Turks speculation thread

[integrated]

Originally Posted by Wandrey

Agreed. If drivers improve performance on VLIW4, which is very likely, a 768sp part would be on par or even better than Barts Pro. Normally, overlapping products is not desirable. So a 10 SIMD (640sp) Turks makes more sense, in a way that it would be cheaper to manufacture and would align better on product portfolio. I only disagree with the Pro variant, I believe they will disable only 1 SIMD, regardless of the total amount, just as they did with smaller Evergreen chips (Juniper and Redwood).

Don't the Cayman-family (dual-graphics architecture) need to add SIMDs in pairs?

How many parts will they be squeezing into the low-end anyway?

6700 seems obvious below Barts @ 640. Could sell at 139-179.
6600 @ 512, using weak parts from 6700 would still be much faster than Llano in the high-volume 100-120 zone.
6500 @ 384 should still beat Llano significantly, especially considering BW, and sell shedloads at <$100.
6400 @ 256 (weak 6500s?) as the low-end retail part.
6300 @ 128 as the ultimate OEM special, or would AMD bother making something smaller?
Aren't the cards more expensive than the GPUs at this point?

5-chip lineup? Seems like a lot...

[eRacer]

Originally Posted by integrated

6700 seems obvious below Barts @ 640. Could sell at 139-179.

Radeon 6950 has 12% fewer shaders than the Radeon 5870 and are roughly equal performance.
Turks @ 640 would have 20% fewer shaders than Radeon 5770 so why would Turks be worth purchasing over a cheaper 5770?

The way the 6000 series has turned out Turks 6770 could be 640SP and a bit slower than the 5770. That would certainly fit the pattern of 6870<5870 and 6970<5970.

[fusion]

Originally Posted by eRacer

Radeon 6950 has 12% fewer shaders than the Radeon 5870 and are roughly equal performance.
Turks @ 640 would have 20% fewer shaders than Radeon 5770 so why would Turks be worth purchasing over a cheaper 5770?

The way the 6000 series has turned out Turks 6770 could be 640SP and a bit slower than the 5770. That would certainly fit the pattern of 6870<5870 and 6970<5970.

The 6850 is also clocked 50mhz slower than the 5870. Drivers will probably be more optimized when it comes out too. Though personally i think it'll still be a VLIW5 chip with a smaller die but same performance as juniper.

[Drunkenmaster]

Originally Posted by eRacer

Radeon 6950 has 12% fewer shaders than the Radeon 5870 and are roughly equal performance.
Turks @ 640 would have 20% fewer shaders than Radeon 5770 so why would Turks be worth purchasing over a cheaper 5770?

The way the 6000 series has turned out Turks 6770 could be 640SP and a bit slower than the 5770. That would certainly fit the pattern of 6870<5870 and 6970<5970.

Why would the 5770 be cheaper, thats the flaw in your assumption.

Cypress to barts, efficiency, 30% smaller MUCH cheaper to make. Heres a hint, a 640 Shader "barts" type core would be MUCH much smaller than a 5770.

LIkewise the 1120 shader Barts is very close to the 1600 shader Cypress, which has over 40% more shaders. 640 x1.4=circa 900 shaders. It would well be faster, much faster in tesselation benchmarks and signficantly smaller meaning they can price it cheaper or similar price for more performance. Considering yields, die size and power it should be pretty easy to be smaller, faster and have higher clocks while hitting the same TDP, it could be 15-20% faster and still 30% + cheaper to make for AMD.

At this stage it looks increasingly likely Cayman took a LOT of time, effort and money to test at 40nm, the lower parts of the line will almost certain be "barts" based and focus on efficiency, increasing performance/mm2, and fitting in against Nvidia's "fixed" cores via lowered costs or improved performance, see where they all end up basically.

[eRacer]

Originally Posted by Drunkenmaster

Why would the 5770 be cheaper, thats the flaw in your assumption.

Radeon 5770 is already selling for under the $139-$179 suggested price I was replying to.

Cypress to barts, efficiency, 30% smaller MUCH cheaper to make. Heres a hint, a 640 Shader "barts" type core would be MUCH much smaller than a 5770.

LIkewise the 1120 shader Barts is very close to the 1600 shader Cypress, which has over 40% more shaders. 640 x1.4=circa 900 shaders. It would well be faster, much faster in tesselation benchmarks and signficantly smaller meaning they can price it cheaper or similar price for more performance. Considering yields, die size and power it should be pretty easy to be smaller, faster and have higher clocks while hitting the same TDP, it could be 15-20% faster and still 30% + cheaper to make for AMD.

I was replying to someone who believed Turks would be VLIW4, unlike Barts.

Also, Radeon 5770 isn't Cypress. Cypress with 1600 SPs isn't twice as fast as Radeon 5770 with 800 SPs (and only half the memory bandwidth). So you can throw out a good share of your Barts vs. Radeon 5870 efficiency advantage because Juniper is also more efficient than Cypress.

[integrated]

If Turks goes with VLIW4, it will likely run with 768 shaders.

I was trying to puzzle out the market segmenting AMD will use and I got in over my head

[charlie]

Originally Posted by Wandrey

I will bet on half Cayman.

768 sp on Turks XT and 720sp on Turks Pro.

Still, I don't want to bet on the name just yet. Who knows if turks will the 6700 or 6600 series?

No, Cayman is the only VLIW4 chip until 28nm.

-Charlie

[Anyone]

So, if Turks is a 28-nm pipecleaner as that is the obviously hanging fruit, it will have VLIW-4 shader, right? 1/2 Cayman?

[eRacer]

Originally Posted by Anyone

So, if Turks is a 28-nm pipecleaner as that is the obviously hanging fruit, it will have VLIW-4 shader, right? 1/2 Cayman?

Turks is scheduled for Q1 2011 which seems early for 28-nm. If I had to guess it will be 640SP, VLIW5, 40-nm. According to Turks location on this roadmap it will be slower than the Radeon 5770.

[sulphademus]

Originally Posted by Nerdmaster

Yeah probably 6700 series. Maybe fusion needs naming space hence amd has to name their cards in a weird manner.

May be weird but AMD's is a $#%^ton better than Intel's scheme.

Originally Posted by integrated

Don't the Cayman-family (dual-graphics architecture) need to add SIMDs in pairs?

How many parts will they be squeezing into the low-end anyway?

6700 seems obvious below Barts @ 640. Could sell at 139-179.
6600 @ 512, using weak parts from 6700

5-chip lineup? Seems like a lot...

Sounds about right to me. I would think a smaller chip than Juniper but maybe 10% quicker?

There is still a market for low end GPUs. It will be a while before Intel's actuallyfinallyworthadamn graphics and AMD's Evergreen based APUs become ubiquitous. The market will start shrinking in 2011 and probably be nill by 2015. Until then you will still have some people buying for prior systems.

Which island is under Turks again? Probably VLIW5 160 SPs.

EDIT: And 2015 is when AMD aims to finally exterminate the DB15 VGA port. Hmmm...