AMD: R9xx Speculation

[Jawed]

Code Rename

http://www.digitimes.com/news/a20100811PD223.html

AMD originally planned to have its next-generation GPUs using 32nm process and codenamed the chips Northern Islands, but Taiwan Semiconductor Manufacturing Company (TSMC) skipped its 32nm R&D for GPUs and advanced directly to 28nm R&D. In response AMD adjusted its plans and decided to continue adopting a 40nm process, while changing the product's codename to Southern Islands, the sources explained.

So we shouldn't be expecting SI and NI, merely that NI has been renamed SI.

One of the things that's been bugging me for a while now is that Gipsel's reports on revised instructions explicitly refer to NI. With rumours consistently referring to SI as the next family, the idea that NI instructions would be "in the drivers" seems peculiar if NI is supposedly much later than SI.

So, treating SI as NI renamed seems to make everything fit together. SI will have NI's instructions and merely be 40nm instead of 32nm.

 

[TKK]

http://www.digitimes.com/news/a20100811PD223.html


So we shouldn't be expecting SI and NI, merely that NI has been renamed SI.

One of the things that's been bugging me for a while now is that Gipsel's reports on revised instructions explicitly refer to NI. With rumours consistently referring to SI as the next family, the idea that NI instructions would be "in the drivers" seems peculiar if NI is supposedly much later than SI.

So, treating SI as NI renamed seems to make everything fit together. SI will have NI's instructions and merely be 40nm instead of 32nm.

Makes sense. If you think about it, porting a 32nm design to 40nm is probably easier and faster to do than the Evergreen-NI hybrid some were expecting SI to be.

 

[Silent_Buddha]

So, basically that would mean...

SI - 40 nm leveraging stuff done for NI.
NI - 32 nm - possibly cancelled.
?? - 28 nm - new gen?

Regards,
SB

 

[Kaotik]

So, basically that would mean...

SI - 40 nm leveraging stuff done for NI.
NI - 32 nm - possibly cancelled.
?? - 28 nm - new gen?

Regards,
SB

IMO more likely:
SI - 40 nm leveraging stuff done for NI.
NI - 32 nm - cancelled.
NI - 28 nm - refresh H1/11

 

[Jawed]

Still have Hecatoncheires as the wild card code name...

 

[trinibwoy]

IMO more likely:
SI - 40 nm leveraging stuff done for NI.
NI - 32 nm - cancelled.
NI - 28 nm - refresh H1/11

Yeah that's what I'm thinking.

Still have Hecatoncheires as the wild card code name...

Hecatoncheires is to NI as Cypress is to Evergreen?

 

[caveman-jim]

except cypress is an evergreen tree. can't find hekatonchieres on map any where...

If we speculate that
NI was the 32nm 2010 generation (plan A)
SI was the 40nm 2010 generation (plan B)
Hakontonchieres was 2011 generation (22nm? 28nm?) plan A

Q4 2010 release is SI. Whatever was in 2010 Plan A is file 13, we go straight to 2011 plan B. SI could be a (downward) bump on the trend line of year on year 'exceeding moores law' for perf/watt that AMD were excited about at Cypress launch. But Heka could be the compensating uptick.

A refresh in H1 2011 doesn't make much sense to me, steps on the toes of all the product refreshes and timed wrongly for market bumps, except US tax return season.

So for 2010 generation release, rumors put mainstream launching Q4 10 with high end Q1 11. Which could be 40nm SI mainstream, 28nm NI enthusiast. But that ties the high end to a new fab process without a nice test part first - risky.

Whichever way it goes I hope AMD pay more attention to the $150-$300 market this time. HD 5770 - 5830 - 5850 left gaps you can drive a steamroller through, and NVIDIA dropped a Fermi powered one right in it (GTX 460).

 

[Jawed]

RV740 was sort of an RV730 refresh, pipe-cleaning at the same time. (Though it's worth noting that RV740 was late - as much as 2 quarters.)

After NI launched this autumn, it seems reasonable to expect the original plan was to refresh with a pipe-cleaner on 28nm in 2011H1. The full replacement family would then hit in autumn 2011, one year after NI.

This refresh/pipe-cleaner chip would have 128-bit bus and be faster than "HD6770". HD6770, on 40nm, might be expected to be around 215mm² (~30% bigger than Juniper) and so around 140mm² on 32nm. The refresh of this would come in at around 110mm² I suppose, on 28nm. That's big enough for a 128-bit bus (Redwood is 104mm²) - though the Anandtech "story of Evergreen" article implies that a little larger than this would be preferable for a pipe-cleaner, e.g. another 20-30%.

I still think SI is a replacement of NI, purely because of 32nm being cancelled. I suppose there's a chance it lost features/capability - in much the same way as Evergreen. The stuff that was lost in Evergreen should appear in SI I reckon.

Hecatoncheires might be a hybrid of NI and the next family that's ~15 months away. RV740 is a hybrid in this sense as the ROP/MCs are re-worked for double fillrate, a preview of Evergreen and there are some ALU changes.

 

[caveman-jim]

I think we're saying the same things just different ways

 

[Jawed]

Whichever way it goes I hope AMD pay more attention to the $150-$300 market this time. HD 5770 - 5830 - 5850 left gaps you can drive a steamroller through, and NVIDIA dropped a Fermi powered one right in it (GTX 460).

I think that's because HD5850 was expected to hit $200 pretty rapidly, so HD5770, HD5830 and HD5850 would have been compressed into a range of ~$75, say.

 

[Jawed]

I think we're saying the same things just different ways

I don't. Specifically I disagree with: "A refresh in H1 2011 doesn't make much sense to me, steps on the toes of all the product refreshes and timed wrongly for market bumps, except US tax return season." and elucidated why. The potential mirroring of RV740 is too striking to ignore.

 

[rpg.314]

I think that's because HD5850 was expected to hit $200 pretty rapidly, so HD5770, HD5830 and HD5850 would have been compressed into a range of ~$75, say.

Yeah. I doubt they were expecting the margins they are getting right now.

 

[Alexko]

They probably weren't expecting to still be supply-constrained on 40nm ~16 months after the process was introduced either…

 

[no-X]

(a few months ago? - I don't remember it exactly) Dave said, that the situation will be better, but demand will be still higher than supplies - so they were aware of this problem.

 

[rpg.314]

They probably weren't expecting to still be supply-constrained on 40nm ~16 months after the process was introduced either…

IIRC, tsmc announced sometime ago that they'll ramp up 40nm capacity in Q3, going up to 160k wpm.

 

[caveman-jim]

I don't. Specifically I disagree with: "A refresh in H1 2011 doesn't make much sense to me, steps on the toes of all the product refreshes and timed wrongly for market bumps, except US tax return season." and elucidated why. The potential mirroring of RV740 is too striking to ignore.

A single product doesn't make a refresh. HD 4770 came out, sold out, then 4830 and 4670 dropped in price to cover the gap so nobody missed it. We might see a 6790 on new process but not a whole generation refresh.

 

[Alexko]

(a few months ago? - I don't remember it exactly) Dave said, that the situation will be better, but demand will be still higher than supplies - so they were aware of this problem.

Well yeah, but when they designed Evergreen and decided on their strategy, they probably didn't know it would turn out this way.

 

[eastmen]

Whats everyone's thoughts on ati adopting 2gig framebuffers on the $300-$500 range products this coming cycle ? They are pushing eyefinity to users and I don't think 1 gig cards will cut it in future games.

 

[Silent_Buddha]

I think that's because HD5850 was expected to hit $200 pretty rapidly, so HD5770, HD5830 and HD5850 would have been compressed into a range of ~$75, say.

Aye, I would expect that AMD just like everyone else at the time thought that Nvidia would be able to get Fermi out within 3 months of Cypress and that it would provide significant competition. Enough that it would forced them to adjust prices downwards. Figured they'd get a least a few months of high margin sales.

But TSMC's 40 nm wafer woes + Nvidia's inability to get a competitive Fermi product out for over half a year conspired to keep 5850 and 5870 prices unnaturally high. And once Fermi did come out, it was somewhat competitive (performance decent, but very high power consumption), but unable to be produced with high enough yield or quantity to impact sales enough.

It was just basically a really really abnormal product cycle. Had it followed a normal product cycle with competition + sufficient wafer supplies, I suspect that 5850 would have ended up in the 200-225 range with 5870 anywhere from 279-329 and 5830 around 149-179.

Regards,
SB

Regards,
SB

 

[Alexko]

Whats everyone's thoughts on ati adopting 2gig framebuffers on the $300-$500 range products this coming cycle ? They are pushing eyefinity to users and I don't think 1 gig cards will cut it in future games.

Yeah I'm pretty sure the high-end will come with 2GB. There might be 1GB versions as well, but I think 2GB is a given for that market.