AMD shows 28nm wafer GlobalFoundries to meet new card?

[mattkosem]

3870 was a midrange GPU.

I believe you're mistaken. 3870 was top end when it was released, and remained there until the 3870 X2 was released.

--Matt

[Behemot]

The RV670 was presented like AMD's first piece of silicon from their "sweet point" strategy as a term for small sized die. RV770 was slightly behind this line. RV870 is so much behind AMD had to try 40 nm on something else.

[Manicdan]

The 3870 was the first product on 55nm, was it not?

--Matt

as far as i recall yes, it was out very shortly after the 2900

3870 was a midrange GPU.

midrange to what? performance was very close to the 2900, but used half the power. the 4870 did beat it, and was bigger and used more power, but came out months later

[ryboto]

as far as i recall yes, it was out very shortly after the 2900

midrange to what? performance was very close to the 2900, but used half the power. the 4870 did beat it, and was bigger and used more power, but came out months later

The 3870 came out 6 months after the 2900, and was a refresh of the R600 on 55nm. Then the 4800's came about 6-8 months after that. Ati was on a pretty consistent release cycle when the process technology was ahead of them...now it seems they're ahead of it.

[LordEC911]

not sure they could do that this time. they already use 128bit on their mid range cards, they cant get that much lower. if anything they might ONLY make mobile gpus on 28nm to test it out, where 64bit super slow gddr5 wont be missed as much.

Remember, they didn't exactly go with the largest bus possible for the die size. A 28nm shrink of Cypress they are probably aiming for ~200mm2 and have some room to play with some specs while being able to fit a 256bit bus on there.
28nm Juniper is looking closer to ~100mm2, so they might have to throw some clusters on there to get a 128bit bus.

[Behemot]

Well, they did at least architecture progress. NVIDIA just added more and more shaders and shrinked. Now they are supposed to release new architecture, but oh, it's problem, they forgot how to do it after those long years or what

On the other hand, AMD does weird things too, they try new things with every new line and sometimes just throw them away right in the another new line. Like they've been just experimenting.

As for 28 nm, I don't think AMD will shrink. They don't do this, they'll more probably release new cards. It's wasting time when you only shrink, I think they learned from NVIDIA. I'd say second half of 2010 is long enough after HD 5000 lanch when you compare it with HD 3000/HD 4000/HD 5000.

[annihilat0r]

Well, they did at least architecture progress. NVIDIA just added more and more shaders and shrinked. Now they are supposed to release new architecture, but oh, it's problem, they forgot how to do it after those long years or what

I don't agree. I think it was more like AMD who did that - take 320 sp's, first shrink it, then shrink + 800, then shrink + 1600.

Nvidia didn't even bother adding SP's to 8800 GTX until the GT200 series. They just shrank once. Of course, that they didn't add shaders doesn't mean they improved a lot in architecture

[Manicdan]

would a shrink of r800 really give them any information on how build the r900 at 28nm?

[generics_user]

I don't agree. I think it was more like AMD who did that - take 320 sp's, first shrink it, then shrink + 800, then shrink + 1600.

Nvidia didn't even bother adding SP's to 8800 GTX until the GT200 series. They just shrank once. Of course, that they didn't add shaders doesn't mean they improved a lot in architecture

not really

amd added direct x 10.1; 11 in the progress; made huge improvements in opencl performance; added additional dispay support; redesigned the UVD and power savings; added GDDR5

AMD: DX 10 > 10.1 > 11 (2900 - 3870/4870 - 5870)
Nvidia: DX 10 > 10 > 10 (8800-9800-280)

the last real core improvement with nvidia was G80-G92 with much faster CUDA performance (G80 sucks in CUDA)

i just wish that they finally come out with fermi but considering the fact that there were no news about it for almost 2 weeks things dont look too good...

[LordEC911]

would a shrink of r800 really give them any information on how build the r900 at 28nm?

Depends on the architecture and it also depends on if they are refreshing at the end of the year @ 28nm and then release a new architecture in 2011. It could go either way.

The real question is, do they really want to try and release a completely new architecture on a new process node with no small test GPUs first?

[HelixPC]

I believe you're mistaken. 3870 was top end when it was released, and remained there until the 3870 X2 was released.

--Matt

The 3870 was never a high end card, it might have been to ATI, but the rest of the enthusiast world did not consider it a high end card. The 8800GT was beating it hands down in every game tested. 8800GT was a $200 card at the time. The 3800 still struggled with AA.

[InCredible]

^ regardless of what it was...ati produced the 3870 as their high end card..just like with their cpus now..just because they arent beating intel's comp..doesnt mean the black edition aren't their high end cpu's..

[Manicdan]

Depends on the architecture and it also depends on if they are refreshing at the end of the year @ 28nm and then release a new architecture in 2011. It could go either way.

The real question is, do they really want to try and release a completely new architecture on a new process node with no small test GPUs first?

and from a new plant, its a new everything, except the word ATI on the front

[mattkosem]

The 3870 was never a high end card, it might have been to ATI, but the rest of the enthusiast world did not consider it a high end card. The 8800GT was beating it hands down in every game tested. 8800GT was a $200 card at the time. The 3800 still struggled with AA.

It was indeed high end in the ATI lineup, replacing the 2900XT. The status of nvidias products at the time is irelevant. The 5870 doesn change the position of the gtx285 in nvidia's lineup, does it? Nor does the i7 on the phenom II line. Each manufacturer has their own product line, and at the time the 3870 was at the top of ATI's.

--Matt

[Particle]

It was indeed high end in the ATI lineup, replacing the 2900XT. The status of nvidias products at the time is irelevant.

--Matt

That's awfully presumptuous. The decision of what product to test on a new process could have been any of the following:

1) Mainstream by own company standards (internal product selection)
2) Mainstream by global standards (market-based product selection if they're aiming for a certain kind of product range)
3) Mainstream by die size (technical product selection)

Nobody knows which of these is true, and #3 is the most likely.

Try it on something small/easier and also cheap. Then, sell it as a market mainstream so it actually makes you some money that late in the product generation cycle. That's what I would do in that position as it makes a lot of sense both from a business and technical point of view compared to other options. It would hardly make sense to shrink a massive chip to a new process late in the cycle unless they're still selling them by the truckload. Mainstream parts move a lot more units.

[god_43]

That's awfully presumptuous. The decision of what product to test on a new process could have been any of the following:

1) Mainstream by own company standards (internal product selection)
2) Mainstream by global standards (market-based product selection if they're aiming for a certain kind of product range)
3) Mainstream by die size (technical product selection)

Nobody knows which of these is true, and #3 is the most likely.

Try it on something small/easier and also cheap. Then, sell it as a market mainstream so it actually makes you some money that late in the product generation cycle. That's what I would do in that position as it makes a lot of sense both from a business and technical point of view compared to other options. It would hardly make sense to shrink a massive chip to a new process late in the cycle unless they're still selling them by the truckload. Mainstream parts move a lot more units.

....i do not see how what you quoted relates to your post? did you miss quote another forum member perhaps?

[mattkosem]

That's awfully presumptuous. The decision of what product to test on a new process could have been any of the following:

1) Mainstream by own company standards (internal product selection)
2) Mainstream by global standards (market-based product selection if they're aiming for a certain kind of product range)
3) Mainstream by die size (technical product selection)

Nobody knows which of these is true, and #3 is the most likely.

Try it on something small/easier and also cheap. Then, sell it as a market mainstream so it actually makes you some money that late in the product generation cycle. That's what I would do in that position as it makes a lot of sense both from a business and technical point of view compared to other options. It would hardly make sense to shrink a massive chip to a new process late in the cycle unless they're still selling them by the truckload. Mainstream parts move a lot more units.

I think you read more than I said. I never made any claims about their selection process. I merely pointed out the fact that the first product on the 55mm node was, at the time, the highest end product on the ATI product line.

The section that you quoted was in response to someone stating that the 3870 was not a high-end card, which was obviously not the case in their product line. Am I missing something?

--Matt

[Behemot]

I merely pointed out the fact that the first product on the 55mm node was, at the time, the highest end product on the ATI product line.

Yeah, but it doesn't matter if it was high-end or whatever. As I pointed, it was part of their "sweet point" strategy, it was small (and 55nm proces had no such problems like 40nm), so no reason why to manufacture RV620 or RV610 sooner than RV670.

Not mentioning they needed something to beat NVIDIA at those times. HD 2900 XT was not bad compared to GTS 320/640 MB, the power consumption was pretty much the same, but you know, less is always better. For that reason RV670 came first.

[saaya]

those wafer shots show a test chip, im pretty sure...
its not a gpu or cpu, its way to heterogenous for that...

[Particle]

I don't honestly understand how there can be any confusion regarding my post, but if nobody else seems to get it I'll just leave it alone.

[=SOC= Admiral]

I might thing that its something that AMD had global foundries do because Intel just released the 32nm and then AMD can go and say we have 28nm in your face Intel. They had one wafer made remember I doubt anything on that wafer is working, it is the first one. and I am also pretty sure that they are releasing bulldozer the new architecture on 32nm and this just says that they really have better R&D now and that they are truly gunning for their old crown again. this means that the released from AMD will be getting faster and faster but also better and better almost equaling Intels speed. But one thing I think AMD should focus on improving is their manufacturing processes quality. Intel still has the best manufacturing process.

[Florinmocanu]

Bulldozer will be 32nm SOI, this has nothing to do with AMD shoving 28 nm waffers in Intel's face. stop thinking in a childish way.


28nm will not be adressed to CPU's, after 32nm we will probably see GF deveoping 22nm, otherwise we cannot imagine AMD being competitive agaisnt Ivy Bridge (refresh of Sandy Bridge in 22nm).

PS: Intel showed 22nm waffers as well, so the smaller these companies go, the better the products we will buy will be.

[=SOC= Admiral]

This is something I did not know so now I say thank you.

[Nedjo]

RV670@55nm, WAS NOT shrink of R600@80nm! It was new design. RV740@40nm WAS NOT shrink of RV770@55nm it was new design. Yes NEW DESIGN did keep same architecture philosophy, but that's about only thing same.

Beside CPUs, AMD do have APUs

[generics_user]

I might thing that its something that AMD had global foundries do because Intel just released the 32nm and then AMD can go and say we have 28nm in your face Intel. They had one wafer made remember I doubt anything on that wafer is working, it is the first one. and I am also pretty sure that they are releasing bulldozer the new architecture on 32nm and this just says that they really have better R&D now and that they are truly gunning for their old crown again. this means that the released from AMD will be getting faster and faster but also better and better almost equaling Intels speed. But one thing I think AMD should focus on improving is their manufacturing processes quality. Intel still has the best manufacturing process.

amd showed 28nm SRAM test wafers serveral months ago so i'm pretty sure that they can produce working ICs by now