AMD: R9xx Speculation
- Thread starter Lukfi
- Start date
As I mentioned earlier, if it was supposed to be sweet-spot @ 32nm (let’s shoot low here and assume something along the lines of 280mm2 instead of RV870's rather big 330mm2), it would (at least based on very simplistic mathematical calculation) end up with around 440mm2 @ 40nm.
That doesn't factor in process-specific peculiarities, of course - but in suggesting the 400mm2 mark, I actually already gave AMD’s engineers a lot of credit for possibly reducing the relative size of the (yield-optimized, low-risk) 32nm design by about 10% due to their expert knowledge of the 40nm process
@Mindfury: Looks like those chiphell-guys can't really decide on Caicos' bus width, he?
Last edited by a moderator:
Well pictures and info from a source known for reliable leaks trumps your doubts in this case but it's only a matter of time until we know the real deal. And what makes you think 6xxx isn't Evergreen plus a few extra-chip goodies - I thought that was what we expected?
As I mentioned earlier, if it was supposed to be sweet-spot @ 32nm (let’s shoot low here and assume something along the lines of 280mm2 instead of RV870's rather big 330mm2), it would (at least based on very simplistic mathematical calculation) end up with around 440mm2 @ 40nm.
That doesn't factor in process-specific peculiarities, of course - but in suggesting the 400mm2 mark, I actually already gave AMD’s engineers a lot of credit for possibly reducing the relative size of the (yield-optimized, low-risk) 32nm design by about 10% due to their expert knowledge of the 40nm process
@Mindfury: Looks like those chiphell-guys can't really decide on Caicos' bus width, he?
I love your posts and I have read all of them.
You seem to be basing your whole hypothesis, that there will be a 1:1 ratio of Evergreen performance/die size related to Islands, though. I am not an engineer or even so much tech savvy as the other guys here, so I perfectly understand your thinking and your descriptions. In all honesty, it makes perfect sense.
Hopefully, AMD has perfected their efficiency and not the same amount of transistors is required for the same level of performance. Yeah I know this does not sit well when comparing with Nvidia's perf/die size, but who knows,, maybe AMD has surpassed themselves.
How about AMD requiring 75% of the transistor count, for Islands to give 100% of Evergreen performance? In such a case, there would be no need for a 440mm2 chip.
Anyone care to comment on this? Particularly the SP redesign. The author claims a reduction to 4-wide SPs which allows ATi to fit an additional 25% SPs into each chip. If true, I find it a bit odd as it would give them slightly less mathematical power at the same die size. Cypress: 320x5 = 1600, Cayman: 320(*1.25)x4 = 1600. Seems a bit odd to me, unless there's more to the story or perhaps more SIMDs will be used, increasing the die size.
It can still be worth it if the overall utilization of the new ALU's is higher. This would particularly help those shaders which had lower utilization on older ALU's.
If the rumored parallel setup/raster is true, then 25 simd's is an odd proposition. I'd have expected something that is a multiple of setup/tess units.
The quoted post says 4 of the new ones would have 98.5% the performance of 5 old ones. So they would increase perf./area.
Supposedly they are less efficient @ 98.5% throughput compared to the old SP design.
Not so. 25% reduction in SP width is canceled out by 25% increase in SPs, hence my request for comment.
Last edited by a moderator:
Maybe I'm just too optimistic, but I understood it as "one new 4D is 98.5% as efficient (clock-to-clock) as old 4+1D", but takes only 75% of the space
That makes sense, guess I shouldn't drink before reading tech news
I can't find 98.5% mentioned anywhere on that post. Also, not sure how he came with that number.
I doubt the "new shader model" part.
Does anyone remember when D3D10.1 was announced? There is not even a hint about 11.1 yet. Also 10.1 brought some very minor improvements mostly for API orthogonality. D3D11 doesn't seem to need such a minor version upgrade as far as I can tell yet.
Does anyone remember when D3D10.1 was announced? There is not even a hint about 11.1 yet. Also 10.1 brought some very minor improvements mostly for API orthogonality. D3D11 doesn't seem to need such a minor version upgrade as far as I can tell yet.
Sound_Card
Regular
I also understood it as 98% as effective per clock. But 25% smaller. Now that is pretty impressive. So that means... for the same amount of space as Cypress's shader core, Cayman can fit 400sp's over Cypress's 320.
More like 66,5 mm².
May be he was just referring to rumored re-organization of ALUs.
NV hasn't even gotten their entire dx11 line out the door. Jeez, give 'em a break, will ya?
Similar threads
