AMD: R9xx Speculation

[Alexko]

I would argue not by that much though (for a full chip with "reasonable" achievable clocks - 5870 doesn't really overclock too much without additional voltage which makes power draw much worse). Of course, if nvidia could produce a full chip in quantity is another question (I have no idea really).
So if GF104 is 10% bigger and 10% slower basically, that wouldn't be too bad. I would say definitely better than any gt2xx comparisons against rv7xx. Though I suspect GF106 against Juniper will look worse again for nvidia (I still think Cypress didn't scale too well compared to Juniper).

On the other hand and if memory serves, GT200b was a bit better than RV770 in perf/W, while GF104 is a bit worse than Cypress, as the GTX 460 1GB draws 10 to 20W more than the HD 5850 while being a good bit slower.

I wonder how things would look with a full GF104 at 1500MHz or so, compared to the HD 5870.

 

[chavvdarrr]

I wonder why noone was able to open the chip and take a shot of the die.
Charlie gives some numbers... but no shots :S

 

[rpg.314]

I wonder why noone was able to open the chip and take a shot of the die.
Charlie gives some numbers... but no shots :S

I don't think getting a die shot is as simple as that. You'll need pretty costly stuff for a die shot, something which will be their with the chip vendors but is unlikely to exist with tech sites/journalists.

 

[AlexV]

I don't think getting a die shot is as simple as that. You'll need pretty costly stuff for a die shot, something which will be their with the chip vendors but is unlikely to exist with tech sites/journalists.

To measure die size you need a ruler, and that's exactly what's needed to put this 460 die size thing to rest (that and a bit of time to de-cap it).

 

[rpg.314]

To measure die size you need a ruler, and that's exactly what's needed to put this 460 die size thing to rest (that and a bit of time to de-cap it).

Yes, But he wasn't angling for the dize size, was he? He seemed a tad more interested int he _die shot_

 

[AlNom]

Problem is that you'd only get the top layers of the chip, which may not be particularly interesting. Not sure what sort of lens you'd need. A microscope camera would be nice?

 

[CarstenS]

We can now tell you about GF104 die, it is 24.9mm x 14.7mm, and the die size is 367mm^2.

Nice rant once again from Charlie. I am looking forward to see which parts turn out to be true in a few months.

While the die size number seems to be a little on the high side of things, I outright doubt that the chip manufacturing and product pricing economics work like Charlie depicts. I don't think, no, I am pretty sure that I couldn't go to TSMCs and buy a wafer for 5000 Dollars, even if that was the price they're asking from Nv or AMD.

edit: Conveniently, he of course forgets that GF104 has also turned of one out of four quad-ROPs/memory controllers/128 kiB L2 Caches, while in Cypress for the higher cost parts, all of those must be present.

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.

I've been adding the data to Davids graph (hope he doesn't mind) for GF100 and Cypress in their respective supercomputing solutions, based on wikipedia data (AMD: 525 DP-GFLOPS, 225 watts, Nvidia: 515 GFLOPS, 225 watts):

http://img682.imageshack.us/img682/1228/computationalefficiency.png
This way, it's getting even clearer.

 

[larrabee]

I don't think getting a die shot is as simple as that. You'll need pretty costly stuff for a die shot, something which will be their with the chip vendors but is unlikely to exist with tech sites/journalists.

you can use 1500 grit sandpaper and get ok results. it wont be as pretty as the dice you see on the internet but it will work. it's not really a simple project and it would take too much time for a review. they could pay a company like chipworks to do it for them. it takes a lot of skill and background knowledge to do properly.

here's a gallery of DIY deprocessed dies, not mine but pretty cool.
http://ctho.org/gallery/bin/category.pl?id=3&category=Computer%20/%20Tech

Problem is that you'd only get the top layers of the chip, which may not be particularly interesting. Not sure what sort of lens you'd need. A microscope camera would be nice?

if you want the pretty stuff sandpaper and polishing will do the trick.

 

[Psycho]

While the die size number seems to be a little on the high side of things, ... Conveniently, he of course forgets that GF104 has also turned of one out of four quad-ROPs/memory controllers/128 kiB L2 Caches, while in Cypress for the higher cost parts, all of those must be present.[/i]

The die size is most probably right, that's easy to get and verify (and if too high nvidia would probably ask some other sites to leak the right number).
He's comparing to 5850/70 combined, so the yield improving by fusing off parts is taking into account, however it's still quite unclear why 5850+70 should yield at 75% and the 460 only at 60%. With the lower density it should have chance of slightly better yield.

I've been adding the data to Davids graph (hope he doesn't mind) for GF100 and Cypress in their respective supercomputing solutions, based on wikipedia data

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).

 

[neliz]

edit: Conveniently, he of course forgets that GF104 has also turned of one out of four quad-ROPs/memory controllers/128 kiB L2 Caches, while in Cypress for the higher cost parts, all of those must be present.

He was always the one hinting that Cypress had a bit more than we would probably ever see.

I've been adding the data to Davids graph (hope he doesn't mind) for GF100 and Cypress in their respective supercomputing solutions, based on wikipedia data (AMD: 525 DP-GFLOPS, 225 watts, Nvidia: 515 GFLOPS, 225 watts):

you used Wikipedia as "facts" on GF100's power usage? 225NVWatts is 325Watt for us.That makes it more close to RV770 in the perf/W figure (seeing as you did pick the right 529mm.

 

[dkanter]

He was always the one hinting that Cypress had a bit more than we would probably ever see.



you used Wikipedia as "facts" on GF100's power usage? 225NVWatts is 325Watt for us.That makes it more close to RV770 in the perf/W figure (seeing as you did pick the right 529mm.

Carsten: I'm actually updating the chart myself, and I hope to put out a revised version on a regular basis. There are a ton of data points that need to be added from both CPU and GPU land.

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

David

 

[Novum]

Well, 5870 is smaller and performs better than GF104.

We haven't seen the full version of GF104 yet. There could be yield problems, but I suspect they are holding it back because of strategic reasons.

 

[eastmen]

We haven't seen the full version of GF104 yet. There could be yield problems, but I suspect they are holding it back because of strategic reasons.

The same can be said of cypress

 

[RecessionCone]

The same can be said of cypress

In what sense? I thought the HD 5870 had all its bits and pieces enabled, and reports seem to indicate that GTX460 has quite a bit more clockspeed headroom. What would a full Cypress look like?

 

[3dcgi]

We haven't seen the full version of GF104 yet. There could be yield problems, but I suspect they are holding it back because of strategic reasons.

What might those strategic reasons be? Low volume due to low yields isn't a strategic reason and I can't think of another reason at the moment.

 

[eastmen]

In what sense? I thought the HD 5870 had all its bits and pieces enabled, and reports seem to indicate that GTX460 has quite a bit more clockspeed headroom. What would a full Cypress look like?

who told oyu it had all its pieces enabled ? Allwe know is that the 5870 is the most complete cypress , we don't know if it is a complete cypress though.

The 5830 and 5850 are both salvage pieces we are almost a year into cypress's life point and even the heavly cut 5830 still exists. We know that they are supply limited so you would think that the 5830s that could work as 5850s would be used as 5850s. So its obvious they still aren't producing 5850s and thus 5870s in large enough numbers. So the question is , is a fully functioning cypress being produced right now?

 

[rpg.314]

While the die size number seems to be a little on the high side of things, I outright doubt that the chip manufacturing and product pricing economics work like Charlie depicts. I don't think, no, I am pretty sure that I couldn't go to TSMCs and buy a wafer for 5000 Dollars, even if that was the price they're asking from Nv or AMD.

Capacity and confidentiality issues aside, why would TSMC NOT sell you the wafers? And I am sure you understand that the wafer prices are a function of demand/supply/history of producer-consumer relationship. With that in mind I haven't really understood your doubts. Can you clarify?

 

[trinibwoy]

Capacity and confidentiality issues aside, why would TSMC NOT sell you the wafers?

Because the price the IHVs pay is inclusive of amortized fixed costs. If you buy a single wafer it's not going to be $5000.

And I am sure you understand that the wafer prices are a function of demand/supply/history of producer-consumer relationship. With that in mind I haven't really understood your doubts. Can you clarify?

That's his point. Charlie's obsession with die size and ridiculous suggestions that he knows AMD's and Nvidia's cost schedules with TSMC are probably far off base given the realities of such contracts - not the least of which are the volume/relationship considerations you mentioned.

 

[rpg.314]

Because the price the IHVs pay is inclusive of amortized fixed costs. If you buy a single wafer it's not going to be $5000.

I know about the fixed costs. I was wondering about whether or not it is possible for Carsten to get wafers from tsmc?

That's his point. Charlie's obsession with die size and ridiculous suggestions that he knows AMD's and Nvidia's cost schedules with TSMC are probably far off base given the realities of such contracts - not the least of which are the volume/relationship considerations you mentioned.

Although it is certainly possible, I am skeptical that there is an appreciable difference in the amortized per wafer price paid by amd vs nv.

 

[CarstenS]

you used Wikipedia as "facts" on GF100's power usage? 225NVWatts is 325Watt for us.That makes it more close to RV770 in the perf/W figure (seeing as you did pick the right 529mm.

TMost 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.
David

Oh, speaking of a hornet's nest…
Well, I double checked my facts of course, but links to wikipedia are so much more convenient. However, both companies give exactly the same numbers as those I've linked in wikipedia.
http://www.amd.com/us/press-releases/Pages/firestream-peak-performance-2010june23.aspx
http://www.nvidia.com/object/product_tesla_M2050_M2070_us.html
Since I don't know for sure what power and FLOPS values David used for his chart I assumed they were the official ratings for all kinds of processors.

Carsten: I'm actually updating the chart myself, and I hope to put out a revised version on a regular basis. There are a ton of data points that need to be added from both CPU and GPU land.

I am looking very much forward to see it.