AMD: Speculation, Rumors, and Discussion (Archive)

[Razor1]

My take was 2x perf/watt compared to Fiji and 2.5 compared to others. Fiji is within the range of efficiency of maxwell. To me this makes the most sense. Tonga isn't really more efficient than other gcn cards. The 390x just has poor efficiency due to the clocks.

Spoiler

2.5x 270x would make sense because they are the same class of chips.

did you factor in the different ram and the AIO?

AMD's most efficient card is the Nano, then its Tonga and AMD CEO/CFO definitely stated they were 2.0 perf/watt increases from today's midrange. That doesn't leave much room to speculate, its either the 280x or the 380x, and the 380x is what I used to come up with numbers is way ahead of the 280x.

Here ya go.

We remain on track to introduce our new 14 nanometer FinFET-based Polaris GPUs midyear. Polaris delivers double the performance per watt of our current mainstream offerings, which we believe provides us with significant opportunities to gain share.

http://finance.yahoo.com/news/edited-transcript-amd-earnings-conference-051600440.html

Yeah I know wtf tech took what was stated by AMD when they saw the 950 and polaris show and made the talk about Maxwell, but that was just wtf tech not understanding what they saw.

 

[Orion]

well lets take 290x as the base line, you will get what with Polaris at 2.5 times perf/watt over 290x?

100 watts at 290x performance level? Is that what it is? yes, at 150 watts you get Fury X level performance then if its perfect scaling frequency to wattage? Yes, where does that leave it when talking about Pascal? Less than Pascal.

I don't really care how its cut, sliced, mashed what ever, you just don't get the best case number provided by AMD to match up with what we know Pascal is doing right now and these are independent tests not best case.

Now you take Tonga, 2.0 times perf/watt. 95 watts for Tonga (380x) performance. Guess what, its damn close to what that 2.5 times pef/watt is to Hawaii.

Not sure if setup bound, but in some dx12 games fury x has been from within 10% to matching 1080 framerate.

Btw, I've heard that in endurance, a few tens of minutes alone can cause significant throttling in founder 1080 if you don't change fan and power settings.

 

[Razor1]

Not sure if setup bound, but in some dx12 games fury x has been from within 10% to matching 1080 framerate.

Btw, I've heard that in endurance, a few tens of minutes alone can cause significant throttling in founder 1080 if you don't change fan and power settings.

What is the differential of those games vs the 980ti? And in DX11 vs the 980ti? most likely you will a large advantage for those games on AMD hardware even in DX11, and you won't see the normal CPU overhead AMD drivers exhibit on typical DX11 gaming. So if AMD has been using a game that heavily favors their architecture to begin with to get their perf/watt numbers I would be extremely skeptical of what happens overall then. Which I haven't been talking about yet nor think its something we can figure out anyways.

 

[3dcgi]

I wonder how you came to this conclusion?

Clock gating is something that happens at the cycle level. One cycle the clock is on, the other it's off. There is no way for us to know who does one better than the other.

There are multiple levels to clock getting. The higher levels save more clock tree power and may take multiple cycles to start and stop depending on how much logic the clock tree drives. It would likely be difficult to measure the effectiveness of an implementation.

 

[silent_guy]

There are multiple levels to clock getting. The higher levels save more clock tree power and may take multiple cycles to start and stop depending on how much logic the clock tree drives.

Of course.

It would likely be difficult to measure the effectiveness of an implementation.

That's really the main point. It's completely impossible to have any observability about clock gating because large parts of it happen at such fine granularity. I'm guessing that tunafish was using clock speed changes (which are observable) as a proxy for clock gating, though it's really completely unrelated.

 

[hoom]

I've heard that in endurance, a few tens of minutes alone can cause significant throttling

NV caught cheating benchmarks again?!

There was a graph in one of the 1080 preview type articles (that I can't seem to find right now) showing clock over time quickly dropping back which which seemed wonky at the time but I couldn't figure out quite why...

 

[Pressure]

Wasn't it hardocp? Clocks dropped sharply after 5 minutes. And not under power virus conditions.

 

[Clukos]

Wasn't it hardocp? Clocks dropped sharply after 5 minutes. And not under power virus conditions.

My guess, after a few minutes temperatures rise and the card is battling with reducing core voltage and clock speeds to reduce the heat output, pretty sure a more robust cooling part (aftermarket or wc) will solve all of these issues and provide headroom for overclocking and more stable clocks in general (if p. limit isn't an issue). We've seen the same thing with Fiji and Hawaii parts from AMD.

My second guess is that Nvidia has clocked the 1080/1070 a bit more aggressively this time around (in comparison to Maxwell, which can easily hit 1.5ghz stable core clock at stock voltage) to impress people more (~20% + from a stock Titan X) at the cost of inconsistent clocks and increased temps. Judging by the fluctuation in the core clock i would assume they could keep the card at 1.5ghz core clock without issues even with the reference cooler, but that would be less impressive than the 1.8ghz that the card is regularly hitting (at least initially) right now.

 

[Pixel]

I remember hearing from one youtube reviewer who said an Nvidia rep told him that the mosfets on the Founders edition cards are not designed for sustained voltage increases. So yeah you can manually increase the threshold of temps and voltage so it doesn't throttle back down to ~1650mhz you might risk killing your card.


Let me try and find the video, I might be wrong about a few details

 

[hoom]

Wasn't it hardocp?

Unlikely, their white text on black kills my eyes.

But this page does show it in action http://hardocp.com/article/2016/05/17/nvidia_geforce_gtx_1080_founders_edition_review/5
On the other hand this is always staying above base clock & nearly all above the quoted boost clock so can't really call it shenanigans.

 

[SimBy]

So NDA lifts exactly mid year.

 

[Deleted member 13524]

http://www.24liveblog.com/live/1331452

(picture seems to have been removed in the meantime)


Whaaaaaaaaaaaaaat?
1 month of NDA? Is it for specs talk or full review?

 

[Love_In_Rio]

http://www.24liveblog.com/live/1331452

(picture seems to have been removed in the meantime)


Whaaaaaaaaaaaaaat?
1 month of NDA? Is it for specs talk or full review?

Polaris is starting to smell like a turd.We can get the new R600 instead what was thought to be the new R300.

 

[3dilettante]

Polaris hasn't really set its sights high enough to produce a pratfall like the R600 launch.
The existence of an NDA period seems architecturally neutral.

 

[Love_In_Rio]

Polaris hasn't really set its sights high enough to produce a pratfall like the R600 launch.
The existence of an NDA period seems architecturally neutral.

doent inspire a lot of confidence either.

 

[Infinisearch]

What do you guys think about this article? http://hardocp.com/article/2016/05/27/from_ati_to_amd_back_journey_in_futility#.V0huzEwzqCg

 

[eastmen]

What do you guys think about this article? http://hardocp.com/article/2016/05/27/from_ati_to_amd_back_journey_in_futility#.V0huzEwzqCg

We will know if its true soon I guess. although Its been known for awhile that Polaris was lower end and the rumors coming out is that it is comparable to the old high end from amd just not NVidia's new high end.

The article itself just seems like a fanboy lashing out at amd. Even in comments he is claiming that AMD offered to send him and his wife out to San fran to be briefed and refused.

Hardocp has a long history of favoring Nvidia all the way back to the FX days and then there is the founder edition rant

NVIDIA either breaks even or loses money on NVTTM/reference cards sold into the market.

I also don't see AMD selling their one ace in the hole to intel. If intel has ATI and AMD doesn't then AMD is gone. There is nothing left for the company. But with AMD they can make attractive APUs and Graphics chips

 

[Kaotik]

How possible scenario would Intel licensing RTG technologies for their CPUs be?

 

[3dilettante]

What do you guys think about this article? http://hardocp.com/article/2016/05/27/from_ati_to_amd_back_journey_in_futility#.V0huzEwzqCg

It seems rather hyperbolic in parts about how badly Polaris can fail.
I can't speak to whatever source he has about the Intel factor.

The possibility of breaking the graphics portion off doesn't seem impossible. It's one of the scenarios I spitballed in the AMD gloom thread as a way to satisfy creditors. The interpersonal drama might not be necessary in that scenario. AMD's finances have deteriorated to a point that there was speculation it might have tripped a covenant in its debt deals that might have forced a reorg.

The idea that it's a graphics mutiny and that AMD's money-sink of a CPU division is being left in the lurch may not be correct. The regression in AMD's "Fusion" concept seems more mutual, given that Zen did not go for an APU, AMD's HPC APU isn't an APU, loss of leadership dedicated to HSA, and AMD's goals for dis-integration with interposers. I've seen some statements on the net that also seem to be retconning HSA into being more concerned about discretes.

 

[Deleted member 13524]

What do you guys think about this article? http://hardocp.com/article/2016/05/27/from_ati_to_amd_back_journey_in_futility#.V0huzEwzqCg

The only thing to take from that article is that Kyle Bennett behaves like an 8-year-old throwing tantrums and he uses his website to do so.
He wasn't invited to the Macau event, so he's trying his best to damage AMD.

In the end, I'm only sorry that I gave that site a click for such an article.