bit-tech.net

ATI's HD 6000 series was originally 32nm

ATI's HD 6000 series was originally 32nm

ATI's Radeon HD 6000 series was originally destined for TSMC's 32nm node, until the company dropped it in favour of the 28nm process, due next year.

At AMD's Technical Forum and Exhibition 2010, we had a quick chat with Matt Skynner, Corporate VP and General Manager of AMD's GPU Division, about the decision to continue using TSMC's 40nm process for ATI's Radeon HD 6000 series. (We'll continue to refer to AMD's graphics division as ATI until Fusion launches.)

We asked if these new graphics cards launched too early, instead of late this year/next year when TSMC's 28nm process was ready.

Mr Skynner admitted that the HD 6000 series was originally set to use TSMC's 32nm process, but that AMD had to opt back to 40nm earlier this year after that process was unceremoniously dumped by TSMC in favour of concentrating on 28nm only. Excluding the yield issues with the more recent 40nm process, TSMC does have great experience in creating high performance, low power half-node processes suitable for GPUs at 110nm, 80nm and 55nm, and this is where 28nm looks to continue.

TSMC claims to have its 28nm process ready pretty much about now, which gave us ammunition to pressure Skynner over the perceived lack of gap between new Radeon HD 6000 series and future 28nm ATI Radeon parts.

The reply was not to TSMC's credit, and although the answer was not direct there was a strong inference that TSMC's 28nm roadmap has been pushed back towards the middle of next year. Ouch. If entirely dropping 32nm wasn't enough, it appears the problems of high-k, metal gate and EUV lithography are significant for a company that hasn't used them before. The next question is therefore how this affects Nvidia's refresh as well.

Other reasons for an October/November launch for the HD 6000 series included the holiday/Christmas season, which is obviously a ripe time for sales, regardless of industry.

Will TSMC's problems continue to be holding the graphics industry back? Let us know your thoughts in the forums.

15 Comments

Discuss in the forums Reply
Snips 22nd October 2010, 10:27 Quote
So who does AMD turn to as an alternative if they wanted to? Global Foundaries?
javaman 22nd October 2010, 11:34 Quote
Intel........joking aside GloFLo seems the only alternative. I'm gonna have to read Richards Article again tho I got the impression that it's not as simple as just turning to another company due to costs in upgrading the process and developing each technique. The more companies that use that process the better since the cost is split between all the companies. On top of that only certain companies can handle the high-k metal gate and EUV lithography too.
TWeaK 22nd October 2010, 11:34 Quote
This is really a shame, as I'm sure this was the ace AMD said they had up their sleeve ages ago before Fermi was brought out. It seems every time AMD gets an advantage (which they no doubt work bloody hard for) it gets taken away through no fault of their own.

@Snips I think the reason is simply that it takes time to do that. They'd already developed the architecture significantly, and it's much easier to just change the die size than the manufacturer. If they'd gone to GloFo they likely would have had months of delays. I'm sure that eventually AMD would like to get all their production moved over though, especially given these events.
azrael- 22nd October 2010, 11:36 Quote
Isn't this what you'd call old news? Although it's nice to know that AMD is saying it officially now.

Also, for anyone believing TSMC has its 28nm process worked out I have some prime beach front property in Florida for sale... ;)
Snips 22nd October 2010, 11:48 Quote
Why is it always "if only" with AMD. Why can't it be "Yes, we set out to do exactly this and you know what, it kicks ass!" I know the marketing sometimes suggests the latter but the end result rarely does unfortunately.

Sorry that's really negative. I'll go and stand myself in the naughty corner and not give me any sweats all day!
Tangster 22nd October 2010, 11:55 Quote
Quote:
Originally Posted by Snips
Why is it always "if only" with AMD. Why can't it be "Yes, we set out to do exactly this and you know what, it kicks ass!" I know the marketing sometimes suggests the latter but the end result rarely does unfortunately.

Sorry that's really negative. I'll go and stand myself in the naughty corner and not give me any sweats all day!

I'll turn the air con up shall I?
Snips 22nd October 2010, 11:58 Quote
lol sorry "Sweets"
Fizzban 22nd October 2010, 12:13 Quote
Senilex 22nd October 2010, 13:18 Quote
Quote:
Originally Posted by Fizzban
http://i56.tinypic.com/ngpsfp.jpg

LoL
Adnoctum 22nd October 2010, 13:48 Quote
AMD can't afford to tell TSMC to go forth and multiply, simply because in the short term there aren't any alternatives.

Last year there was a GF briefing that roadmapped their processes, and what is important for this conversation.
Fab 1 - Module 2 in Dresden is to be refitted for 28nm bulk (as used for graphics cards) for Q4 2010, but the capacity will be 25000 wafers per month (wpm) max. I can only guess that this would not be enough to satisfy AMD's HD7xxx series needs from top to bottom, especially if AMD is in competition with others for wafers (nVidia, maybe?).
GF Fab 8 (formerly Fab 2) will be 28nm bulk at 35000 wpm, but won't be ready until 2012, which will be too late for HD7xxx.

In case you are wondering, in the same briefing they said that Fab 1 - Module 1 will be 32nm SOI/40nm bulk starting 25000 wpm, and this year they announced that production at Fab 1 will be expanded from 50000 to 80000 wpm,, and that after Phase 2 expansion Fab 8 will move from 35000 to 60000 wpm. So in 2012, that is 140000 wpm on an advanced process (the former Chartered 200mm fabs are on older processes, with the 300mm Fab 7 to be refitted for 55/40nm).

TSMC in comparison has two 300mm fabs each starting 100000-120000 wpm right now, and they have just broken ground on a third of a similar capacity which will be ready for 2012 (and the new Fab 15 will cost TSMC...wait for it...US$9.4 billion. Yeah, now we can see why AMD is getting out of the fab business).
So, 300000-360000 300mm wpm in 2012.
GF has a long way to go before they can challenge TSMC on capacity.
rickysio 22nd October 2010, 13:50 Quote
nVidia probably went a little : http://i184.photobucket.com/albums/x34/rickysio/7b1a4f1d.gif

before realising that they were also relying on the 32nm shift to help Fermi.
alwayssts 22nd October 2010, 22:08 Quote
All I would say is to look at GF roadmap.

28nm: Mass Production in Q211, so with tape-out products could be available in Q411. (similar to TSMC)

28 HPP: Mass production in Q212, so with tape-out products could be available in Q412.

20nm: Mass production in Q213, so with tape-out products could be available in Q413. (similar to TSMC).


Now, I'm not saying AMD will use GF for graphics in 2011 because there are a lot of things to consider (TSMC versus GF gate approach, die size vs clock speed design decisions, ability to produce enough chips) but it's interesting that the timetables surely line up well with a yearly graphics refresh in Q4...something ATi has become known for.
technogiant 23rd October 2010, 08:49 Quote
About GF's ability to produce enough chips.
Was wondering if we will see a split with the 7000 series, we are seeing AMD rebranding the lower 5000 series as low end 6000 series this time around, wondering if in the 7000 series we may see just mid/high end chips made on 28nm at GF and the lower end rebranded 6000-to-7000 series still on 40nm made at TSMC, that would certainly help with the initial supply problem and enable AMD to start to transition to 28nm and GF earlier than waiting for full capacity.
Krayzie_B.o.n.e. 23rd October 2010, 21:01 Quote
AMD must be doing something right as we always seem to discuss AMD's manufacturing process and wafers and silicon and blah blah blah GloFlo vs. TSMC Deathmatch.

Where as Nvidia makes something that requires a nuclear reactor to power and LN2 to cool in the current age of electronics efficiency and Ozone preservation.
or
Intel Makes something that is never backwards compatible and requires a Swiss Bank account to purchase.

Trial and error then improve is the staple of manufacturing since the beginning of mass production.
ill regardless the end product (HD6000) so far performs dam awesome, uses less watts and produces less heat.

Can't wait to hear all the ball droppings of AMD's 28nm process (in the old news section) but will happy to see even better performance and efficiency.
drunkenmaster 26th October 2010, 01:04 Quote
Quote:
Originally Posted by Adnoctum
AMD can't afford to tell TSMC to go forth and multiply, simply because in the short term there aren't any alternatives.

Last year there was a GF briefing that roadmapped their processes, and what is important for this conversation.
Fab 1 - Module 2 in Dresden is to be refitted for 28nm bulk (as used for graphics cards) for Q4 2010, but the capacity will be 25000 wafers per month (wpm) max. I can only guess that this would not be enough to satisfy AMD's HD7xxx series needs from top to bottom, especially if AMD is in competition with others for wafers (nVidia, maybe?).
GF Fab 8 (formerly Fab 2) will be 28nm bulk at 35000 wpm, but won't be ready until 2012, which will be too late for HD7xxx.

In case you are wondering, in the same briefing they said that Fab 1 - Module 1 will be 32nm SOI/40nm bulk starting 25000 wpm, and this year they announced that production at Fab 1 will be expanded from 50000 to 80000 wpm,, and that after Phase 2 expansion Fab 8 will move from 35000 to 60000 wpm. So in 2012, that is 140000 wpm on an advanced process (the former Chartered 200mm fabs are on older processes, with the 300mm Fab 7 to be refitted for 55/40nm).

TSMC in comparison has two 300mm fabs each starting 100000-120000 wpm right now, and they have just broken ground on a third of a similar capacity which will be ready for 2012 (and the new Fab 15 will cost TSMC...wait for it...US$9.4 billion. Yeah, now we can see why AMD is getting out of the fab business).
So, 300000-360000 300mm wpm in 2012.
GF has a long way to go before they can challenge TSMC on capacity.

Theres something you're missing, TSMC can most certainly NOT make 100k wafers of 40nm.

Like GloFo, who have spent(are in the process of) building an 8billion fab in New York State thats frankly going to be the single best fab on earth, with the option of 2 more fabs on site to be started probably a lot sooner than expected, which will likely cost 4-5billion each. On top of that they are expanding Chartered Silicon, a company they bought off the Singapore government for a few billion, which is a series of smaller 200mm fabs that are due to be upgraded, or again are in the process of being moved to 300mm.

Right now GloFo are pumping out close to 80-90k a month I believe just out of the Dresden fabs.


TSMC started at the end of last year with 9000 wafers a month, max, at 40nm. 25k wafers on 40nm for GloFo would be massively massively more than AMD could ever use alone. GloFo are fully capable of producing everything AMD need, though its a big risk to move an entire range. Its more of a hedge your bets situation, any process at any fab can be delayed.

Its rumoured, and quite likely that AMD will do some production on GloFo of the 7xxx series, and some at TSMC. The designs do need to be different, but not massively so. Its more layout of transistors as opposed to actual architectural difference.

Each individual chip needs its mask to be "taped out" , so it really won't cost much more to have chips made for different processes at different plants, because, each chip needs a team that works out the best arrangement of transistors, which design you do it for doesn't matter "that" much.

AS you said, one of the main reasons for the GloFo spin off was its very hard for AMD to take orders from Arm for cpu's that will eventually compete with their own products, while a 3rd party producing chips for both is very viable. In that way, manufacturing costs get spread across a massively higher production scale, which makes it much cheaper, and quicker.

AMD actually did phenomonelly well to stay in as little debt as they did. They probably spent between 4-6billion every couple years on process R&D, equipment and fab upgrades, but while Intel spread that across more labs and 80% of the market sales, AMD did it on 20%, 2 fabs and had logistical problems from only having two fabs at max capacity.

Intel for a new process increase capacity for a while at 3 fabs, shut another down, upgrade to new process, and ramp up production, then move onto the next fab, etc, etc. AMD are at pretty much 100% capacity on both fabs, shutting down one to upgrade hurts them very badly, Intel upgrade to a new process the second its yields are good enough, AMD do it at the last second they can't wait anymore because theres no longer a choice.

This is again where GloFo will come through, because 3rd high capacity fab in NYS, Singapore operations, a several billion new fab in Dubai fairly soon, and another two modules at NYS in the near future mean switching to new processes will happen as quickly as Intel manage it.

TSMC are big though, with a lot of lead time.

It seems quite possible GloFo will be first to 28nm and real production volume this round, but if its out early will they be ready for it? If AMD/Nvidia are planning new products around August-Oct next year, it might not matter if one is a few months earlier in the year.

Thing is, TSMC has a long long track record of being late, and providing worse quality than they promise, with some severe problems of late.
Log in

You are not logged in, please login with your forum account below. If you don't already have an account please register to start contributing.



Discuss in the forums