one to run at most times. As seen in following screenshot, this yields a consistent 2050Mhz in Dirt Rally Benchmark - really not giving much performance away here!


Apologies once again for incorrect conclusion I drew in my quoted post.
@D2 Ultima , @Mr. Fox , @Coolane , @Papusan , @dspboys , @Ionising_Radiation , @Miguel Pereira , @hmscott , @ThePerfectStorm (in case you guys aren't subscribed, and you've all been active in the latest convos with likes or comments).

EDIT: and if anyone wants to post Volta news than please do so, because I'm aware that I've been instrumental in derailing this thread, although to be fair no one had posted in this thread for months, but welcome to post any Volta news & thoughts!

 

Very interesting, thanks for the tag. I certainly hope Volta overclocks this well. I do look forward to seeing what Eurocom can do with a Volta + Coffee Lake Tornado F5 / Tornado F7. @Mr. Fox @Talon @Papusan @Diversion @hmscott @Ashtrix @D2 Ultima

Sent from my SM-G935F using Tapatalk

 

Very interesting, thanks for the tag. I certainly hope Volta overclocks this well. I do look forward to seeing what Eurocom can do with a Volta + Coffee Lake Tornado F5 / Tornado F7. @Mr. Fox @Talon @Papusan @Diversion @hmscott @Ashtrix @D2 Ultima

Sent from my SM-G935F using Tapatalk

 

Suddenly the world makes sense again xD

There was an update earlier this year where OCs crashed a lot on Firestrike. I think it was since resolved. There were people with superclocked cards who couldn't finish a run at stock.

 

So any word on Volta Gpu's?What can we expect? Improved power efficiency and more power?
While the 10** series has been pretty amazing for the most part I think its time for new series to release.

 

Most people think Volta is gonna be released sometime in 2018 - me too. Here's one of the latest articles I could find on Volta news from a quick google:
https://www.pcgamesn.com/nvidia/nvidia-volta-gpu-specifications

And here's another one I posted earlier in this thread (but in the article ignore the 32 inch squared - that must be a typo or something!):
https://www.notebookcheck.net/Nvidia-reveals-more-details-on-the-upcoming-Volta-GPUs.244332.0.html

NVidia are seeming to be very tight lipped about Volta, doesn't seem to be much news nor many leaks about it recently. Personally, I'm not particularly interested in Volta, but that's probably because I won't be upgrading to it - Pascal performs stellar at 1080p (not only) and I just don't think there's much hunger in the market for a new generation of cards right now - on top of that you have AMD struggling to be competetive with Vega - BAHH, who needs a new architecture right! Ha!

 

I feel where at a strange point now.
A 1080ti handles 1440p 120hz+ remarkably - even a 1080 for that matter (to make it relevant for laptops). I can't think of a reason I'd want to upgrade a GPU unless it's going to give me the same frames at 4k - which I doubt the next generation will.
I love high frame rate gaming but it's a real Pandora's box. 60fps/Hz almost immediately becomes a deal breaker after about 1 weeks use.
Those who are well content (as I use to be) with 60fps/Hz... "don't go 120hz" unless you're prepared to be disappointed with anything less down the road. It's totally limited my selection in future laptops. Even not gaming laptops. Sometimes I wish I never opened that box.

 

Ark Survival Evolved notebook and desktop benchmarks (09/02/2017)

"Survival Evolved has been officially launched it is time to take a look at its system requirements to figure out whether or not the game gulps up computing power as fast as it does player avatars."

 

Ha, exactly, it's hard to go back when you've experienced 144Hz for example! To be honest I need a 1080ti if I'm gonna crank up details to the max in all games and want a stable 144fps in some of my games, so in that kind of a market there's certainly some reason for Volta to come sooner rather than later. As it stands though, with my GTX 1070 I'll turn down game details to achieve 144 fps in the one or two games that require high frame rates (competetive first person shooters), but for all others, especially story driven ones like Mass Effect Andromeda, I just use a frame rate limiter set to 70 fps and G-sync does the rest - so I've got no need for Volta. Die hard no compromise 144Hz owners are for sure gonna want Volta to come tomorrow though! I did say in my previous post that there would be no real market for Volta, so I suppose I need to reconsider that, but I guess 144Hz owners are a small proportion (although a growing one), same goes for 4K.

Yeah, I saw that review yesterday, and I remember thinking: damn that game runs badly for how bad it looks! That's another game that needs Volta if you want high frame rates at max details, I was gonna say it's probably badly optimised given how rubbish it looks for the frame rate, but it's UE4 so shouldn't be badly optimised per say. I'm not gonna buy that game, not my thing.

 

Yeah, as someone who hasn't been spoiled by high refresh rates and enjoys gaming at 1080p, I see little reason to upgrade to Volta. A laptop with a GTX 1060 can max out most titles at 1080p@60fps. Eventually I may want to upgrade to a laptop that can handle 4K@60fps at max settings, but I doubt Volta will offer that--at least not on mid-range GPUs.

 

The high end will easily have 4k 60fps covered and I expect the mid range might even be reasonably close.
But your right about a 1060 being plenty for 1080p @ 60fps.

 

If they can make their Volta version of 1070 easily push 4k @ 60FPS, I'm ready. Otherwise, meh. I'd much rather see 3K/2.5K (2880x1620 / 2560x1440) quality LCD's with current tech at cheaper prices. Or even better, 16:10 equivalents of 3k/2.5k. The current 1070 and 1080 can easily push 2.5K @ 60FPS.

 

I think I'm going big this time, whether they call it the 1180 or 2080, I want it in my next laptop.

Good enough is never enough.

 

Man, a GTX 1170/2070 upgrade over my 860M is going to be huge. Imagine cramming a GTX 1080 Ti into a 15" notebook...

Fire Strike Graphics scores jumping from ~ 5000 to ~ 30000. That's a sixfold improvement.

 

It'll blow your mind, you're definitely ready for an upgrade!

 

Are you planning to get Volta regardless of whether it uses HBM2?

 

Yep—I will get Volta regardless of the VRAM technology used. I hope Volta uses HBM (probably HBM2) at the highest level, maybe the 1180/2080 Ti, but I get a feeling the cut down GV104, and the GV106, 107 and 108 chips will use different memory technology depending on the bandwidth. If it turns out the GTX 1130 is only as powerful as a GTX 750 Ti, then there's no point in using memory with so much bandwidth because the GPU couldn't feed it, anyway. Here's a quick prediction (read: pulled out of my ar*e):

GTX Titan Volta: HBM2
GTX 1180 Ti: HBM2/GDDR6
GTX 1180: GDDR6
GTX 1170: GDDR5X
GTX 1160: GDDR5X
GTX 1150 Ti: GDDR5
GTX 1150: GDDR5
GTX 1130/150MX/1140: GDDR5

GDDR6 is already postulated to be as fast, and HBM2—as demonstrated by AMD Vega—is prohibitively expensive. I don't think Nvidia will want to use it widely until prices fall to a more reasonable level. By then, we'll have HBM3, which is supposedly faster, less power-hungry and also cheaper.

 

Yeah I need a volta titan for good gaming experience.

Titan xp isnt enough for 5k(5120 x 2880). I also want to move to 8k in the future.

 

Jokes aside... I think the pared-down GV104 chip, whatever it's called, will more than suffice for whatever gaming I want to do. Going by history, the X70 series has always outdone the previous generation's X80 Ti. 970 > 780 Ti, 1070 > 980 Ti, 770 > 590/690. Like I said earlier, the cut-down GV104 chip will very likely perform as well as the GP102 (1080 Ti) we have today. 4K 60 FPS gaming will become a reality, and 1440p 120 FPS gaming will become commonplace...

 

Ehh you want proof for my 5k monitor and the lack of fps I get with a GP102?

I think you are expecting too much. I need a minimum of a Volta Titan. 8k will likely require the one after that.

 

8K will probably require 2x Titan V for smooth 60fps+ experience.

Single Titan V should be able to at least provide 4K 60fps+ maybe 120fps+

5K @ 60fps will probably be doable with single Titan V

I don't think the 1180 will be much faster than a 1080Ti though, maybe 15%.

 

770 was Not faster then 690... It was a pimped 680.

That tendency you talk of is recent.

Enviado do meu MHA-L29 através de Tapatalk

 

Progress toward consumer release now that commercial release has been reached?

Nvidia Ships first Volta-based DGX Systems
http://www.anandtech.com/show/11824/nvidia-ships-first-volta-dgx-systems

"This Wednesday, NVIDIA has announced that they have shipped their first commercial Volta-based DGX-1 system to the MGH & BWH Center for Clinical Data Science (CCDS), a Massachusetts-based research group focusing on AI and machine learning applications in healthcare. In a sense, this serves as a generational upgrade as CCDS was one of the first research institutions to receive a Pascal-based first generation DGX-1 last December. In addition, NVIDIA is shipping a DGX Station to CCDS later this month.

At CCDS, these AI supercomputers will continue to be used in training deep neural networks for the purpose of evaluating medical images and scans, using Massachusetts General Hospital’s collection of phenotypic, genetics, and imaging data. In turn, this can assist doctors and medical practitioners in making faster and more accurate diagnoses and treatment plans."

NVIDIA Ships First Volta DGX-1
https://www.top500.org/news/nvidia-ships-first-volta-dgx-1/

"CCDS is well-versed with NVIDIA gear. It received the first-generation DGX-1 last December, which was equipped with eight Pascal-based Tesla P100 GPUs. With 170 teraflops of 16-bit floating point performance, it was the most computationally dense GPU-based platform you could buy at the time

However, that P100-based appliance pales in comparison to the new V100-powered DGX-1, which, courtesy of Volta’s Tensor Core technology, offers 960 teraflops of mixed FP16 and FP32 matrix math. That’s nearly a petaflops-worth of deep learning performance in a 4U box."

NVIDIA Ships First Volta-based DGX Systems
https://www.reddit.com/r/hardware/comments/6ynvrt/nvidia_ships_first_voltabased_dgx_systems/

 

@Papusan—yeah, I still want to get a desktop

Just that I want a notebook, too. Can't bring my desktop into lecture halls, can I?

 

I know

Spoiler: You can have both :)

Or maybe a smartphone. They are alredy passed 6" and can maybe be a outsider

 

So, thoughts? @Papusan @Mr. Fox

 

What's the synopsis? That's a 24 min video, I started watching it, and it seems to be an NVidia bashing video with regards to releases starting back all the way to the Fermi range - at least provide a synopsis or a time stamp to view the video from with regards to the pertinent point for your reason for posting. As much as I like the Scottish accent, I don't really want to listen to 24mins of it to garner the point you're not making - not on a Sunday anyway! (I skipped through parts of it, to try to quickly get the point, but to me it's not worth the effort, what's your point?). I thought as a lawyer you might like to try to get to the facts as efficiently as possible, or do you like the small print too much.

 

Here's 27 minutes more of Scottish joy for you, boyo.

A History of Nvidia GeForce, Part 1 - Fierce Competition


It's not as good when jumping around, you'll miss the subtle nuances, be sure and watch Part 1 first, and then Part 2 afterwards

 

I was about to sit down and watch this. I usually like his analysis. Theyre usually quite contrary and well articulated.

 

Ha, well at least now there's a Part 1! Not sure I can stomach it though, will have a quick listen now though to see if there's a satisfying intro that puts it into picture quickly! Thanks.

EDIT: Sh*t, same approach with the only intro being that it's a 'history' of NVidia GPUs, yet still no main synopsis given at the beginning of his vid - I get the feeling he's trying to compare increases in GPU performance from NVidia generation to generation, but I'm not willing to spend 40+ minutes of my time listening to an annoyingly syncopated voice over to finally get to the point the guy's trying to make in the end. Now, if he'd have come out at the beginning of his vid saying now this is what I've found in a couple of short sentences & now I'm gonna show you the detail -then I could live with it - otherwise it's just a waste of time. Anyone know what the final/ultimate point this guy is making in his two 25min vids?

 

First, as a lawyer, you present evidence and lead people to your conclusions. If they refuse to take in evidence or give it weight, all you can do is attempt to persuade them.

Second, part two discusses how Nvidia pulled the switcheroo, making the 80 series filled with the mid-range chips when it used to be the number for high-end. It then further discusses Nvidia's techniques to milk and bilk the market. It does note where Nvidia did very well, such as Maxwell, but also shows how they were screwing consumers due to that. He addresses fanboys common statements (both sides), while concluding AMD gave up and decided to bilk their fanboys as well (so there is some AMD bashing in with the video).

Finally, you don't pay me to perform analysis, so don't act like a sniveling twit and respond wanting me to do the work for you. If you have no interest in it, move on with your day. Don't respond. Don't tell me what I should do. Just let it go, as you made up your mind not to watch it, which is your right.

 

Yeah, you like the small print!

 

The term is fine print. And forgive me if I take the little respite of my day, from working on a deadline, having a cousin's son die yesterday, having 5 family members in shelters just south of Tampa, Fl with a hurricane coming in, and not taking the time to accommodate you on your Sunday with your equivalent response to TL;DR. That is all on me. I should accommodate you at every turn. I'm sorry masta'.

 

Hey, I don't know about your immediate situations, I based my reply on your immediate post (and it's vagueness & lack of point), how I know you from your other posts, my overall impression of you, and how you've responded to my own posts in the past - the fact that you're having a rough time at the moment, I understand that. You're obviously not here to accommodate me, of course.

I hope you guys get through those awful storms, they do sound horrendous!

 

And sorry. TL;DR is one of my pet peeves, and equivalent. So, when my profession is brought up and I'm told how to spend my free time, while stressed (the part you couldn't have known about), I can get a little pissy. To that, I'm sorry.