With the recent news of ATI's upcoming GPU family in the works, and nVidia desperately trying to salvage Fermi dies from a seemingly flawed design, I thought it prudent to apply the same (nearly spot on) logic to Southern Isalnds, that was applied to estimating a release date for nVidias troubled launch. We shall call this the
Demerjian Conjecture for the purpose of this post.
It has been reported that ATI's upcoming Southern Islands GPU
has taped out. This article was posted April 21, 2010 and while it is possible the tape-out was completed days, or even a week or two prior to the announcement, we'll use the Apr 21 date as a worst case scenario here.
Source of following numbers.
Once tape-out has occured it is on to the 6-8 week process (we'll call it 7 weeks) to run the first design through TSMC's machines and get back the first 'hot lot'.
If everything works the first time around, production silicon can start to be produced. This process takes 10-12 weeks (we'll call it 11 weeks) to get parts back.
Once production silicon is recieved it takes approximately 2 more weeks to slap it on a board, box it up and ship it out.
This gives us a best case scenario of 20 weeks from April 21st to the public, hard-launch of a Southern Islands board(s). (This would be the second week of September 2010 for those counting)
Assuming all is not perfect the first time around however and a respin is necessary to fix some errors, boost clock rates, improve yields, etc... each spin costs another 8 weeks to produce.
It has been posted in these forums that some believe the launch variant of Cypress to be the
A13 stepping of the GPU core. This would seem to indicate that two respins were necessary after the initial 'hot lot' was received.
If we assume that S.I. will also require 2 metal layer revisions before it is ready for retail production, that would push our launch date back by another 16 weeks to the very last week of December 2010, missing Christmas.
If the worst happens and a base layer respin should be necessary, that would
well over a quarter to pull off, we'll give it a value of 16 weeks for this example (a quarter is 13 weeks)
Now, from all the above we arrive at the following simple equation:
Demerjian Conjecture:
GPU Launch Week = t + 8(r) + 16(b) + 20
Where: t = tape out date week, r = the number of metal layer revision required, and b = the number of base layer respins required.
So, using our equation, assuming 2 metal layer revisions before launch (same as Cypress) and knowing that April 21st would be Week 16 we arrive at the following:
GPU Launch Week = 16 + 8(2) + 16(0) + 20 = 52 (Last week in December)
Since this will be essentially the third generation of GPU's produced on the 40nm process at TSMC, perhaps we can assume that ATI has a good grip on the process by now and will be able to produce S.I. with only 1 metal layer revision before production. This would yield:
GPU Launch Week = 16 + 8(1) + 16(0) + 20 = 44 (First week of November)
So to Recap:
No revisions (A11): 2nd week September, 2010 [20 weeks]
1 Revision (A12): 1st week November, 2010 [28 weeks]
2 Revisions (A13): Last week of December, 2010 [36 weeks]
From all I've read so far, I believe southern islands will be a bit better than 5870 but with much better tesselation. I have no doubts that Northern Islands will use a different archetecture (different shaders) than Cypress or Southern Islands. We won't really know anything about Northern Islands (save fab process, things like that) until we see SI.
