AMD's Tonga - R9 285 (Specs) and R9 285X (Partial Specs)

[RussianSensation]

Your link is the power consumption for one game. You can't draw final conclusions from one game.
It doesn't work like that. Its a crazy assertion. Surely AMD methods in finding the 7970 tdp differed from nvidia and kepler but in no case would they base their figures from simply loading a save point in a single game (crysis 2).

No one said anything about using Crysis 2 as end all be all game for power usage. However, if you go browse 30 reviews of the cards outlined below in games, you will see that all of them consume different average and peak power. TDP actually has more to do with guidance for cooling/heatsink design due to heat dissipation. TDP != power consumption.

HD6970 = 250W
R9 280 = 250W
HD7970 = 250W
HD7970GE = 250W
780 = 250W
780Ti = 250W

All of these cards use a completely different amount of power but all have the same TDP rating.

7970 and 6970 are 190-200W cards and 780 is a 220W not a 250W card, while 780Ti easily exceeds 250W:
http://www.techpowerup.com/mobile/reviews/NVIDIA/GeForce_GTX_780_Ti/25.html

Stand back and just think about it: are you going to claim now that an R9 280 uses up to 250W but 7970Ghz also uses 250W? The TDP ratings for AMD and NV tell us little about the card's real world power usage in games.

R9 280 is barely higher clocked than 7950 but has a 250W TDP. The real world power usage of cards such as 280 or 7950 or 7970 is far below 250W.

Look at the TDP of 280X vs. 770. You would think that 770 uses way less power in games but it's not even remotely true.

http://techreport.com/review/25466/amd-radeon-r9-280x-and-270x-graphics-cards/10

Or

http://www.anandtech.com/show/6994/nvidia-geforce-gtx-770-review/16

In fact, the variance in power usage among 280X models alone shows how worthless the TDP is for gauging real world power consumption in games:

http://www.techspot.com/review/841-radeon-r9-280x-roundup/page11.html

There are countless cases of cards which use much less power than their TDP, about the same power and way more power. The fact that AMD and NV even define TDP differently makes it even more pointless to compare them.

Then we get to the part where after-market cards use better cooling, digital power delivery and overall more efficient components. As a result, you can have an after-market card that uses less power than the reference design. Alternatively, the board could be designed specifically to support high overclocking (Classified/Lightning/Matrix/Vapor-X) and these tend to use more power than the reference design. Unless you test a specific card in question, often times the TDP rating doesn't align at all.

First it took gamers a long time to accept that Furmark was a waste of time for testing a GPU's real world power usage but using TDP to mean power consumption is another one of those old errors/myths that long needs to go away. We have tools that allow us to measure the card's real world power usage where it's no longer relevant to look at some arbitrary number on the box.

 

[redzo]

First it took gamers a long time to accept that Furmark was a waste of time for testing a GPU's real world power usage but using TDP to mean power consumption is another one of those old errors/myths that long needs to go away. We have tools that allow us to measure the card's real world power usage where it's no longer relevant to look at some arbitrary number on the box.

I believe that furmark is suitable for worst case scenarios. Worst case scenarios begin to make sense in the long run. For example, as games become more demanding, the gpu will be loaded at almost 100%. That's where furmark tests begin to make sense: when you're stuck with the same GPU for a few years. I always take into account max energy consumption. While simulating 100% load with furmark may seem artificial at launch, in the long run I think is important.

 

[Erenhardt]

I believe that furmark is suitable for worse case scenarios. Worse case scenarios begin to make sense in the long run. For example, as games become more demanding, the gpu will be loaded at almost 100%. That's where furmark tests begin to make sense: when you're stuck with the same GPU for a few years. I always take into account max energy consumption. While simulating 100% load with furmark may seem artificial at launch, in the long run I think is important.

You see... That is exacly why he wrote that.
Furmark doesn't peg GPU to 100%. Its known that (nv) drivers recognize furmark and throttle cards a bit. Ploting performance/watt graphs based on performance in games and power consumption in furmark is misleading.

 

[3DVagabond]

You see... That is exacly why he wrote that.
Furmark doesn't peg GPU to 100%. Its known that (nv) drivers recognize furmark and throttle cards a bit. Ploting performance/watt graphs based on performance in games and power consumption in furmark is misleading.

^^This. The IHV's target the .exe and throttle the cards. It means nothing anymore.

 

[redzo]

You see... That is exacly why he wrote that.
Furmark doesn't peg GPU to 100%. Its known that (nv) drivers recognize furmark and throttle cards a bit. Ploting performance/watt graphs based on performance in games and power consumption in furmark is misleading.

Thought he was stating that energy consumption at max load is not that important and that somehow it has to go away. Didn't know about nv thorttling. Still some objective method of simulating max load across both brands has to be used as I find energy consumption at max load an important metric for testing a GPU card.

 

[boxleitnerb]

I believe AMD drivers also throttle Furmark, for example here (290W under gaming load and 300W - PT limit) under Furmark load):
http://ht4u.net/reviews/2013/amd_radeon_r9_290_-_hawaii_pro_im_test/index16.php
Last paragraph, no graph.

Anyway, best indicator for Perf/W is the 3DC efficiency index.
The first link shows the averaged power numbers and a performance index, both based on multiple reviews. Both parameters are taken and averaged from 5+ reviews that measure graphics card power consumption only (not the whole system but using a specially modified mainboard/testing equipment). So for example they take the TPU power value and the respective TPU performance rating, then the hardware.fr power value and the hardware.fr performance rating and so forth:
http://www.3dcenter.org/artikel/stromverbrauch-aktueller-und-vergangener-grafikkarten

Then both values are combined for each card, creating an index that you can see in the market overview, which can be found in the second link below. Precisely, it is the value
"Perf./Verbr.: x,xx" (engl.: performance over power consumption), for example 2.11 for the GTX 780 Ti (530% performance index divided by 251W). Higher is better.
http://www.3dcenter.org/artikel/grafikkarten-marktueberblick-august-2014
Don't be fooled by the percentage numbers for the performance index, the 100% baseline is a lower end card, hence current cards are listed as 500+%.

I've mentioned this site and index multiple times in the past, yet people still cling to a single site like TPU or TDP numbers. I don't know why the most complete overview of efficiency results in the whole web is ignored on a continuing basis...

 

[ShintaiDK]

You see... That is exacly why he wrote that.
Furmark doesn't peg GPU to 100%. Its known that (nv) drivers recognize furmark and throttle cards a bit. Ploting performance/watt graphs based on performance in games and power consumption in furmark is misleading.

Really? This seems to show otherwise. No throttle, full load.

 

[antihelten]

Really? This seems to show otherwise. No throttle, full load.

Isn't that example running with a dynamic overclock though (increased TDP limit)?

 

[ShintaiDK]

Isn't that example running with a dynamic overclock though (increased TDP limit)?

Its completely stock with reference design.

 

[Skurge]

Is it normal for it to say 117% TDP? I am not familiar with nvidia GPUs.

 

[Rvenger]

Really? This seems to show otherwise. No throttle, full load.

That is throttling. Stock 680 clocks are
Core Clock: 1006MHz
Boost Clock: 1058MHz

The real issue is the voltage which should be at 1.175v. Furmark is not running boost state. If I was developing drivers, I'd be throttling Furmark too. It kills cards.


When I get home I can run some furmark on my HD 7950 to see if AMD throttles on 14.7 Catalyst. I only paid $100 for the GPU but running it for 10 minutes shouldn't kill the card.

We probably should create a new thread if we want to delve into furmark further.

 

[Silverforce11]

Furmark isn't representative of gaming loads the same as mining isn't.

Hammering the same few instructions over and over isn't gaming load.

AMD cards will hit their respective TDP under mining load, but gaming is much less.

 

[RussianSensation]

http://www.3dcenter.org/artikel/str...nce/watt of a GPU on a standalone basis. [/U]

 

[ShintaiDK]

That is throttling. Stock 680 clocks are
Core Clock: 1006MHz
Boost Clock: 1058MHz

Boost clock is never garanteed. Baseclock however is. Throttle would assume to be below baseclock.

 

[boxleitnerb]

And the card can certainly boost above 1058 MHz at stock. 1058 MHz is the average (by Nvidia observed) clockspeed. If you put the rig into the fridge (aside from the condensing water) it would boost much higher without one modifying the card or the software itself.

 

[AtenRa]

I believe that furmark is suitable for worst case scenarios. Worst case scenarios begin to make sense in the long run. For example, as games become more demanding, the gpu will be loaded at almost 100%. That's where furmark tests begin to make sense: when you're stuck with the same GPU for a few years. I always take into account max energy consumption. While simulating 100% load with furmark may seem artificial at launch, in the long run I think is important.

You can have 100% load on the GPU in different applications/games but not the same power consumption.
Furmark doesn't represent gaming scenarios not even in the worst case. It is only suitable for maximum power consumption evaluation of the graphics card.

 

[tollingalong]

My theory:

R9 285 is designed to get people to buy the existing cards thinking they're a better deal. When the existing inventory goes down then proof! The prices of the 285 plummets down to $180.

I hope that's AMD's tactic anyway.

 

[toyota]

Boost clock is never garanteed. Baseclock however is. Throttle would assume to be below baseclock.

no not really. if the card is under max load then it will actually go above rated boost clock even. that card is only hitting base clock at full load which clearly shows the clocks are being slightly throttled for one reason or another. in that screenshot temps and power limit seem mostly responsible for the lowered clockspeeds.

 

[Gloomy]

My theory:

AMD needed a new flagship notebook GPU and Tahiti and Hawaii aren't suited for that form factor.

They've been riding Pitcairn too long already in that space. This chip will probably end up rebranded to R9 370. The desktop cards are probably using the leakier bins, and higher binned parts will go to notebooks.

Some of you are taking this launch as a personal insult. That's cool. :thumbsup:

 

[DeathReborn]

Around the time these launch I would like to see benchmarks with a R9 280 & 285 using Mantle BF4 to see how the 2GB vs 3GB fares with the memory leaks. I think that maybe even 1080p will be choking the 2GB 285. It may well be 10-20%+ faster than Tahiti but it may struggle with VRAM until 3rd party 4GB cards arrive.

The 285 might be the "mainstream" card of choice for Star Citizen come the turn of the year, currently it's probably Pitcairn XT.

 

[blackened23]

Yeah, Tonga is definitely better suited for small (ie ultrabook) form factors than the Tahiti was. But I was still thinking the PPW improvement would be more substantial based on what various rumor sites hinted at, what do they know though I guess.

I can't be sure entirely what AMD's intent was here, that (better efficiency for portables) could be a factor. Like you said....pitcairn has been used forever and a day as AMD's mobile dGPU. I also think that a cheaper to produce card to position itself against the lower tier 2nd gen maxwell cards (ie GTX 860) was also a factor. They can be more flexible with pricing with Tonga, since Tahiti would be more expensive to produce and severely limit margins. (therefore why the 280/280X will be EOL at some point).

Who knows though!

 

[Rvenger]

Tonga may actually be good here. Let's wait to see some reviews. It can be a step in the right direction for the fact that AMD is starting to focus on efficiency which is not their strong suit.

 

[caswow]

it sound like you actually cant buy amd cards because they melt your desk...its not even close as beeing bad as evergreen and fermi:thumbsdown: not even close\

Infraction issued for thread crapping.
-- stahlhart

 

[Enigmoid]

Yeah, Tonga is definitely better suited for small (ie ultrabook) form factors than the Tahiti was. But I was still thinking the PPW improvement would be more substantial based on what various rumor sites hinted at, what do they know though I guess.

I can't be sure entirely what AMD's intent was here, that (better efficiency for portables) could be a factor. Like you said....pitcairn has been used forever and a day as AMD's mobile dGPU. I also think that a cheaper to produce card to position itself against the lower tier 2nd gen maxwell cards (ie GTX 860) was also a factor. They can be more flexible with pricing with Tonga, since Tahiti would be more expensive to produce and severely limit margins. (therefore why the 280/280X will be EOL at some point).

Who knows though!

Tonga isn't going in an ultrabook period with a 190W TDP. Maybe a cut down version can fit in a beefcake notebook.

 

[FatherMurphy]

http://www.tomshardware.com/news/tsmc-apple-nvidia-denver-finfet,27538.html

Apparently TSMC is moving up its 16nm (20nm w/ FINFET) process by a quarter to the first part of 2015 for some customers.

The 285/285x may simply be a stop gap for AMD until it can release 300 series parts at the smaller process node.

We'll need to wait for the reviews, but it certainly doesn't seem (based on the specs) like AMD's evolution of GCN with Tonga is an adequate answer for what is anticipated for 2nd Gen Maxwell. Perhaps AMD's "Maxwell," that is, AMD's big architectural shift toward efficiency, will be coupled with the process node shift in the first part of 2015.