Gtx 1080 benchmarked

JohnC0817 said:
Thanks for the find. Didn't hear of 5x till the 1080 was announced. Always thought that they'd go with HBM2.


I was wondering the same thing unless they plan on releasing with HBM2 but it may not be mature yet.
Click to expand...

Nahh, they have no competition without Vega, and Vega is dependent on HBM2. One cannot launch without the other :D

And the longer they hold-off releasing the Titan, longer they can charge $600 for a midrange part.

See, there's no reason to do a rush job on GP102 with GDDR5x. If they've already announced GP100, then HBM2 is going to be available in quantity soon.
 
well if gp102 is coming or its going to have gddr5x, it would have to be planned well before hand, speculating it was planned for, looking at Fiji with HBM, I think they would go with HBM2 so it would preserve their perf/watt advantage since Fiji (Nano) came close to the perf/watt of the higher Maxwells.
 
zauba shipments to India are in, so they are validating gp 106 and reference boards already. So I would say a quarter away maybe less maybe a little bit more but expect end of summer fall time.
 
razor1 said:
zauba shipments to India are in, so they are validating gp 106 and reference boards already. So I would say a quarter away maybe less maybe a little bit more but expect end of summer fall time.
Click to expand...

1080 is confirmed to be GP104-400. follows naturally that 1070 is -200, waiting on mystery -150 to show up. I sincerely hope there's a 1060ti announced soon
 
thats also a possibility but I think they will let AMD announce Polaris and see if they need it right now.
 
razor1 said:
thats also a possibility but I think they will let AMD announce Polaris and see if they need it right now.
Click to expand...

I'm sure they have a pretty good idea of what it can do given it's already in a product (PS4K). Maybe, like us, they're waiting to see how aggressively they clock them and how they're priced effectively (considering stock)

It's still unclear how availability for 1070/1080 will play out
 
razor1 said:
well if gp102 is coming or its going to have gddr5x, it would have to be planned well before hand, speculating it was planned for, looking at Fiji with HBM, I think they would go with HBM2 so it would preserve their perf/watt advantage since Fiji (Nano) came close to the perf/watt of the higher Maxwells.
Click to expand...

It depends on margins. I would bet that GP102 won't be bottlenecked by 12 GBps GDDR5x so if it's a much cheaper play than HBM2 I don't see why nVidia would go with the more expensive option.

The 1080 has 43% more bandwidth than the 980 at just 10 GBps.
 
its not all about bandwidth though, its also about power draw, if they use HBM2, they can use that power to push more performance out of their GPU's, it all depends on how they approach it. And since these cards are going to be more than $600 bucks, they can use HBM2 while maintaining margins and keep their perf/watt ahead of Vega.

Once we start seeing tests of bandwidth overclocking with gp104, I think we will get an idea of what a full gp100,102 or what ever it is, would need in terms of bandwidth.
 
In the Pascal white paper nVidia didn't mention HBM power savings. They instead attributed it all to TSMC and chip architecure improvements.

I'm not convinced HBM is a significant contributor to perf/w.
 
HBM is a very large influence, the bus for GDDR memory takes up a good deal of space and it has to be clocked at the frequency of the memory.
 
Yeah, I'm not convinced of HBM1 being the ONLY driver behind the perf/watt improvements on Fiji.

Uncut Tonga manages to use 50w less at the same performance level as the old 290x:

ASUS Radeon R9 380X Strix 4GB Review

I'm sure you can peg maybe 10w of that on the loss of 128-bits of memory controller, but the rest comes form efficiency improvements. And considering the price-point of under $250, I don't think it was more efficient VRMs. That's a > 20% performance/watt improvement nothing else can account for! And those same improvements went into Fiji.

I'm sure HBM1 helped with power consumption on Fury X, but the best estimates I recall seeing on Real World Tech said they were saving maybe 25w going from fast-clocked 512-bit GDDR5 to HBM1 (IIRC, it was a 36w 512-bit GDDR5 estimate down to 12w HBM1 4GB estimate from data sheets) That's only 10% power reduction on a 250w part. So unlikely to be the BIGGEST contributor to efficiency improvement - otherwise Fury X would have used over 400w at stock!

And Nvidia is coming from just a 384-bit bus, so the power reduction would be under 20w. Starting to look less and less like the sole source of efficiency improvements.

So HBM2 would have to be CHEAPER to implement than GDDR5x, and that's not so likely this year. So Perhaps Nvidia will stay away from HBM2 for GP102.
 
Last edited:
If there's a bit more headroom with GDDR5X, I believe Nvidia might just stay with it for the Ti version and perhaps even the next Titan and just up the power consumption a little. Seeing as how GDDR5X will only just be released, I can't imagine them using it on only the 1080 and then jump to HBM1 or 2. Stranger things have happened but I'll just wait it out and see what they do.
 
If it doesn't have any advantage over GDDR 5, 5x, there would have been no reason for AMD and Hynix to spend millions of dollars to create it. HBM 1 doesn't have advantages over GDDR5x with bandwidth, why wouldn't the have just helped hynix create a GDDR5x or something similiar instead of HBM? Cause that wouldn't be logical enough, they wanted to waste money. I guess in AMD's defense that would ok with their recent history. And tonga is still way behind in Perf/watt over Maxwell, so although it was a slight improvement over other AMD chips using GDDR, it wasn't enough of an improvement. Nano got very close to maxwell's perf/watt though what does that tell ya.
 
Like I was talking the other day, those numbers look fantastic but to me, sound alot like Nvidia bullsh*t marketing numbers.
Not saying that they are not looking great, they do, but I think what most of you dont realize is that those presentation numbers are mostly on VR mode.

Of course it will offer amazing fast performance, but check this video that actually explains pretty well what I was thinking.
 
Increased memory bandwidth only provides increased performance if the GPU was being bottlenecked by lack of bandwidth in the first place. I doubt that GDDR5X is going to constrain the 1080's performance given that it should be within spitting distance of a heavily overclocked 980 Ti, which did just fine with GDDR5. So this notion that the 10x0 series is just an intermediate step to HBM2 Vega which will redefine performance on par with the 3dfx Voodoo, GeForce 256, Radeon 9700 Pro, or 8800 GTX is wishful thinking.
 
razor1 said:
If it doesn't have any advantage over GDDR 5, 5x, there would have been no reason for AMD and Hynix to spend millions of dollars to create it. HBM 1 doesn't have advantages over GDDR5x with bandwidth, why wouldn't the have just helped hynix create a GDDR5x or something similiar instead of HBM? Cause that wouldn't be logical enough, they wanted to waste money. I guess in AMD's defense that would ok with their recent history. And tonga is still way behind in Perf/watt over Maxwell, so although it was a slight improvement over other AMD chips using GDDR, it wasn't enough of an improvement. Nano got very close to maxwell's perf/watt though what does that tell ya.
Click to expand...

Because GDDR5X is only a bridge before things go full-on HBM. But for the next couple years it'll do.

And I'm pretty sure AMD thought two years go that GDDR5 was as fast as things could get, so they bet on the next big thing. They've made that mistake before, testing but ultimately dropping GDDR4. Just like Nvidia made that mistake betting the farm on GDDR2.

Sometimes you bet right. Sometimes you bet wrong.

And razor, you and I BOTH know I stuck specifically to AMD parts on the same process node to asses the efficiency improvements of GCN 1.2. Why you're bringing up the fact that it's aced by Maxwell is beyond me. I only brought it up because there is no equivalent part for Fiji
 
Last edited:
The voltage drops for HBM vs GDDR5 is very real hence the low clock speed of HBM type memory, there is no denying that, if they were only getting 25watts savings on FuryX or p100 or gp102, for using HBM 1 or 2 respectively, it would be a waste to even make it.
 
razor1 said:
The voltage drops for HBM vs GDDR5 is very real hence the low clock speed of HBM type memory, there is no denying that, if they were only getting 25watts savings on FuryX or p100 or gp102, for using HBM 1 or 2 respectively, it would be a waste to even make it.
Click to expand...

Not much lower voltage, just 1.2v versus 1.5v. That drops dynamic power by 35%. The I/O lines are much shorter, probably having a much bigger impact. The overall reduction in power is around 3x versus the 512-bit GDDR5.

But the memory interface has to be much wider to make up for it. That's 8 times as much I/O lines, so this doesn't run for free. And the memory is 4x lower density, which means higher power consumption. Splitting hairs I know, but you asked :D

Now that GDDR5X has managed to find a way to eventually double bandwidth while having almost the same voltage as HBM2, it's reducing the need to have HBM2 TODAY. But we will eventually transition to it, since bandwidth on external chips will soon hit an impassible wall.

YES, HBM1 was a waste on Fiji. That's the whole reason we did the analysis at RWT: to figure out what kind of miracle savings AMD got t out of this crappy tech demo. Answer was: under 30w :D
 
Last edited:
defaultluser said:
Because GDDR5X is only a bridge before things go full-on HBM. But for the next couple years it'll do.

And I'm pretty sure AMD thought two years go that GDDR5 was as fast as things could get, so they bet on the next big thing. They've made that mistake before, testing but ultimately dropping GDDR4. Just like Nvidia made that mistake betting the farm on GDDR2.

Sometimes you bet right. Sometimes you bet wrong.

And razor, you and I BOTH know I stuck specifically to AMD parts on the same process node to asses the efficiency improvements of GCN 1.2. Why you're bringing up the fact that it's aced by Maxwell is beyond me. I only brought it up because there is no equivalent part for Fiji
Click to expand...


The reason I bring up maxwell is because there are no other chips that are at the same level of transistors and performance. If we take a look at Tonga and Fiji, Tonga sucks up a lot of power compared to the performance it delivered compared to Fiji both in perf/mm and perf/watt (excluding FuryX of course)
 
defaultluser said:
Not much lower voltage, just 1.2v versus 1.5v. That drops dynamic power by 35%. The I/O lines are much shorter, probably having a much bigger impact. The overall reduction in power is around 3x versus the 512-bit GDDR5.

But the memory interface has to be much wider to make up for it. That's 8 times as much I/O lines, so this doesn't run for free. And the memory is 4x lower density, which means higher power consumption. Splitting hairs I know, but you asked :D

Now that GDDR5X has managed to find a way to eventually double bandwidth while having almost the same voltage as HBM2, it's reducing the need to have HBM2 TODAY. But we will eventually transition to it, since bandwidth on external chips will soon hit an impassible wall.

YES, HBM1 was a waste on Fiji. That's the whole reason we did the analysis at RWT: to figure out what kind of miracle savings AMD got t out of this crappy tech demo. Answer was: under 30w :D
Click to expand...
Can ya link me to the analysis?
 
Found it.

there was much tossing of numbers about in the thread, but eventually nobody contested this post from juanrga, sourced from AMD and Hynix.

Real World Technologies - Forums - Thread: AMD's Fiji pros and cons

The HBM on Fiji uses 12w.

The entire memory subsystem uses somewhere between 12w,to 26w guesstimate. The 12w is a hard spec sheet number for the HBM, and the rest of the wattage is a guesstimate about the GPU controller power. Charlie Demerjianis claiming 26w here:

Real World Technologies - Forums - Thread: AMD's Fiji pros and cons

The claimed efficiency factor from AMD was 3x, so it's simple math to figure out the GDDR5 power. So somewhere between 25 and 50w saved, based on their combined guesstimate.

Sorry if my memory only remembered the bottom of the range, and not the top end.
 
Last edited:
Ok went through it a bit, for the available bandwidth, HBM power reduction is substantial. Yeah ~30 watts savings for a 2x the increase in bandwidth. GDDR5X doesn't double the bandwidth without an increase in bus size also higher frequency memory well eat up more wattage too (10 vs 12 ghz). If lets say the 1080 had a 384 bit bus with 12 ghz memory, The effective bandwidth is 576 Gb/s, this is far off of the lower end HBM2 with its ability to push over 750 Gb/s-1Tb/s. To reach the bandwidth levels of HBM2, they would need a 512 bit bus, and these chips are large enough to support that (vega and p100 size)

I think you would end up seeing around ~50 maybe a bit more watts savings of HBM2 vs GDDR5x if bandwidth amounts are similar. Which would be a substantial amount of your power envelope ~25%

Saying that would there be a need for so much bandwidth (either a 512 bit bus or HBM2)? Going from 2560 shaders to ~3800 shaders, there doesn't seem to be a need for either of those two since a 384 bit bus and GDDR5x would provide I think and seems more reasonable since its only 30watts but that still is 10% of your power envelope which is high and would that be something to cover the additional cost of HBM2, don't know.
 
Last edited:
razor1 said:
Ok went through it a bit, for the available bandwidth, HBM power reduction is substantial. Yeah ~30 watts savings for a 2x the increase in bandwidth. GDDR5X doesn't double the bandwidth without an increase in bus size also higher frequency memory well eat up more wattage too (10 vs 12 ghz). If lets say the 1080 had a 384 bit bus with 12 ghz memory, The effective bandwidth is 576 Gb/s, this is far off of the lower end HBM2 with its ability to push over 750 Gb/s-1Tb/s. To reach the bandwidth levels of HBM2, they would need a 512 bit bus, and these chips are large enough to support that (vega and p100 size)

I think you would end up seeing around ~50 maybe a bit more watts savings of HBM2 vs GDDR5x if bandwidth amounts are similar. Which would be a substantial amount of your power envelope ~25%

Saying that would there be a need for so much bandwidth (either a 512 bit bus or HBM2)? Going from 2560 shaders to ~3800 shaders, there doesn't seem to be a need for either of those two since a 384 bit bus and GDDR5x would provide I think and seems more reasonable since its only 30watts but that still is 10% of your power envelope which is high and would that be something to cover the additional cost of HBM2, don't know.
Click to expand...

But what about if you downclock HBM2 and its boost to efficiency figures?
Still seems to work out the more ideal.
Will be interesting to eventually find out why HBM2 was clocked lower in the Tesla P100, considering it was already on the limits of acceptable TDP (maybe this helped).
Cheers
 
razor1 said:
I think you would end up seeing around ~50 maybe a bit more watts savings of HBM2 vs GDDR5x if bandwidth amounts are similar. Which would be a substantial amount of your power envelope ~25%.
Click to expand...

You know, just because my memory's rusty, doesn't mean you have to inflate the math here. I posteed the right numbers eventually, so now I'm going to fix your numbers :D

First off, the GP102 level card will target 250w TDP, so that drops your math down to 20%. 50/250 * 100% = 20%.

Second, if you're doing the percentage IMPROVEMENT you have to start with the UNMODIFIED power of 300w, not the target power of 250w. So 50/300 * 100%= 16% POWER REDUCTION.

Also, the power difference will be lower because we're comparing a theoretical 384-bit GDDR5X bus here (instead of the 512-bit GDDR5 bus it would have taken to power Fury X). So let's just call it more like 40w and 15% even? Again, a contributor to lower power consumption, but not the undefeated breadwinner it's played out to be. And still far more expensive to implement.

You can get the same power savings by down clocking the core maybe 5-7%, which will significantly reduce the power consumption. I mean, just look at what AMD did with the Nano: cut power by a 65w by down-clocking 10%, because the PEAK core voltage is very high. And you slightly reduce your bandwidth needs as well, so it's a double win!

So is the extra cost WORTH that extra 5-7% performance?
 
Last edited:
defaultluser said:
You know, just because my memory's rusty, doesn't mean you have to inflate the math here. I posteed the right numbers eventually, so now I'm going to fix your numbers :D

First off, the GP102 level card will target 250w TDP, so that drops your math down to 20%. 50/250 * 100% = 20%.

Second, if you're doing the percentage IMPROVEMENT you have to start with the UNMODIFIED power of 300w, not the target power of 250w. So 50/300 * 100%= 16%.

Also, the power difference will be lower because we're comparing a 384-bit bus. So let's just call it more like 40w and 15% even? Again, a contributor to lower power consumption, but not the undefeated breadwinner it's played out to be. And still far more expensive to implement.

You can get the same power savings by down clocking the core maybe 5-7%, which will significantly reduce the power consumption. I mean, just look at what AMD did with the Nano: cut power by a 65w by down-clocking 10%, because the PEAK core voltage is very high. And you slightly reduce your bandwidth needs as well, so it's a double win!

So is the extra cost WORTH that extra 5-7% performance?
Click to expand...


I'm not even talking about max board power, I'm talking about just the GPU and memory changes, nothing about GP 102 power envelope, just possible changes based on what memory they use and if they are gong to try to equalize memory bandwidth based on that type of memory to HBM2 bandwidth amounts.

If you want equal bandwidth from GDDR5x to HBM2 you need a 512 bit bus with 12Gb memory no way around that, that 30 watts on Fiji, is not the same as the case I just stated. They didn't look at what equalizing bandwidth levels would do to power usage (at the time they wrote that the fastest GDDR5 even with a 512 bit bus wouldn't get to 500GB.sec). So lets say Fiji didn't need to clock HBM at the clocks they did so they only had 350gb/sec. what will happen to the power differential than, that 30 watt difference might be 40 watts now.

I can go into many more issues that guy didn't even go into. Single strength based on trace lengths, also affect power usage. GDDR memory this an issue, the longer the trace length the more power is used to keep the single coherent. But over all we can disregard this as its a minor %, 2 or 3% maybe.

And the FuryX the design of that chip was not designed to sustain the clock speed and thus they needed that AIO system to keep the leakage at an acceptable level for the thermal envelope they wanted to get to. This is why you see with a 10% increase in core clocks it ends up eating 30% more power. There is a leakage problem, this is what happens with leakage, the chip needs more power to keep things going than it normally would. If leakage wasn't so much of an issue, you should get a fairly linear increase of frequency to wattage, didn't happen right? This has nothing to do with the memory used it has everything to do with leakage and chip design.
 
Last edited:
razor1 said:
I'm not even talking about max board power, I'm talking about just the GPU and memory changes, nothing about GP 102 power envelope, just possible changes based on what memory they use and if they are gong to try to equalize memory bandwidth based on that type of memory to HBM2 bandwidth amounts.
Click to expand...

When you're talking about replacing GDDR5 with HBM2, then we ARE talking about board power. Because the board consumes most of that power. Communication over long traces is power-hungry.

If you want equal bandwidth from GDDR5x to HBM2 you need a 512 bit bus with 12Gb memory no way around that, that 30 watts on Fiji, is not the same as the case I just stated. They didn't look at what equalizing bandwidth levels would do to power usage (at the time they wrote that the fastest GDDR5 even with a 512 bit bus wouldn't get to 500GB.sec). So lets say Fiji didn't need to clock HBM at the clocks they did so they only had 350gb/sec. what will happen to the power differential than, that 30 watt difference might be 40 watts now.
Click to expand...

Fiji didn't need HBM1 bandwidth. OR are you telling me that the 380X doesn't have enough bandwidth on that 256-bit bus? The delta color compression makes a difference, so don't just asssume w're talking about a more compute-heavy Hawaii here.

Speaking of which, the GP102 can't be more than 60-70% faster than a GTX 980 Ti, then it needs 60-70% higher bandwidth. So YES, the insane HBM2 bandwidth would be way overkill. GDDR5x 10Gt/s gets us up to 42%, and you can add additional bits and pieces through other optimizations.

I can go into many more issues that guy didn't even go into. Single strength based on trace lengths, also affect power usage. GDDR memory this an issue, the longer the trace length the more power is used to keep the single coherent. But over all we can disregard this as its a minor %, 2 or 3% maybe.

And the FuryX the design of that chip was not designed to sustain the clock speed and thus they needed that AIO system to keep the leakage at an acceptable level for the thermal envelope they wanted to get to. This is why you see with a 10% increase in core clocks it end up eating 30% more power. There is a leakage problem, this is what happens with leakage, the chip needs more power to keep things going than it normally would. That had nothing to do with the memory used.
Click to expand...

See you took the water cooler bullshit PR just like the rest of the press did.

Did you know that the Fury Nano managed to reduce power consumption by 65w while raising the temperature up to a toasty 75C? IMPOSSIBLE YOU SAY!

But possible. The leakage power consumed by the GPU is also a function of voltage. That's another reason why there's a race to zero. It's also the reason why the temperature got lost in the noise with a simple downclock/downvolt.

And let me make this PERFECTLY CLEAR: leakage affects everyone equally. It's how your power management handled it, and how you design your architecture around it that determines your performance/watt, and that is where Maxwell excelled.
 
defaultluser said:
Fiji didn't need HBM1 bandwidth. OR are you telling me that the 380X doesn't have enough bandwidth on that 256-bit bus? The delta color compression makes a difference, so don't just asssume w're talking about a more compute-heavy Hawaii here.

Speaking of which, the GP102 can't be more than 60-70% faster than a GTX 980 Ti, then it needs 60-70% higher bandwidth. So YES, the insane HBM2 bandwidth would be way overkill. GDDR5x 10Gt/s gets us up to 42%, and you can add additional bits and pieces through other optimizations.
Click to expand...

It doesn't matter if it didn't need it or not (it definitely didn't really need it), it had it.

Which I agree to that HBM2 would be over kill they really only need ~500 gb/s bandwidth of a gp102.



See you took the water cooler bullshit PR just like the rest of the press did.

Did you know that the Fury Nano managed to reduce power consumption by 65w while raising the temperature up to a toasty 75C? IMPOSSIBLE YOU SAY!
Click to expand...

There is a direct correlation of heat from power usage and leakage all three are connected.

If you overclock the nano 10% you get a 30% increase in power usage (sorry didn't mean to allude that to Fury X, it was nano that got those results)

GPU Clock Speed and PowerTune - Radeon R9 Nano Small Form Factor Overclocking Review

Pretty much what that is telling us Nano is at the limit of where leakage starts to become a problem with the cooling solution it has (even if it was a specially binned chip for lower power consumption)

But possible. The leakage power consumed by the GPU is also a function of voltage. That's another reason why there's a race to zero. It's also the reason why the temperature got lost in the noise with a simple downclock/downvolt.
Click to expand...

Leakage doesn't consume power, to overcome leakage you need to increase power so the signal propagation isn't lost over the gates.Yes there is a race to zero. Also if voltage needs to be increased there is an exponential increase in wattage *and exponential decrease if voltage is dropped
And let me make this PERFECTLY CLEAR: leakage affects everyone equally. It's how your power management handled it, and how you design your architecture around it that determines your performance/watt, and that is where Maxwell excelled.
Click to expand...

Depending on how the architecture is made, there can be a big change on how leakage affects different chips. Perfect example, Fiji and GM200, both similar size chips one that doesn't' have the issues as much as the other.
 
Last edited:
I posted this in the Doom thread, but it is probably just as relevant here:
So one real-world game performance result we have for 1080 albeit still just summary info goes back to the announcement last week.
They showed Doom obviously using beta drivers with Vulkan and the game on absolute top settings (nightmare) running between 130fps to a brief peak near 190fps.
pcgameshardware has done a recent review of Doom and the results are interesting even at 1920x1080 (which was the presentations setting as well); they used the setting below nightmare so less shadow detail.
In their test an AIB 980ti Palit Super Jetstream had a minimum 125 and average 158.7 - See OP chart, thanks Ieldra - DOOM (2016) first benchmark review

So that is a surprising result, OK the 1080 was using Vulkan but considering the optimisation benefits are yet to translate to healthy boost in DX12 for NVIDIA this still looks impressive (or at least Vulkan is working better than DX12 lol), especially as the 1080FE was on the very top setting and the 980ti was a notch below and also 1080FE on beta drivers.
Fingers crossed pcgameshardware will repeat their test once the Vulkan patch is rolled out for a better comparison.

But looks like it should raise some eyebrows regarding initial real-world result for Founders Edition 1080, anyway better information IMO than any supposed "leaked" results popping up on the internet.


Cheers
 
Last edited:
I usually wait a generation to buy new cards. But I am thinking about jumping on a 980 Ti when the 1080 Ti drops, that or a 1070.
 
heatshiver said:
I usually wait a generation to buy new cards. But I am thinking about jumping on a 980 Ti when the 1080 Ti drops, that or a 1070.
Click to expand...

Well if you get a 980Ti after the 1080Ti drops, you have waited a generation.
 
You must log in or register to reply here.
Back
Top