News [WCCFTech] AMD to reveal Navi GPU at Computex (partial rumor)

[Hitman928]

Nice story about your 2 samples...look at the curves from Toms Hardware, the GTX1060 even increases its perf/power lead when undervolting compared tp Rx480. GTX1060 is taking 50% power at 70% performance while RX480 still uses 65% power at 70% performance. This makes the GTX1060 a much better under-clocker than the RX480 - and is the main reason you find the GTX160 at a reasonable TDP in laptops without much performance regression.

It is pretty much given that both NVidia and AMD are working with the same voltage margins for shipping products and both are "undervolting" for their laptop products.

Do you have a link to the Toms article to show how they're undervolting? The graph only shows clocks and power. I asked for it before but no reply. . .

Yes my example is anecdotal, I've asked for any data sets that show Nvidia cards undervolting in the same way you can AMD cards, have yet to receive any.

Even still, once again, people are so invested in Nvidia must be superior that they're missing the point.

 

[maddie]

Nice story about your 2 samples...look at the curves from Toms Hardware, the GTX1060 even increases its perf/power lead when undervolting compared tp Rx480. GTX1060 is taking 50% power at 70% performance while RX480 still uses 65% power at 70% performance. This makes the GTX1060 a much better under-clocker than the RX480 - and is the main reason you find the GTX160 at a reasonable TDP in laptops without much performance regression.

It is pretty much given that both NVidia and AMD are working with the same voltage margins for shipping products and both are "undervolting" for their laptop products.

Who's talking about underclocking?

 

[tviceman]

Looking through those thread, pretty much all are undervolting and underclocking. Again, not the same situation. Even those that say they're at stock it's difficult to say with the way Nvidia boost works. They'd have to run actual benchmarks to know for sure.

A GTX 2080 TI at 1800-1900 mhz is not underclocking. The boost clock for the FE 2080 TI is listed as 1635mhz.

 

[ZGR]

Back when Polaris was first unveiled, it was demoed @ 1080p60 vsync running at its optimal voltage curve. It had great perf/w compared to Nvidia's 900 series in this cherrypicked scenario.

When Polaris was released, it was clocked way higher and sucked a lot more juice.


I think AMD got it right with Fury by releasing a Nano and a full size. While the Nano was a fully unlocked chip which cost less than the air cooled fury, it was advertised for its perf/w.

Nano used around 175w stock vs 275w stock for normal fury.

I think it is really important to know the voltage frequency curves of all our hardware and just how efficient (or inefficient) they can be.

AMD needs to release a GPU with excellent stock perf/w alongside their performance lineup imo.

They did something similar with the R7 1700 65w.

 

[Head1985]

As if undervolting Maxwell/Pascal isn't a thing.

I made some videos back when i bought 1080TI.
I am using 0.800v@ 1700Mhz(FE clocks) and card pulling only 150w and it is fast as 1080TI FE in reviews.So yeah pascal can be undervolted like AMD cards.
I have even manage undervolt it to 115-120w(GTX1060 TDP) and performance was in far cry 5 at GTX1080 level.

 

[Hitman928]

I made some videos back when i bought 1080TI.
I am using 0.800v@ 1700Mhz(FE clocks) and card pulling only 150w and it is fast as 1080TI FE in reviews.So yeah pascal can be undervolted like AMD cards.
I have even manage undervolt it to 115-120w(GTX1060 TDP) and performance was in far cry 5 at GTX1080 level.

1080Ti FEs are known to boost up to just shy of 1.9 GHz at stock.
https://www.youtube.com/watch?v=dG2Az0PclQA
https://forums.evga.com/1080-Ti-Clock-Speeds-m2631181.aspx

Yours is locked at 1709 MHz when undervolted. Again, very impressive increase in perf/w, but not the same situation.

 

[Hitman928]

A GTX 2080 TI at 1800-1900 mhz is not underclocking. The boost clock for the FE 2080 TI is listed as 1635mhz.

Just like the 1080Ti, Nvidia lists the boost clock as kind of a "base" boost clock. The actual boost clocks are much higher and a bit card dependent. That's why I said you'd need actual benchmarks or core clock monitoring to be sure of what clocks are actually being run. A 2080Ti boosting up to 1.9 GHz at stock settings is not uncommon so it depends on how to clock speeds are being reported, the details of which were missing from the threads.

https://forums.geforce.com/default/...rtx-20-series/2080-ti-fe-core-clock-question/

From what I've seen, it does look like Turing has the ability to undervolt without dropping clocks a bit better than previous generations, but it's hard to say as good Turing undervolt data is even harder to find.

 

[Head1985]

1080Ti FEs are known to boost up to just shy of 1.9 GHz at stock.

https://forums.evga.com/1080-Ti-Clock-Speeds-m2631181.aspx

Yours is locked at 1709 MHz when undervolted. Again, very impressive increase in perf/w, but not the same situation.

Not even close to 1900mhz if you dont mess with power limit or fan curve
https://www.anandtech.com/show/11180/the-nvidia-geforce-gtx-1080-ti-review/16

 

[Hitman928]

Not even close to 1900mhz if you dont mess with power limit or fan curve
https://www.anandtech.com/show/11180/the-nvidia-geforce-gtx-1080-ti-review/16

I don't have the card so I can only go off of the reports of others. Anandtech shows max boost at 1898 MHz but much lower average clocks during gaming. Others online (which I linked a thread for) have reported higher clocks during gaming. It is definitely system/card/game/scene dependent which makes it hard to compare across different users. In the video I linked, when overclocked he is sustaining clocks at just over 2 GHz. He doesn't show sustained clocks at stock during benchmark, but based on the score increases when he adjust the power limit and fan curve as well as the score increase after overclock, he should be running at about 1.8 GHz at stock during the benchmark.

It's the same for AMD cards as well, the only difference, that I've seen, is that AMD users have reported the exact same behavior time and time again whereas I haven't seen the same for Nvidia cards where most of the time they are sacrificing a little performance to get significant power savings. It could be just because it's more compelling for AMD owners so there's a sampling bias, I don't know. I can only work with what I see online and the experience I had between a 1060 and 480 which seemed to match the majority of experiences I've read about.

Edit:

Here is a good example of actual testing. This user (AiB model with higher than FE stock clocks) tests at stock and then benchmarks the same card with the same set core clocks while lowering voltage or raising voltage. Each time he uses a lower voltage, the benchmark score goes down because despite not touching the clocks, the card is actually boosting to lower clocks during the benchmark. This is the same behavior I saw with the 1060 and mirrors pretty much every test case I've seen that has bothered to check clock/performance changes with undervolt though obviously there may be other examples, I just haven't seen them and for sure does it does not reflect the vast majority of examples I've seen.

https://forums.evga.com/FindPost/2811013
https://forums.evga.com/FindPost/2811126

 

[Yotsugi]

Each time he uses a lower voltage, the benchmark score goes down because despite not touching the clocks, the card is actually boosting to lower clocks during the benchmark

Boost 3.0 is a double-edged sword, yes.

 

[Markfw]

Using Xprecision, my 1080TI's run at 1936 or more AT STOCK:

Or am I reading this wrong ?

And GPUZ to confirm

 

[coercitiv]

It's the same for AMD cards as well, the only difference, that I've seen, is that AMD users have reported the exact same behavior time and time again whereas I haven't seen the same for Nvidia cards where most of the time they are sacrificing a little performance to get significant power savings. It could be just because it's more compelling for AMD owners so there's a sampling bias, I don't know. I can only work with what I see online and the experience I had between a 1060 and 480 which seemed to match the majority of experiences I've read about.

The same happens now with both AMDs Vega iterations - people who play around with Vega's voltage curve know it tries to adjust clocks based on voltage (within some thresholds ofc, and you can still push clocks further). AMD had mechanics in place to do this ever since Polaris came out, although based on my experience with both Polaris and Vega it looks to me like AVFS implementation on Polaris was more conservative (and/or less transparent).

The fact that Nvidia cards do this as well shouldn't surprise anyone, as the technique allows for a drastic reduction in the voltage guard band required for safe operation.

This tech is likely also responsible for the more impressing undervolts we read about (from both camps), as it manages to keep the card stable through small workload segments where the thinner voltage guard band would normally be insufficient.

 

[DrMrLordX]

The same happens now with both AMDs Vega iterations - people who play around with Vega's voltage curve know it tries to adjust clocks based on voltage

My Vega FE would increase clocks as I lowered voltage. You had to lower Vega's voltage to below 1.0v to get it to actually drop clocks. My sample was so volt-happy that it would crash at any voltage below 1.1v so it was impossible for me to actually lower clocks by lowering voltage (power limit was an entirely different story). The funny thing is that power usage would not drop by much, thanks to AVFS trying to boost to higher clocks from the reduction in temperature (from lowering voltage). AVFS can be pretty weird sometimes.

My Radeon VII will not change clocks at all from reductions in voltage. I can drop volts all the way down to .990v and it exhibits no change in performance at all. Any voltage lower than that and it begins to artifact @ stock speeds.

So no, AMD cards do not necessarily behave the same way as NV cards when reducing voltage.

 

[coercitiv]

My Vega FE would increase clocks as I lowered voltage. You had to lower Vega's voltage to below 1.0v to get it to actually drop clocks. My sample was so volt-happy that it would crash at any voltage below 1.1v so it was impossible for me to actually lower clocks by lowering voltage (power limit was an entirely different story). The funny thing is that power usage would not drop by much, thanks to AVFS trying to boost to higher clocks from the reduction in temperature (from lowering voltage). AVFS can be pretty weird sometimes.

My Radeon VII will not change clocks at all from reductions in voltage. I can drop volts all the way down to .990v and it exhibits no change in performance at all. Any voltage lower than that and it begins to artifact @ stock speeds.

So no, AMD cards do not necessarily behave the same way as NV cards when reducing voltage.

Here's what Vega 56 does under a constant compute load and 1450Mhz target clock:

1050mV -> 1406Mhz
1000mV -> 1419Mhz (example of what you experinced with your Vega FE)
950mv -> 1412Mhz
925mV -> 1405Mhz
900mV -> 1387Mhz (example of what I'm talking about)

The values tend to fluctuate based on chip quality, power , temps & fan settings etc. but even this short test is enough to prove a point.

Spoiler: Wattman & HwInfo Screenshots

 

[DrMrLordX]

Here's what Vega 56 does under a constant compute load and 1450Mhz target clock:

Seems about right. Radeon VII is different though. NV cards seem to scale downward in clockspeed similarly to Vega56 @ .925v .

 

[IntelUser2000]

Even still, once again, people are so invested in Nvidia must be superior that they're missing the point.

I think some of the advantages are hidden because they are much more locked down. With AMD you can do all sorts of changes. So they are actually pretty decent once you use 3rd party programs to avoid Wattman's weird behavior and do undervolting.

 

[Ottonomous]

Remember this?

https://gfxbench.com/compare.jsp?be...&hwtype2=GPU&hwname2=AMD+Radeon+RX+580+Series

This PCI ID does indicate Vega 20 but there's a possibility it was a forced driver installation for immature silicon. Its Gen 10 and has 10-15% performance on the 580 using only 20 CUs

In the miniscule chance that its true, it would mean they've achieved performance parity all the way up to the 2080

 

[Yotsugi]

it would mean they've achieved performance parity all the way up to the 2080

Depends on power targets, SP count and such.
Its never perfect scaling going up for any vendor.

 

[Thala]

Who's talking about underclocking?

Thats because frequency-voltage curves are strongly monotonic. There is no lower voltage, where the circuit can operate faster - assuming there are no thermal constraints. Under thermal or power constraints this could of course happen.