NVIDIA Pascal Thread

[Sweepr]

To clarify yes using four cards can have negative scaling when your cpu isn't fast enough and we are not using graphically demanding options.While the second option is easy to fix the first not so much, I believe even Intel's fastest isn't fast enough to feed these four beasts.From the conclusion of the review you linked

" Especially for 4-way SLI, we noticed that performance will often times decrease rather than increase."

I believe some site cant recall now tested and concluded that 3 way was prolly the best way to go but that doesn't mean you can't grab one more

That makes sense, probably explain the poor scaling results from some reviews. But I don't expect anything less than a heavily OCed Intel HEDT powering a 4-way SLI system, and of course maxed out settings.

 

[MrTeal]

That makes sense, probably explain the poor scaling results from some reviews. But I don't expect anything less than a heavily OCed Intel HEDT powering a 4-way SLI system, and of course maxed out settings.

DX12 will certainly bring some interesting changes in terms of multi-GPU setups. For very well coded engines I don't think it's unlikely that quad scaling will improve. I'm much less bullish on multi-GPU overall though, as niche markets that require extra effort from the dev could see some pretty crappy support given the publishers desire to push out a console port as quickly as possible to soak up a few extra PC sales.

 

[antihelten]

If that's not enough, more results: http://uk.hardware.info/reviews/603...-way-sli--4-way-sli-review-insane-performance

Zero negative scaling in any of the 9 games tested (up to 4-way Titan X SLI). Better than random tests using different VGAs with 4GB VRAM. Cya!

Eerm, 2 of the 9 games clearly showed negative scaling for 4-way SLI relative to 2 or 3-way SLI (3 if you count Civ: BE, but that is only at lower settings):

Also you and Arachnotronic should probably be careful with the benches from Digitalstorm, a company that has a clear interest in selling you as many GPUs as possible, and as such is anything but unbiased in this regard. I mean you wouldn't trust benches directly from AMD or Nvidia would you?

 

[Sweepr]

Eerm, 2 of the 9 games clearly showed negative scaling for 4-way SLI relative to 2 or 3-way SLI (3 if you count Civ: BE, but that is only at lower settings):

This is explained in the article. Minor negative scaling can happen in CPU intensive titles and/or if you're not running at extreme resolutions with high settings (like you should if you own four Titan X's).

Do note that this does not necessarily mean that Nvidia's drivers aren't up to snuff: another explanation that is just as plausible is that even the fastest Intel CPU simply isn't fast enough to provide four of these monstrous graphics cards with data. However, with performance scaling being excellent in certain aspects of benchmarks such as 3DMark, overclockers will (as is tradition) certainly benefit from using a 3-way or 4-way SLI configuration.

They're not blaming NVIDIA here, and as long as you're GPU-limited, these monstruous systems can perform better. Probably doesn't make sense to spend that much when 2-way Titan X SLI (or even GTX980 Ti's) can deliver most of that performance, but then again many enthusiasts don't care about this.

 

[antihelten]

This is explained in the article. Minor negative scaling can happen in CPU intensive titles and/or if you're not running at extreme resolutions with high settings (like you should if you own four Titan X's).

They're not blaming NVIDIA here, and as long as you're GPU-limited, these monstruous systems can perform better. Probably doesn't make sense to spend that much when 2-way Titan X SLI (or even GTX980 Ti's) can deliver most of that performance, but then again many enthusiasts don't care about this.

I wouldn't really call the stuff you see in CIV:BE minor negative scaling (it's a 25% performance drop versus 2-way SLI), and besides the claim that Erenhardt made was never that there is major negative scaling, simply that there is some amount of negative scaling, something which you denied (prior to your edit):

Zero negative scaling in any of 9 games tested @ 4K (up to 4-way Titan X SLI).

 

[Sweepr]

I wouldn't really call the stuff you see in CIV:BE minor negative scaling (it's a 25% performance drop versus 2-way SLI), and besides the claim that Erenhardt made was never that there is major negative scaling, simply that there is some amount of negative scaling, something which you denied (prior to your edit):

All he did was attack/mock AdamK47's choice, probably because it's four NVIDIA cards instead of Fury X's. He didn't say major, but he implied it was substantial ('downgrade'), and overall (out of the 9 games tested at 4K) we can see that he is wrong.

Nice. I really like your sense of humor.

Do you have 2 of those titans X disabled, or are you enjoying the downgrade in performance with more than 2way sli negative scaling?

 

[antihelten]

All he did was attack/mock AdamK47's choice, probably because it's four NVIDIA cards instead of Fury X's. He didn't say major, but he implied it was substantial ('downgrade'), and overall (out of the 9 games tested at 4K) we can see that he is wrong.

I agree that his post was mocking in nature, and as such it should just have been ignored or marginalized (which ironically is exactly what AdamK47 did, when he said that negative scaling wasn't an issue in the majority of games), instead some people started arguing that negative scaling doesn't exist at all, when it quite obviously does.

As to whether or not he implied the downgrade was substantial I couldn't say, since I'm not a mind reader.

á propos mocking (your reply to Erenhardt):

... Don't be jealous.

Either way this is all far off topic, so I'll leave it here.

 

[Sweepr]

á propos mocking (your reply to Erenhardt):

Either way this is all far off topic, so I'll leave it here.

Other than that, coupled with brand preference, I don't understand why he would mock/attack Adam's choice.

 

[thesmokingman]

Other than that, coupled with brand preference, I don't understand why he would mock/attack Adam's choice.

A lot of times when you are heavily invested in something it's hard to look at it objectively so we tend to glass over the downsides. And there are a lot of downsides to running quads. I've been running quads for a few generations now, been there done that. Anyone who says that they get great performance all around is not telling the truth. It's that simple, and it's not the whole story. You have to have a lot of criteria met, like having a lot of psu to supply them gpus, a big cpu clocked enough to supply them gpus when the game is poorly threaded which is like most dx11 tittles, and then all of the other supporting systems, fast ssd's, lotsa ram, big cooling, etc etc. It's no small feat to support quad gpu whether its 4 titans or 4 furys or what have you. And then often times it still sucks, so well we tend to only look at the bright side which is like a handful of games in each genre.

 

[Sweepr]

Trying to get back on topic, launching the mobile Pascal GPUs first makes sense (if current rumours are true). Smaller dies are cheaper, perf/watt counts the most there and they don't want to lose their 'momentum'.

Nvidia’s desktop discrete GPU shipments were down -7.56% from last quarter; and the company’s notebook discrete GPU shipments increased 34.2% increased 8.4% from last quarter.

 

[Qwertilot]

Some pressure from Intel pushing big IGP's too. Combine mobile stuff with a huge Tesla card (turning into a titan at some point), and I think they could easily cope with a delay on the stuff in between.

If they have to then the 980/70/60 etc are small/power efficient enough to drop a market bracket or two with price cuts.

 

[airfathaaaaa]

i saw somewhere that hynix will start production of hbm2 on november? is that true? that means the big ones from both companies will come no late q1 2017?

 

[Qwertilot]

Maybe. Don't forget that there's production and production

Like with the availability of Fury(X) for the first few months after its launch vs now say. The Tesla's could probably cope with that sort of production quantity, the 980ti equivalent probably couldn't.

 

[Silverforce11]

i saw somewhere that hynix will start production of hbm2 on november? is that true? that means the big ones from both companies will come no late q1 2017?

Volume in Q3 2016. Samsung started on HBM2 but we don't know about volumes.

 

[Game_dev]

Volume in Q3 2016. Samsung started on HBM2 but we don't know about volumes.

Samsung has already started mass production. This may be why Pascal is arriving shortly

 

[IEC]

Samsung has already started mass production. This may be why Pascal is arriving shortly

Desktop Pascal isn't arriving "shortly".

The first launch will be mobile GPUs.

 

[Mahigan]

The big GPUs will likely arrive before X-Mas for AMD and Q1 2017 for NVIDIA.

The GTX 970/980 replacements will likely arrive around August of 2016 for NVIDIA. But will be announced around June.

The R9 390/390x replacements will likely arrive before their NVIDIA counterparts so around July of 2016. Probably only a few weeks separating both GP104 and Polaris 11.

AMD has shown one of their low end parts already (Polaris 10) and has, behind closed doors, shown off their Baffin XT (Polaris 11 or R9 390/390x replacement). The Vega 10/11 line (formerly Greenland) have yet to surface.

 

[HeXen]

So in other words, it's more likely I'll find one in stock after Xmas sometime next year. That kind of sucks, I'll be wanting moar power for my Oculus Rift before then.

 

[Game_dev]

The high end desktop Pascal is looking like June. I expect a paper launch of Polaris in the fall. Since AMD has to wait for hynix. They don't have the money to place a large chip order so expect a soft launch until well into 2017.

 

[crisium]

^You heard it here first, folks. GP100 980 Ti / Titan X replacement in 3 months.

 

[AnandThenMan]

The high end desktop Pascal is looking like June. I expect a paper launch of Polaris in the fall. Since AMD has to wait for hynix. They don't have the money to place a large chip order so expect a soft launch until well into 2017.

Nvidia is being incredibly secretive for a card that is going to launch in 3 months.

 

[Osjur]

@crisium: Don't forget that Polaris paper launches in fall because AMD used all their money on that Polaris chip which they demoed two months ago @CES so they can't produce more.

They also have to wait for hynix to produce HBM2 chips for reasons only known by few on this board.

Sooo... where can I get those crystal balls :hmm:

 

[steve wilson]

So in other words, it's more likely I'll find one in stock after Xmas sometime next year. That kind of sucks, I'll be wanting moar power for my Oculus Rift before then.

Mine is due in July and I'm rocking a gtx770 at the moment. All I play is CS GO and rocket league, so it's fine for now, but it isn't going to cut it come July. I feel it's pointless buying a 980ti now, so I'm just going to have to wait it out.

 

[airfathaaaaa]

Samsung has already started mass production. This may be why Pascal is arriving shortly

but nvidia doesnt have a contract with samsung.. amd has samsung started on january 19 if i remember correctly

 

[stormkroe]

Dude, admit when you're wrong, be a man. You asked for examples of negative scaling and @Erenhardt provided it, 4 GPUs scaled backwards. -_-

Some people...

LOL @ 0 fps min ... some major stutter issues there. haha

Those aren't mins, just another res that presumably failed (which would be worse than stuttering I suppose
Unless you're talking about the graph one of the posters mentioned a few posts back, in which case if you don't mind, please drop a link (in a PM if you want since it's so OT now) I'd love to see what kind of frame delays the massive overhead causes in this case.
Thanks