• We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.

Leaked AMD Catalyst driver codenames for Volcanic Islands GPU's: Hawaii confirmed

Page 3 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
^ of course you can, that is what all those sites in my link to. They measure the power through the slot AND through the PCIe power cables. It is expensive and difficult to do this, but it can be done.
 
boxleitnerb said:
^ of course you can, that is what all those sites in my link to. They measure the power through the slot AND through the PCIe power cables. It is expensive and difficult to do this, but it can be done.
Click to expand...

Well I don't read german and have no chrome at work 🙁.

But I had no idea one could buy one of the risers in this photo:
http://www.behardware.com/articles/781-1/report-the-true-power-consumption-of-73-graphics-cards.html

Learn something everyday. Teach me for assuming noone makes such a device (and that reviewers wouldn't bother spending the time and money to build a way to do it), though I suppose it would just need to be a terminal block to a current meter. At any rate, neat stuff. Sorry for my silly assumptions.

That being said if there is some reason the rest of the system draws more power due to one card or the next that is rather interesting.
 
Last edited:
boxleitnerb said:
Why pointless? If we want to know how much power a graphics card uses, we measure the graphics card. Plain and simple. This applies to all measurements that take place in a multi-component system.
Click to expand...

The rest of the components in the system end up consuming ~40 watts more as a result of being paired with said graphics card.I think it's much more useful to measure its consumption as part of a whole
 
sontin said:
Proof? For what? That the GTX780 uses 30-50W less power when Boost 2.0 is kicking in?
there is really no proof for that:
http://www.computerbase.de/artikel/grafikkarten/2013/nvidia-geforce-gtx-780-im-test/9/
http://ht4u.net/reviews/2013/asus_matrix_hd_7970_platinum_im_test/index14.php
http://www.hardware.fr/articles/894-3/consommation-efficacite-energetique.html
Click to expand...

Your post fails even harder considering that you cannot buy reference HD7970GE cards on the market and those 3 sites you linked used them.

After-market 1.05ghz 7970 uses less power than a 925mhz 7970 Reference:
http://www.techpowerup.com/reviews/VTX3D/Radeon_HD_7970_X-Edition/26.html

How about looking at real world after-market 7970GE cards like Anandtech used in the review?

What about 1180mhz 7970 HIS that uses 211W on average
http://tpucdn.com/reviews/HIS/HD_7970_X_Turbo/images/power_average.gif

vs. ASUS DCUII GTX780 that uses 191W on average:
http://tpucdn.com/reviews/ASUS/GeForce_GTX_780_Direct_Cu_II_OC/images/power_average.gif

For the last 18 months, you can't comprehend that there is no such thing as a reference HD7970GE and yet keep linking reviews of flashed 7970 925mhz cards with GE bioses. Stop trolling the forums already. My card at 1150mhz uses less power than reference 7970GE's in those reviews. 🙄

boxleitnerb said:
And much more prone to errors. I'm sorry, but as an engineer, I like to do things as exact as possible.
Click to expand...

If you want to be scientific, you wouldn't be discussing scores of cards no one can buy on the market - i.e., reference 7970GE. AT pitted HIS 7970GE and it uses less power in the system than GTX780 did. Therefore, sontin's 30-50W difference on average against a real world 7970GE is purely a pigment of his imagination.
 
Last edited:
RS, I thought you would have already learned your lesson about arguing with sontin? 😀

OT, I hope AMD releases something to compete with the 780/Titan soonish...they shouldn't let nV keep the crown for too long. Should go back and forth IMO.
 
RussianSensation said:
It's understandable how AMD could reduce the GPU voltage at a similar clock speed OR reduce the GPU voltage with a larger die size but end up with a much lower clock speed (similar to 780 vs. 680). However, what you are proposing is a lot lower voltage than HD7970GE and a much larger die size on the same node with barely a 50mhz reduction in GPU clock I am not sure that can be done with 250W of real world power consumption.
Click to expand...

Firstly the reference HD 7970 cards at 925 mhz came with 1.175v and easily hit 1125 mhz at stock voltage. This was a chip which launched in Jan 9, 2012. It must have taped out in late Q1 2011. The 28nm process would have been very immature and the AMD design would have been conservative. AMD did not tapeout a big 440 sq mm chip after that. Assuming a late Q3 or early Q4 2012 tapeout and a Sep / Oct 2013 launch you are talking about a much more improved TSMC 28nm process , improved learning on AMD's side. Add to it a new power management scheme with more voltage states and finer power control as found in Bonaire and I don't see why a 5.5 billion transistor , 1 Ghz boost clocks, 250W chip is not possible. the GTX Titan with 7 billion transistors is hitting 900+ mhz boost speeds within a 250w TDP so you are overly pessimistic. Moreover Hawaii will arrive almost a year after the first GK110 Tesla chips. so AMD's design will make the best of TSMC process improvements.

I strongly disagree. 35% more over 7970GE would put it faster than the Titan. If AMD "easily" release such a card (i.e., not difficult), what were they doing since February 2013 when they already knew Titan is dropping? It's obviously not easy at all since it's already near end of June and AMD has no such card in sight.

Also, Bonnaire's die size went up 30% to get that 30% increase in performance, while GPU clock remained at 1Ghz like Cape Verde. 30% increase in die size over 365mm2 is nearly 475mm2, not 420-440mm2.
Click to expand...

first of all die size increases are not always proportional. the increase in sp from cape verde to bonaire is 256 sp. the front end was doubled from 1 ACE, 1 geometry engine / raster engine to 2 ACE and 2 geometry / raster engines. it took 37 sq mm extra die space. Looking at 512 sp more and increase in front end resources and extra 16 ROPs we can expect 90 - 100 sq mm extra die space. But due to denser packing of transistors which is possible as AMD will take advantage of a very mature TSMC 28nm process and a more aggressive design we can expect a die size increase from 365 sq mm to 440 sq mm. If GK110 with 7 billion transistors can fit in 560 sq mm ( 560 x 5.5/7 = 440 sq mm) Hawaii can fit 5.5 billion transistors in 440 sq mm.
 
Last edited:
Given recent leaks I suppose 20nm-GPUs in small doses in Q4/2013 is possible.
Availability of Cypress (40nm) was bad and Tahiti (28nm) was clocked quite conservatively, and yet AMD launched both of them successfully.

If the "HAWAII GL40" entry truly points to 40 CUs for the salvage part, this would make sense:

Tahiti Pro: 28 CUs, 1792 SP
Tahiti XT: 32 CUs, 2048 SP (+15%)

Hawaii Pro: 40 CUs, 2560 SP
Hawaii XT: 48 CUs, 3072 SP (+20%)

With a better front end, Hawaii XT could be about 60% faster than Tahiti XT per clock. I would expect clocks to be quite low (1-1.1 GHz) in the beginning for yield reasons.

Big question is bandwidth. A 512-bit interface is expensive and consumes lots of power. With GDDR5, you can only achieve a 17% higher bandwidth compared to current top cards. How are GDDR6 and HMC coming?
 
boxleitnerb said:
Given recent leaks I suppose 20nm-GPUs in small doses in Q4/2013 is possible.
Availability of Cypress (40nm) was bad and Tahiti (28nm) was clocked quite conservatively, and yet AMD launched both of them successfully.

If the "HAWAII GL40" entry truly points to 40 CUs for the salvage part, this would make sense:

Tahiti Pro: 28 CUs, 1792 SP
Tahiti XT: 32 CUs, 2048 SP (+15%)

Hawaii Pro: 40 CUs, 2560 SP
Hawaii XT: 48 CUs, 3072 SP (+20%)

With a better front end, Hawaii XT could be about 60% faster than Tahiti XT per clock. I would expect clocks to be quite low (1-1.1 GHz) in the beginning for yield reasons.

Big question is bandwidth. A 512-bit interface is expensive and consumes lots of power. With GDDR5, you can only achieve a 17% higher bandwidth compared to current top cards. How are GDDR6 and HMC coming?
Click to expand...

20nm is out of the question. read TSMC Q1 2013 earnings call. if you are not registered with seekingalpha here is the excerpt

Steven C. Pelayo - HSBC, Research Division
All right, then, if you could just remind me, what -- when is 20-nanometer revenues going to start being 1%, or 2%, or 3% of revenue? I think -- I forgot if you guys could comment on it?
Morris Chang - Chairman and Chief Executive Officer
20-nanometer?
Steven C. Pelayo - HSBC, Research Division
Yes.
Morris Chang - Chairman and Chief Executive Officer
20 has not started.
Elizabeth Sun - Director of Corporate Communication Division
20 has not started.
Steven C. Pelayo - HSBC, Research Division
I know. When do you expect?
Elizabeth Sun - Director of Corporate Communication Division
When? First half next year.
Lora Ho - Chief Financial Officer and Senior Vice President of Finance
You're asking about 20?
Morris Chang - Chairman and Chief Executive Officer
Well, she answered already, she said first half of next year. You actually set a pretty low hurdle. You said 2% or 3%, right?
Steven C. Pelayo - HSBC, Research Division
What quarter will be the first few percentage of revenues will come from 20-nanometer for TSMC?
Morris Chang - Chairman and Chief Executive Officer
What quarter will be the first 2% quarter?
Lora Ho - Chief Financial Officer and Senior Vice President of Finance
Well, based on our current estimation, it will be roughly second quarter 2014.

fyi Q4 2011 was the first quarter when 28nm production hit 2% of total TSMC production volume.

http://www.xbitlabs.com/news/other/..._Fourth_Quarter_Totaled_Nearly_1_Billion.html

HD 7970 shipped in early Q1 2012. so the earliest we can expect 20nm graphics cards is June or July 2014.

AMD would ship Hawaii on 28nm in September 2013 with very high volume. AMD might not be keen in shipping 20nm cards first. AMD saw how launching slightly later helped Nvidia postition their GTX 680/670 cards very well. AMD might do the same with 20nm and could launch in late August 2014 (a couple of months after GM104 Maxwell aka GTX 880) .

also for 20nm AMD will aim for a 4096 sp chip with 384 bit GDDR6 . 2014 is the year when GDDR6 should launch with AMD.

http://vr-zone.com/articles/gddr6-memory-coming-in-2014/16359.html
 
boxleitnerb said:
Hm, but what about risk production?
Click to expand...

you cannot ship products until volume production begins. here is the press release on Oct 24, 2011 talking about TSMC beginning 28nm volume production. retail HD 7970 was launched on Jan 9, 2012.

http://www.tsmc.com/tsmcdotcom/PRListingNewsAction.do?action=detail&newsid=6181

so the earliest you can expect 20nm Maxwell GM104 should be late June / early July 2014. AMD might be even later given that Hawaii on 28nm will launch only in Sep / Oct 2013. I expect AMD's 20nm flagship product to launch with GDDR6 in Aug or Sep 2014. :thumbsup:
 
Good point. 40nm volume production startet in November 2008, Cypress launched in September 2009. Though that seemed a bit late, but 40nm had problems iirc.

But...from October 24th 2011 to January 9th 2012 it's less than 3 months. The earning call was from Q1 which is over by now. For a 20nm launch in September/October, mass production should start in June/July at the latest. Which is still possible I guess.
 
Last edited:
boxleitnerb said:
Good point. 40nm volume production startet in November 2008, Cypress launched in September 2009. Though that seemed a bit late, but 40nm had problems iirc.

But...from October 24th 2011 to January 9th 2012 it's less than 3 months. The earning call was from Q1 which is over by now. For a 20nm launch in September/October, mass production should start in June/July at the latest. Which is still possible I guess.
Click to expand...


The 4770 was the first 40nm and launched April 2009.
 
But is also was nothing more than a "test chip", small and relatively easy to manufacture. We're talking about a supposed Tahiti successor here with 350mm2 die size and 6-7 billion transistors.
 
boxleitnerb said:
But is also was nothing more than a "test chip", small and relatively easy to manufacture. We're talking about a supposed Tahiti successor here with 350mm2 die size and 6-7 billion transistors.
Click to expand...

True. I was just pointing out that the 5870 wasn't the first 40nm chip like the 7970 was. Remember 32nm was cancelled and changed to 28nm. That doesn't instill confidence that 20nm will be a walk in the park. It could be a while before we see anything on 20nm.
 
Why was that, incidentally? With most new nodes, special features are introduced like HKMG, FinFET, strained Si etc.
Maybe with 32nm, this particular new addition was so troublesome and the reason for cancelling? If they fixed it with 28nm, I don't see why 20nm should be as problematic as 32nm or 40nm. But I don't know if 20nm brings a more radical change to the table. FinFET is due for 14/16nm I believe.
 
boxleitnerb said:
Good point. 40nm volume production startet in November 2008, Cypress launched in September 2009. Though that seemed a bit late, but 40nm had problems iirc.

But...from October 24th 2011 to January 9th 2012 it's less than 3 months. The earning call was from Q1 which is over by now. For a 20nm launch in September/October, mass production should start in June/July at the latest. Which is still possible I guess.
Click to expand...

please did you even read the TSMC Q1 2013 earnings excerpt. it said 20nm volume production in Q2 2014. so products can be expected in June or July 2014 at the earliest. and you are saying 20nm products in Sep 2013 is possible 😱
 
raghu78 said:
please did you even read the TSMC Q1 2013 earnings excerpt. it said 20nm volume production in Q2 2014. so products can be expected in June or July 2014 at the earliest. and you are saying 20nm products in Sep 2013 is possible 😱
Click to expand...

They didn't really clarify, because the question was about 1-3% volume. It depends on what you mean by volume production. For a limited Hawaii launch in Q4 2013, even less than 1% volume might be ok.

And don't forget - we don't know what percentage of those 1-3% actually go to AMD on 20nm and how that number was for 28nm. There are other companies, too, which makes estimates based on this difficult if not impossible.

Yet one thing makes 20nm unlikely:
If it really were ready for Q4 2013, Nvidia would never have launched their "refresh" right now. They could neither let AMD lead with 20nm for 6 months like with Fermi nor could they replace their products only 3 months after their launch.
 
Last edited:
boxleitnerb said:
They didn't really clarify, because the question was about 1-3% volume. It depends on what you mean by volume production. For a limited Hawaii launch in Q4 2013, even less than 1% volume might be ok.

And don't forget - we don't know what percentage of those 1-3% actually go to AMD on 20nm and how that number was for 28nm. There are other companies, too, which makes estimates based on this difficult if not impossible.
Click to expand...

generally the first quarter of volume production is the one which has 2% volume production. also TSMC has promised 20nm ramp to be more steeper than 28nm. so even though 20nm production volume in Q2 2014 is 2% of total volume , TSMC will ship more 20nm wafers in 2014 than 28nm wafers in 2012. so Q2 2014 is going to be the first quarter of volume production. from the xbitlabs table 28nm production in

Q1 2012 - 5%
Q2 2012 - 7%
Q3 2012 - 13%
Q4 2012 - 22%

( 5 + 7 + 13 + 22) / 4 = 47/4 = 11.75 % of total 2012 production volume was 28nm volume

here is another question from TSMC Q1 2013 earnings call

Andrew Lu - Barclays Capital, Research Division
Dr. Morris Chang and Lora, I have 2 questions. Last time you mentioned 20-nanometer in next year revenue will be larger than 28-nanometer last year. How about the percentage in each quarter, which means the total percentage revenue will be higher compared to year 2012? For example, last year, first quarter percentage, 28-nanometer is about 5%, but 22% by Q4.

Morris Chang - Chairman and Chief Executive Officer
I'm not going to go into those details. But I repeat, I'll reiterate, I stand by what I said last time. That is, that the volume -- of production volume upward of 20-nanometers next year will be greater than 28-nanometer in 2012. That, I stand by. But as to the percentage and so on...

TSMC intends to in fact have more 20nm wafers shipped in year 2014 than 28nm wafers in year 2012. so its easy to see that TSMC's 20nm volume production starts in Q2 2014 with a very fast and steep ramp in H2 2014.
 
Well, in April 20nm already passed risk production, so that means you can order wafer runs right now (albeit in low volume I guess):
While TSMC has four "flavors" of its 28nm process, there is one 20nm process, 20SoC. "20nm planar HKMG [high-k metal gate] technology has already passed risk production with a very high yield and we are preparing for a very steep ramp in two GIGAFABsTM," Sun said.
Click to expand...

http://www.cadence.com/Community/bl...-20nm-16nm-finfet-and-3d-ic-technologies.aspx
 
Just throwing this out there but you can not believe what is said on quarterly conference calls all the time for a number of reasons. They will purposely under estimate and over deliver on forward looking guidance at times for a number of reasons on earnings and new products so I wouldn't put 100% on what was is said on the conference call imo.
 
Last edited:
BallaTheFeared said:
Nope, GPU mining is all but dead and next gen consoles and their respective games are not out yet.

Unless something dramatic happens, I'll probably wait for DX12. Assuming of course AMD is fixing their drivers, and given the total lack of any real improvements over the last 4-5 months I sure hope that is the case. :thumbsup:
Click to expand...

Are you serious??? http://www.youtube.com/watch?v=3gB9-eIPLzM with latest drivers, even the R7950 outperforms the GTX680... no improvement over the last months, right...
 
The 7950 certainly does not perform better than the 680. You can see this in any current review from AT, TPU, CB etc. Who needs Linus...?
 
You must log in or register to reply here.

TRENDING THREADS

Back
Top