Perf/watt/IHV man hours/posts *split*

[Deleted member 13524]

Oddly enough if you factor out the max perf/watt changes for the node, you get a worse Polaris over Hawaii when you look at perf/watt. Outside of the front end changes, Polaris really was just the node.

No. Polaris 10 is vastly superior in geometry performance to Hawaii, not to mention it gets similar performance on ~60% of Hawaii's theoretical fillrate and memory bandwidth.

Yeah having Doom running on Pascal in nV's Pascal presentation, is not the same thing as sending Pascal to the developers and having them code a path for it.

Yeah sure, nvidia just brought a couple of id devs to their headquarters and put the Doom demo up and running within a couple of hours on their brand new hardware.
That's totally the amount of effort nvidia and id software put forth for their first ever public Vulkan demo on both companies.

 

[Razor1]

No. Polaris 10 is vastly superior in geometry performance to Hawaii, not to mention it gets similar performance on ~60% of Hawaii's theoretical fillrate and memory bandwidth.

That's what I stated, outside of the front end changes........ :/

Yeah sure, nvidia just brought a couple of id devs to their headquarters and put the Doom demo up and running within a couple of hours on their brand new hardware.
That's totally the amount of effort nvidia and id software put forth for their first ever public Vulkan demo on both companies.

According to Id they didn't, I don't know should I believe you on this matter or Id? Also so far with the Doom updated Vulkan reviews, all the changes in performance from nV cards have been due to drivers, since the game hasn't been patched again from the Vulkan patch release. So guess what, looks like nV decided to do what ever they had to do in their drivers, at least for the time being.

 

[Deleted member 13524]

You claimed "if you factor out the max perf/watt changes for the node, you get a worse Polaris over Hawaii when you look at perf/watt."
It's simply untrue (and a bit ridiculously so, to be honest), regardless of the following sentence.

Feel free to create a thread called "I think if it wasn't for node changes, Polaris would get worse perf/watt than Hawaii", though. I'm not contributing to that discussion in the Vega thread anymore.

 

[ImSpartacus]

Take out the 60% savings of the perf/watt for the node and see what ya get. That is what the max theoretically would be, if per transistor Polaris is worse in performance than Hawaii/Granada, which it is, do I have to pull out the calculator or can you do it? Its pretty simple you should be able to do in your head too because transistor amounts for Polaris and Hawaii/Granada are close and Polaris overall performance wsie only hits the lowest end Granada and highest end Hawaii, you don't even need to look at the 60% savings to figure it out.

That 60% figure seems too good to be true, but that's roughly what Anandtech reported in a table of advertised benefits of various processes. The article is about glofo's 7nm roadmap, but it compiles lots of other goodies.

http://www.anandtech.com/show/10704/globalfoundries-updates-roadmap-7-nm-in-2h-2018

And now I'm wondering where they found all of those, lol.

 

[Deleted member 2197]

I believe both AMD and Nvidia include updated versions of the Vulkan API with their driver releases. Initial Nvidia Doom drivers were using a much older Vulkan API then AMD's Doom driver. The API for both has since been updated and as of now the latest drivers from both AMD and Nvidia are using the 1.0.26.0 Vulkan runtime redistributable.

 

[monstercameron]

Curious is there a source for the ID devs saying that they never tested on Pascal -or Nvidia for that matter?

 

[Silent_Buddha]

What is weird is how AMD kept talking about Polaris as the largest step in many years through its marketing campaign (improved IPC, improved geometry, improved color compression, etc.), whereas Vega kept having only "HBM2" in its description. To me, this suggested that Polaris would be the largest architectural leap since GCN1 and Vega would be little more than larger Polaris + HBM2.

It could have just been marketing talking about the transition from 28 nm, which GCN had been stuck on for ages, to 14 nm. That would result in the "largest step in many years." They also touted other architectural changes that were supposed to help with power consumption which would have been another large step compared to previous GCN incarnations.

Pushing the chip so hard on the voltage front, however, wiped out virtually any potential power savings from architectural changes. As well, there's no way of knowing just how good GloFo's 14 nm node is. Considering AMD's willingness to pay them a LOT of money to avoid producing their GPUs there, I'd say that could also have contributed to wiping out any gains from architectural changes.

Regards,
SB

 

[Razor1]

Curious is there a source for the ID devs saying that they never tested on Pascal -or Nvidia for that matter?

Gotta find it was in one of the interviews, right now Pascal runs the same path as Maxwell. AMD has a separate path with AMD intrinsics. Now nV Vulkan drivers if updated with nV intrinsics that should be all they need, but there no more information on the matter. But if those drivers aren't available the developer can't put those features in cause there is no way to test them out.

 

[CSI PC]

It could have just been marketing talking about the transition from 28 nm, which GCN had been stuck on for ages, to 14 nm. That would result in the "largest step in many years." They also touted other architectural changes that were supposed to help with power consumption which would have been another large step compared to previous GCN incarnations.

Pushing the chip so hard on the voltage front, however, wiped out virtually any potential power savings from architectural changes. As well, there's no way of knowing just how good GloFo's 14 nm node is. Considering AMD's willingness to pay them a LOT of money to avoid producing their GPUs there, I'd say that could also have contributed to wiping out any gains from architectural changes.

Regards,
SB

By default the voltage is within 5% of Nvidia's Pascal; 1.1V against 1.15V.
1.15V is enough to maintain 1266MHz, and some report they can undervolt that a little bit and still hit the 1266MHz.
The ideal voltage for this 14/16nm silicon-node is slightly higher than the 28nm with regards to GPU and some FPGA/SoCs.

The issue is not necessarily that they are too high but the relationship they have for their voltage against frequency and the silicon-node.
Worth noting that Polaris-GF is more resilient than Pascal-TSMC in terms of higher stable-tolerant voltages, which may explain the 1.1V against 1.15V
Reputable extreme OCers removing thermal constraints and all protections suggest the limit is around 1.3V for Pascal-TSMC and 1.4V to 1.45V for Polaris-GF, of course these are limits at the very fringe of specialist benchmark use-risk.
Cheers

 

[ImSpartacus]

Pushing the chip so hard on the voltage front, however, wiped out virtually any potential power savings from architectural changes. As well, there's no way of knowing just how good GloFo's 14 nm node is. Considering AMD's willingness to pay them a LOT of money to avoid producing their GPUs there, I'd say that could also have contributed to wiping out any gains from architectural changes.

Yeah, what's up with that amendment to allow amd to fab their stuff elsewhere (and still pay glofo)?

I don't know fabs well, but I was wondering if Polaris was so disappointing that amd wanted to move Vega to tsmc. Or is that just out of the question if Vega already taped out?

Maybe it is a Navi thing? Or some cpu (zen?)?

It just seems weird that they would continue to pay glofo without utilizing that capacity. I never quite wrapped my head around why that would make business sense for amd.

For others that missed this, check out news of the deal here:

http://www.anandtech.com/show/10631/amd-amends-globalfoundries-wafer-supply-agreement-through-2020

 

[Silent_Buddha]

Yeah, what's up with that amendment to allow amd to fab their stuff elsewhere (and still pay glofo)?

I don't know fabs well, but I was wondering if Polaris was so disappointing that amd wanted to move Vega to tsmc. Or is that just out of the question if Vega already taped out?

Maybe it is a Navi thing? Or some cpu (zen?)?

It just seems weird that they would continue to pay glofo without utilizing that capacity. I never quite wrapped my head around why that would make business sense for amd.

For others that missed this, check out news of the deal here:

http://www.anandtech.com/show/10631/amd-amends-globalfoundries-wafer-supply-agreement-through-2020

There's a good discussion about the amendment recently in the AMD thread in the Industry forum. Basically boils down to - AMD has obligations they have to meet with regards to GloFo. And GloFo aren't particularly keen on being generous when they kind of got suckered into buying the foundry.

Regards,
SB

 

[Silent_Buddha]

By default the voltage is within 5% of Nvidia's Pascal; 1.1V against 1.15V.
1.15V is enough to maintain 1266MHz, and some report they can undervolt that a little bit and still hit the 1266MHz.
The ideal voltage for this 14/16nm silicon-node is slightly higher than the 28nm with regards to GPU and some FPGA/SoCs.

You can't compare the AMD solution to the NVidia solution WRT voltages as they are on 2 completely different processes.

What you can do is look at the power curve and see which solution is operating in a safe or pragmatic zone (at or below the knee of the power curve) and which is pushing their silicon beyond the knee of the power curve. Nvidia isn't and AMD is. That makes AMDs power consumption significantly worse than it needs to be. However, AMD felt that they need to hit X performance level and thus sacrificed perf/watt by pushing the silicon beyond what would be considered a good voltage and frequency level. It's the same thing they did with Fury X.

Regards,
SB

 

[seahawk]

I am still not seeing how this makes the situation better. So AMD has the choice to lack in performance or perfromance/w.

And even undervolted Polaris still makes use of the power limit and gains only ~5% performance, which is about the same when you undervolt a 1060.

http://www.hardwareluxx.de/index.ph...39615-amd-radeon-rx-480-im-test.html?start=27

I stand by my point that limit to 120W RX480 stands no chance against the 1060 and the difference in perf/w would not look much better. Maybe somebody can do some benchmarks with a RX480 limited to 120W.

 

[Esrever]

Polaris 10's best efficiency performance is probably slightly less than a 470 on 100W or so. AMD would be hard pressed to sell a 480 with that performance and a 470 with even less. They wouldn't even hit the minimal VR requirement.

 

[CSI PC]

You can't compare the AMD solution to the NVidia solution WRT voltages as they are on 2 completely different processes.

What you can do is look at the power curve and see which solution is operating in a safe or pragmatic zone (at or below the knee of the power curve) and which is pushing their silicon beyond the knee of the power curve. Nvidia isn't and AMD is. That makes AMDs power consumption significantly worse than it needs to be. However, AMD felt that they need to hit X performance level and thus sacrificed perf/watt by pushing the silicon beyond what would be considered a good voltage and frequency level. It's the same thing they did with Fury X.

Regards,
SB

Err yes you can if you see the envelope spec (voltage-frequency-performance) for the silicon-node as produced by some of the manufacturers.
How are you deciding 1.15V is too high if you yourself are not basing the figure on any actual silicon performance envelope for products on 14/16nm (I am basing this off actual engineer product notes and can see the difference between this generation and the previous) or analysis looking at the frequency-voltage-power across the whole envelope for the actual GPU?
You do know Polaris-GF is perfectly linear up to 1.15V (from a total power curve-frequency perspective) while Pascal-TSMC starts to break that rule just above 1.05V (Pascal's limit of 1.1V actually is in the envelope zone of non-linear behaviour)?

Both are set to the point where they are at the optimal envelope use, if you want to go above 1266MHz and the 1.15V it breaks down fast (same goes for Nvidia above 1.1V), there are several parameters that are causing AMD challenges to get the most out of the silicon and also from an overall thermal-efficiency perspective.
But as mentioned earlier Polaris-GF silicon is more tolerant of higher voltages, as proven by extreme OCers in the real world once you remove all safeguards.
Anyway Polaris-GF is linear between 0.8V and 1.15V, they could reduce the upper voltage and indirectly frequency but it would not influence that linear performance relationship and the optimum voltage-frequency-performance, but would help coping with thermals/dynamic-static power/waste energy.
However many custom AIB Polaris cards reviewed are OC'd above the 1266MHz envelope region so it is a bit academic.
Cheers

 

[DavidGraham]

Polaris can be made more power efficient by undervolting and downclocking, but so can the 1060, it is already power efficient, and can be made even more so by the same undervolting and downclocking.

Tomshardware experimented with the concept here:

GTX 1060 requires 61W to achieve 130 FPS at 1920x1080, whereas the AMD Radeon RX 480 needs 139W for the same frame rate in spite of its adjusted voltage.

In order to average 90 FPS at 2560x1440, the GeForce GTX 1060 draws just under 62W, whereas the Radeon RX 480 comes in at 146W.

And:

Lowering the Radeon RX 480’s power consumption results in significantly worse frame times though, whereas the GeForce GTX 1060’s do improve a bit with overclocking. This means that those who want to save power and still game at 1920x1080 without perceptible drawbacks shouldn’t go below 90W.

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-1060-pascal,4679-7.html

 

[CSI PC]

Polaris can be made more power efficient by undervolting and downclocking, but so can the 1060, it is already power efficient, and can be made even more so by the same undervolting and downclocking.

Tomshardware experimented with the concept here:



And:



http://www.tomshardware.com/reviews/nvidia-geforce-gtx-1060-pascal,4679-7.html

Yeah I touched on that before when mentioning Pascal 1060 can achieve its base clock just with 61W earlier in a couple of threads and highlights part of the efficiency of their design.
However importantly that chart relative to fps is particular to a game and I am not keen trying to correlate fps to watts as this figure can depend upon how well a game is optimised for a particular IHV, meaning it can be skewed, ok for a rough idea but not too much more.
Metro Last Light is originally used by Tom's Hardware at 4k because from their experience it was equally power demanding for AMD and Nvidia cards, rather than fps perspective.

You also should consider the linear behaviour of voltage-frequency-power, AMD maintains its linear performance window while 1060 starts to break down at its peak, meaning both are set to about the right parameters in terms of optimal peak voltage-frequency-performance.
For RX480 the just under 900MHz is 0.8V and the 1266MHz is 1.12 to 1.15V.

Cheers

 

[Anarchist4000]

I don't know fabs well, but I was wondering if Polaris was so disappointing that amd wanted to move Vega to tsmc. Or is that just out of the question if Vega already taped out?
...
It just seems weird that they would continue to pay glofo without utilizing that capacity. I never quite wrapped my head around why that would make business sense for amd.

They have Zen and other products to worry about as well. That 480 running 1475MHz@1.17V and using less power would seem a strong indication there is quite a bit of variation in the process or something not working correctly. Assuming no respins or chip changes, that's 10-15% higher clocks and less power, which obviously moves the curves a bit. So it may have very well been worth it to be competitive.

 

[CSI PC]

...
They have Zen and other products to worry about as well. That 480 running 1475MHz@1.17V and using less power would seem a strong indication there is quite a bit of variation in the process or something not working correctly. Assuming no respins or chip changes, that's 10-15% higher clocks and less power, which obviously moves the curves a bit. So it may have very well been worth it to be competitive.

Regarding the 1475 at 1.17V, reading XFX they are using a specific tuned BIOS for that specific OC model, not available for any of their other Polaris models.
I would wait and see how many other models and cards can maintain that, because there is a person on another OC forum who has shown their 1080 or 1070 running at just under 2300MHz on air, I really doubt anyone else will ever hit that and no-one else could at the time, that also was using a certain BIOS that did nothing for other models from same or other AIB partners.
Cheers

 

[Anarchist4000]

Regarding the 1475 at 1.17V, reading XFX they are using a specific tuned BIOS for that specific OC model, not available for any of their other Polaris models.
I would wait and see how many other models and cards can maintain that, because there is a person on another OC forum who has shown their 1080 or 1070 running at just under 2300MHz on air, I really doubt anyone else will ever hit that and no-one else could at the time, that also was using a certain BIOS that did nothing for other models from same or other AIB partners.
Cheers

Maybe it is a perfect sample, but just the fact that it exists shows how much variation there is in the process. It might not even be a good sample, just better than most as it only takes one transistor to pull down the others. Judging by how much AMD was willing to pay GF, I'd imagine there is an issue.