AMD: R9xx Speculation

[neliz]

TDP: 247W http://www.nvidia.com/object/product_tesla_C2050_C2070_us.html

The V3 document is from July 16th, when was that page last updated? 9W for 3GB of RAM sounds reasonable though.

 

[Silent_Buddha]

That's what Charlie seems to be saying too.



Now how do you reconcile that with the fact that Nvidia has no CPU business yet has higher gross margins than AMD (45.6% vs 45%) in the second quarter of this year with those huge economically unviable dies they keep making? Inconvenient facts?

I'd say Tesla and Quadro have a fair bit to do that with. They command much higher price premiums than ATI's FireGL line. Likewise AMD's margins are hampered by their CPU line where competition with superior (generally) products from Intel severly limit possible margins.

Regards,
SB

 

[Entropy]

May be you should look here

http://www.realworldtech.com/page.cfm?ArticleID=RWT090909050230&p=2

Your views on perf/W of GPU's might change a bit.

That chart does not affect my argument, which was about adding GPU capabilites for co-processing in a general purpose system.


It does however illustrate the state of things.
That chart has been constructed to show FLOPs/W and FLOPs/mm^2.
But those FLOPs are marketing numbers. No application generated the FLOPs data. No computational kernel, or even LINPACK. There is nothing there that connects to the real world.
And the article demonstrates no understanding of utilization, such as distinct differences between serial, vector, parallel and vector parallel codes, what limitations there are in data organization and usage for different cases. Even then, we are still in the domain of theory, in the realm of physical devices we have to deal with the memory subsystem (which is what typically defines what performance you can wring out of a small computation kernel), communications et cetera. To actually produce code, you need tools that allows you to optimally access the hardware, so now we have gotten to the software side of things which is its very own can of worms. Of course the application is rarely just a tight computational kernel, so we have Amdahls law to deal with, and the further up you go in specialized capabilities, the harder it applies.

All of the above concerns new code, and specifically targeted at that. Legacy codes in this case gain nothing and old code, and old code reused in new applications, constitute just about 100% of what is run on general purpose computers. I could go on and on.

 

[Gipsel]

Most other numbers are for graphics cards, so the 5870/1G (544@188w) would be more comparable than the 4gb firestream. On the other hand the C2070 is (now) specified at 247w, not 225 (unfortunately no gf100 graphic cards to use here).

And I strongly recommend not using wikipedia as a source for anything.

So, here you have it with the consumer cards added:

I've taken the official numbers for the TDP straight from the manufactures websites (and wikipedia's numbers Carsten used were actually correct, if one considers the M2050/70 modules, which are better comparable to the Firestream 9370 either way as they also come without fan). The same goes for the GTX480 (250W), so one may want to correct the GTX480 point a bit down, as we know the official TDP is an average and nothing close to the maximum consumption as in case of the Teslas or Firestream cards.

 

[CarstenS]

That's Easy, C2070 runs it's GF100@1.0V and has a "board power" greater than 238W. (actually.. I think it's still listed as "TBD" in the official documents.) GTX480 default is 0.950 right the "C2070" documents from last year say 1.05v? What are you getting at?

Tesla C2050/2070 Specs v1 November 2009:


Tesla C2050/2070 Specs v3 July 2010

I've linked the official specs twice (once via wikipedia, once via the respective IHVs homepage). The Specs you linked didn't even have a fix on the clockrates, that's enough for me to completely disregard them for the obviously newer ones on the website, when it comes to TDP. Apart from that, I was linking to Tesla M2070 for a reason. The Tesla card comes with a 22 watts fan (and that's why it's TDP is exactly that amount higher) - something which isn't factored in into CPUs (apart from the memory issue, which I tend to view as a working set cache) and also not on Firestream 9370, as you can see here. We don't want to spoil our data, don't we?
http://www.3dnews.ru/tags/FireStream 9370



Besides, GTX 480 isn't 0,95 volt - that's 470. 480 is 1,0 Volts.

After having read you talking about Fermi exceeding its TDP under certain workloads, i couldn't help but wonder why you haven't hinted at this also wrt to other chips which are doing the same (RV770). Coincidence?

 

[neliz]

The Specs you linked didn't even have a fix on the clockrates, that's enough for me to completely disregard them for the obviously newer ones on the website,

Here's Version 3 of the C2050/2070 Board Specification document, dated July 16th 2010, how do you know your website numbers are newer/better/more real/less disregardable?

www.nvidia.com/docs/IO/43395/BD-04983-001_v03.pdf

Besides, GTX 480 isn't 0,95 volt - that's 470. 480 is 1,0 Volts.

Thanks, I couldn't find any official reference to the stock voltages on the consumer boards..

After having read you talking about Fermi exceeding its TDP under certain workloads, i couldn't help but wonder why you haven't hinted at this also wrt to other chips which are doing the same (RV770). Coincidence?

Because we're not talking about two years ago, this is the R9xx topic. RV7x0 has nothing to prove against the upcoming architectures, Fermi does.

 

[Gipsel]

Here's Version 3 of the C2050/2070 Board Specification document, dated July 16th 2010, how do you know your website numbers are newer/better/more real/less disregardable?

How about linking to the M2050/70 board specification document, as this is what the dot in Carsten's (and also my) graph referred to?

 

[neliz]

How about linking to the M2050/70 board specification document, as this is what the dot in Carsten's (and also my) graph referred to?

http://www.nvidia.com/docs/IO/43395/BD-05238-001_v02_M2050_boardspec.pdf

225W isn't the board power in that case, it's the "Board Power Dissipation" according to that document.

I was confused since people first place a GPU in that chart and later on start to specify which model of card they want in there to have to most desirable specifications. Like Cypress is now split in Radeon and Firestream, but GF100 only as a specific Tesla part.

 

[Alexko]

Sure making up facts is always an easy way out.

Uh? What am I making up, here? Obviously I was exaggerating about three and a half GF100s, but Steam's hardware survey clearly shows that they've sold very few of them: http://store.steampowered.com/hwsurvey/videocard/

So if they've sold very few GF100s, their margins depend mostly on the rest, and the rest is mostly DX10.1 parts these days.

Tesla and Quadro have very high margins, that's a fact.

 

[Gipsel]

http://www.nvidia.com/docs/IO/43395/BD-05238-001_v02_M2050_boardspec.pdf

225W isn't the board power in that case, it's the "Board Power Dissipation" according to that document.

What do you think is the difference? None?

I was confused since people first place a GPU in that chart and later on start to specify which model of card they want in there to have to most desirable specifications. Like Cypress is now split in Radeon and Firestream, but GF100 only as a specific Tesla part.

If you look closely, you will also find the GTX480 in my chart

 

[neliz]

If you look closely, you will also find the GTX480 in my chart

At it's mythical 250W power?

 

[aaronspink]

That's what Charlie seems to be saying too.



Now how do you reconcile that with the fact that Nvidia has no CPU business yet has higher gross margins than AMD (45.6% vs 45%) in the second quarter of this year with those huge economically unviable dies they keep making? Inconvenient facts?

They own the vast majority of the WS market and AMD is fire selling 6K series cpus to stay competitive (6100 is 2x the silicon have < 1/2 the price as what it replaced).

 

[rpg.314]

That chart does not affect my argument, which was about adding GPU capabilites for co-processing in a general purpose system.


It does however illustrate the state of things.
That chart has been constructed to show FLOPs/W and FLOPs/mm^2.
But those FLOPs are marketing numbers. No application generated the FLOPs data. No computational kernel, or even LINPACK. There is nothing there that connects to the real world.
And the article demonstrates no understanding of utilization, such as distinct differences between serial, vector, parallel and vector parallel codes, what limitations there are in data organization and usage for different cases. Even then, we are still in the domain of theory, in the realm of physical devices we have to deal with the memory subsystem (which is what typically defines what performance you can wring out of a small computation kernel), communications et cetera. To actually produce code, you need tools that allows you to optimally access the hardware, so now we have gotten to the software side of things which is its very own can of worms. Of course the application is rarely just a tight computational kernel, so we have Amdahls law to deal with, and the further up you go in specialized capabilities, the harder it applies.

All of the above concerns new code, and specifically targeted at that. Legacy codes in this case gain nothing and old code, and old code reused in new applications, constitute just about 100% of what is run on general purpose computers. I could go on and on.

So now you are onto legacy code, architectural differences, Amadhal's law, memory subsystem. While in the last post you said (paraphrasing) "GPUs are horribly inefficient, even if by magic all sw was ported to them".

From the horses mouth.

Viewed in that light, a GPU is incredibly inefficient even if by magic all x86 software was rewritten from scratch due to the very limited set of problems it can be applied to.

Singing a lot different tune about GPU's now aren't we?

If by magic all sw is ported to them, then architectural differences and leagcy code arguments don't apply.

If you "waste" your transistors on x86, you still have to port your code to multi cores and make sure that it scales. So the need to port code argument is quite weak.

And I am going to go out on a limb and claim that the set of consumer-segment workloads that need performance but are unsuitable to GPU's is quite comparable to the set of of workloads that need performance and are suitable to GPU's.

And as time goes on, the former set will keep shrinking. Fermi was already a huge step forward in this regard.

 

[aaronspink]

I saw a Q2-10 report on EETimes which mentioned a median 300mm wafer price of $3200 which was up from $2900 in Q1, are those numbers realistic? And are these price increases a factor when ordering large amounts like AMD/NV do?

Depends what they were reporting on aggregate 300m wafer costs (which would be the median wafer cost across all of TSMC's processes on 300mm wafer) or 40nm wafer cost. It is likely the former which just means the extra cost of 40nm wafers is starting to affect the average overall 300mm price.

 

[eastmen]

I think we can all agree that they are going to sell a boatload of GF104 based cards for ~$200. They're projecting a gross margin improvement in the next quarter. So a couple thousand sales of professional cards are going to offset poor margins on millions of GTX460's? Charlie didn't even go to those lengths trying to justify his misdirected nonsense.

Even with poor margins , as long as the gtx 460 is a money maker (even if its $10 bucks a card) they will still be able to use the high margin professional market and the low end gpu sales to help increase margin improvement.

The real question comes with ati. If they do decide to drop prices is the gtx 460 still competetive when put up against the 5850 ? If thier chip prices cascade downward and the 5830 becomes $175 and the 5770 goes further under $150 will nvidia's new parts keep up ?

I think the ball is fully in ati's court and its up to ati to decide if they are going to fumble or make a second touch down out of the cypress parts.

 

[mboeller]

according to heise.de the new HD6000 cards will be introduced between the 15th October and the 29th October.

Tatsächlich will AMD die Southern-Islands-Chips wohl zwischen dem 15. und 29. Oktober vorstellen.

http://www.heise.de/newsticker/meldung/AMDs-neue-Grafikchips-kommen-Ende-Oktober-1044422.html

 

[ECH]

Even with poor margins , as long as the gtx 460 is a money maker (even if its $10 bucks a card) they will still be able to use the high margin professional market and the low end gpu sales to help increase margin improvement.

The real question comes with ati. If they do decide to drop prices is the gtx 460 still competetive when put up against the 5850 ? If thier chip prices cascade downward and the 5830 becomes $175 and the 5770 goes further under $150 will nvidia's new parts keep up ?

I think the ball is fully in ati's court and its up to ati to decide if they are going to fumble or make a second touch down out of the cypress parts.

But if their partners are reporting average/above average units sold do they really need to cut prices just to undercut nvidia? IMO, this is perhaps why there has been no change in price for the 5850?

Another problem I'm seeing is the notion that a review of an ATI card maybe benched certain way that may not show an actual representation of the card's performance. So now I have to worry about making sure that:
-AI is enabled
-control panel settings are set to default
-etc
Who can you trust once the reviews do come out?

 

[3dcgi]

The answer to both is common sense. Nvidia would cannibalize the expensive GTX 470 with a fully blown GF104 with clocks in the 750/1500 range and AMD would (or should) really DIE to see their fully blown Cypress with a little higher clocks and 1920 ALUs best GTX 480 in most games.

If I were Nvidia I'd be happy if GTX460 could perform close to GTX470 because it's much cheaper to make and I'd sell it for around the same price.

Bear in mind that ATI's rasteriser is still based on R520 (though the fixed function interpolator unit has been deleted). Which was only 16 fragments per hardware thread. Packing multiple triangles' fragments into a thread wasn't a priority back then.

I'm not sure why you think it's based on R520 as it has no more in common with R520 than anything else. It's more closely related to Xenos which was 8 pixels per clock feeding a 64 wide vector. Even that's irrelevant though as there have been multiple generations since then which is plenty of time to change the internals.

 

[Harison]

AMD would (or should) really DIE to see their fully blown Cypress with a little higher clocks and 1920 ALUs best GTX 480 in most games.

Thats the thing, not only AMD arent DIEing over this, they dont even care. Why? I expected AMD to launch GTX480 killer as 5890 (newer stepping with high clocks, which would be low R&D and high reward line, spoiling GTX480 sales), yet AMD didnt bothered. I guess they think its enough to have the fastest card (even if its dual) and selling all the chips they make. Its not what I would have done in AMD place in this particular case, but its their business, and until NV becomes competitive again, AMD is concentrating on Fusion, mobile market, etc.

 

[Harison]

And I'll bet they're not. Based on a similiar amount of information that you provided

I'm sorry, what are you talking about? When orders are huge from two long term customers, the prices they get are very similar, any difference is negligible. Thats common sense as well as common business practice, thats what they are talking about. Yet I have no clue what YOU are talking about. Maybe you want to talk "we havent seen their exact contracts", but common sense should remain, and if you argue against common sense, you better come up with something more than... your opinion.