RV850 + Samplers per SIMD?

[ElMoIsEviL]

Juniper + Samplers per SIMD?

I'm curious to know.

As I understand it Cypress maintains 20 Samplers per SIMD and I do know that Juniper is a cut down variant of Cypress (nearly half of Cypress in most respects).

Does Juniper maintain 20 Samplers per SIMD or is it cut down to 10?

Thanks.

 

[Dave Baumann]

<sigh> Please, there is no such thing as RV850. Instead of making up numbers, we should use their proper codenames - Cypress and Juniper!

 

[ElMoIsEviL]

<sigh> Please, there is no such thing as RV850. Instead of making up numbers, we should use their proper codenames - Cypress and Juniper!

All right.. Juniper.

How many Samplers per SIMD for Juniper?

 

[Dave Baumann]

Exactly the same as Cypress.

 

[trinibwoy]

16 man. Don't know why Dave is being so mean though.

 

[ElMoIsEviL]

Dave shouldn't be mean considering I am a former ATi employee myself.. lol (from way back when).

I'm simply working towards creating my own block diagrams in an article comparing subsequent generations of ATi/AMD GPU architectures (so the entire lineup from R600 to Cypress).

16 Samplers while Dave says 20... lol

So which one is it

 

[trinibwoy]

Where did Dave say 20?

 

[OpenGL guy]

Dave shouldn't be mean considering I am a former ATi employee myself.. lol (from way back when).

I'm simply working towards creating my own block diagrams in an article comparing subsequent generations of ATi/AMD GPU architectures (so the entire lineup from R600 to Cypress).

16 Samplers while Dave says 20... lol

So which one is it

Your initial post is wrong. Cypress has 1 sampler per SIMD. Each sampler handles 4 pixels per clock, so effectively 80 texture lookups per clock since Cypress has 20 SIMDs. Juniper also has 1 sampler per SIMD, but only 10 SIMDs, so Juniper has the same number of samplers per SIMD as Cypress.

 

[trinibwoy]

Depends on what you consider a sampler. In normally accepted (layman) terms it's got 4, with each capable of one filtered or 4 unfiltered samples per clock. Hence 4 filtered or 16 unfiltered texels per SIMD.

 

[ElMoIsEviL]

Your initial post is wrong. Cypress has 1 sampler per SIMD. Each sampler handles 4 pixels per clock, so effectively 80 texture lookups per clock since Cypress has 20 SIMDs. Juniper also has 1 sampler per SIMD, but only 10 SIMDs, so Juniper has the same number of samplers per SIMD as Cypress.

That makes more sense. So Juniper has 10 Samplers total (1 per SIMD). I was rather confused when looking at the Block diagrams from this website in the Cypress architecture article. That's my fault as I didn't take the "," into consideration (without the "," it seems to indicate 20 Samplers for each SIMD).

Thanks guys

 

[OpenGL guy]

Depends on what you consider a sampler. In normally accepted (layman) terms it's got 4, with each capable of one filtered or 4 unfiltered samples per clock. Hence 4 filtered or 16 unfiltered texels per SIMD.

80 texture units is usually how Cypress is presented as few care about the unfiltered rate.

 

[Dave Baumann]

That makes more sense. So Juniper has 10 Samplers total (1 per SIMD). I was rather confused when looking at the Block diagrams from this website in the Cypress architecture article. That's my fault as I didn't take the "," into consideration (without the "," it seems to indicate 20 Samplers for each SIMD).

In a very high level sense Juniper is a straight chop down the middle - half the SIMD's, half the ROPs, half the mem-I/F. So, at a per SIMD level things stay the same. Of course, where it doesn't change is the triangle and tesselation rate!

 

[trinibwoy]

 

[ryta1203]

There are 16 thread processors per SIMD with 4 texture fetch units per SIMD.

As far as I am aware the difference in cards (as far as SIMDs go) is scaling. I'm sure there is some other minor differences though (like in LDS between 870 and 770).

 

[fellix]

I'm sure there is some other minor differences though (like in LDS between 870 and 770).

Careful with that!
Dave is watching...