AMD: R8xx Speculation

[Silent_Buddha]

I thought there were ways to get DXVA under VMR9 even in Win7? I haven't actually looked to see if it's possible or not. Doesn't matter too much on my HTPC though since it's using WHS which is based off the same kernel as XP, so I have to use DXVA through VMR9.

Regards,
SB

 

[swaaye]

Well I think that VMR9 works pretty much the same from a quality perspective as EVR.

If you are watching anything other than Bluray, check out Haali Renderer or madVR. madVR isn't as stable but it might be best. But you don't have any choice other than VMR9 or EVR with BDs.

 

[GZ007]

From fudzila http://www.fudzilla.com/content/view/17745/1/.
The planned launch date and NDA lift is 25th February 6.01 AM CET and late night on the 24th of February on 21.00 PM Pacific time for the 5830.
They posted this too http://buy.fudzilla.com/a505712.html.

 

[Silent_Buddha]

Well I think that VMR9 works pretty much the same from a quality perspective as EVR.

If you are watching anything other than Bluray, check out Haali Renderer or madVR. madVR isn't as stable but it might be best. But you don't have any choice other than VMR9 or EVR with BDs.

VMR9 as well EVR both utilize whatever media splitter you want. And the rendering quality is fine for me, haven't had any complaints bout them for the past 5+ years. Heck, Windows Media Player is perfect for me, other than fast forward, which is why I prefer GOM player for most playback (4x fast forward). MPC-HC is just my backup player for things that don't work well in GOM.

If WMP had the same playback features as GOM player I'd go with WMP fulltime myself. For the HTPC I'm still waiting on XBMC to mature on Windows, still not quite up to snuff.

Regards,
SB

 

[TKK]

From fudzila http://www.fudzilla.com/content/view/17745/1/.
The planned launch date and NDA lift is 25th February 6.01 AM CET and late night on the 24th of February on 21.00 PM Pacific time for the 5830.
They posted this too http://buy.fudzilla.com/a505712.html.

Well, the German magazine GameStar contains a review already.
Their numbers don't quite make sense though: According to them, the card has 1280 SP/64 TMU as predicted, and even a core clock of 800, not 750.

Now with those specs, I would expect the gap between 5830 and 5850 to be rather small, clearly smaller than the gap between 5850 and 5870. But according to their benchmarks, the opposite is true. Granted, they tested only Crysis in DX10 and Dirt 2 in DX9, but in those tests, the card was much slower than the 5850, slower than the GTX 285, even slower than the GTX 275 and just about the same as the 4890, and not all that much faster than a 5770.

These results are puzzling at best. Only 160SP/8TMU less than a 5850, but same bandwidth and a ~10% higher core clock should result in a less than 10% difference normally (more like 5% IMO), but the real world difference is much bigger. In Crysis at 1920x1200 + 4xAA/16xAF the 5850 is a whopping 20% faster, for example. And the 5830 is only ~10% faster than a 4870 512MB here. Something is quite fishy about those scores... Either there's some driver problems, some throttling, or the cut ROPs and memory interface in half. Those are the only reasons I can think of that would put some sense into these numbers. Neither the number of SIMDs, nor the clock speeds can be the culprit(s) here.

 

[Jawed]

Interesting numbers even if they're fishy. Since Cypress is "two shader blocks", each of 10 SIMDs, what we might be seeing there is a GPU with four SIMDs turned off in one of the shader blocks. Such an unbalanced configuration would be problematic...

Jawed

 

[Babel-17]

I was under the impression that the memory of the rumored HD 5830 was to be significantly slower than that of the 5850 and the 5770.

Edit: Sorry, I see now that my impression was based on older rumors.

 

[GZ007]

Interesting numbers even if they're fishy. Since Cypress is "two shader blocks", each of 10 SIMDs, what we might be seeing there is a GPU with four SIMDs turned off in one of the shader blocks. Such an unbalanced configuration would be problematic...

Jawed

And why would it be a problem ? It shouldnt make much difference for the dispatch processor if it has 16 , 18 or 20 SIMD-s at all. The gpu is balancing work per SIMD-s the blocks shouldnt have any meaning there beside the fact that they needed to place the SIMD-s on the die some accesible way. There isnt any die shot to see it unfortunately.

 

[Jawed]

And why would it be a problem ? It shouldnt make much difference for the dispatch processor if it has 16 , 18 or 20 SIMD-s at all. The gpu is balancing work per SIMD-s the blocks shouldnt have any meaning there beside the fact that they needed to place the SIMD-s on the die some accesible way. There isnt any die shot to see it unfortunately.

Rasterisation is split equally across the two shader blocks and that splitting in ATI is done according to screen-space tiles. I presume these tiles are shared equally by the two halves, despite the fact one half is significantly hobbled.

Any render target (e.g. a shadow buffer) cannot be completed until the slower half has finished its tiles. The faster half can't help with those tiles. The slower half just has to struggle through.

This means the faster half runs pixel shading at the same effective speed as the slower half. Other shaders (e.g. compute shaders or vertex shaders) shouldn't experience much difference - it's a matter of cache behaviour. Can't think of anything else that might be relevant.

Of course, this is just a theory.

Jawed

 

[Kaotik]

No-one can deny there's SOME driver problems at least, otherwise 2x HD5850 in CF couldn't be faster than 5970, while in some benches the HD5850 combo is faster

 

[GZ007]

Rasterisation is split equally across the two shader blocks and that splitting in ATI is done according to screen-space tiles. I presume these tiles are shared equally by the two halves, despite the fact one half is significantly hobbled.

Any render target (e.g. a shadow buffer) cannot be completed until the slower half has finished its tiles. The faster half can't help with those tiles. The slower half just has to struggle through.

This means the faster half runs pixel shading at the same effective speed as the slower half. Other shaders (e.g. compute shaders or vertex shaders) shouldn't experience much difference - it's a matter of cache behaviour. Can't think of anything else that might be relevant.

Of course, this is just a theory.

Jawed

It still doesnt mean that the screen space tiles need to be equal in case of 6-10 parts and that they cant be changed. I think the card wouldnt make muchs sense at 230-240 EUR with that performance. The 5850 doesnt show any difference with 8-10 SIMDs. AMD surely thinked ahead in case of disabled SIMDs.

 

[rpg.314]

Surely, AMD can think this through and has taken care to disable SIMD's equally from both blocks. :?

 

[GZ007]

Yep maybe there will be only chips with only 8-8 SIMDs.
But they could salvage more parts with combinations up to 6-10 if it wouldnt hurt performance. With improving yields with time the equal SIMD blocks looks like the best way to go.

 

[CarstenS]

Well, the German magazine GameStar contains a review already.
Their numbers don't quite make sense though: According to them, the card has 1280 SP/64 TMU as predicted, and even a core clock of 800, not 750.

Nice one by Gamestar. After their preview titled "Geforce GTX 380" they're wrong on the Ati side of things again?

 

[Jawed]

It still doesnt mean that the screen space tiles need to be equal in case of 6-10 parts and that they cant be changed. I think the card wouldnt make muchs sense at 230-240 EUR with that performance. The 5850 doesnt show any difference with 8-10 SIMDs. AMD surely thinked ahead in case of disabled SIMDs.

My understanding of the architecture leads me to believe that the screen-space tiles cannot be anything other than equally split.

We don't know that HD5850 is 8-10, it could be 9-9. AMD might disable SIMDs evenly even when only one SIMD needs disabling due to defects.

Another possibility is HD5830 has only 16 ROPs.

Jawed

 

[CarstenS]

Has there been any hint discovered if Evergreen has intra-SIMD redundancy the way RV770 had? Just wondering.

 

[TKK]

Nice one by Gamestar. After their preview titled "Geforce GTX 380" they're wrong on the Ati side of things again?

Besides the fact that they seem to have broken the NDA as well (unless they didn't sign one, but then that begs the question where they got that card from), the review contains one more contradiction: In the article itself, they write that the card has a 256 bit interface like the other Cypress cards. In spec table at the end of the article where they list the specs of all Cypress and Juniper parts, however, it says 128 bit. Might be a simple copy-paste error, though.

On the other hand, it could simply be that they got some stuff wrong (wouldn't be the first time, as CarstenS already pointed out).

 

[CarstenS]

Besides the fact that they seem to have broken the NDA as well (unless they didn't sign one, but then that begs the question where they got that card from), the review contains one more contradiction: In the article itself, they write that the card has a 256 bit interface like the other Cypress cards. In spec table at the end of the article where they list the specs of all Cypress and Juniper parts, however, it says 128 bit. Might be a simple copy-paste error, though.

On the other hand, it could simply be that they got some stuff wrong (wouldn't be the first time, as CarstenS already pointed out).

If they'd have got a real card and not simulated one themselves, they'd not have made those specific mistakes.

 

[GZ007]

There isnt any die shot to see it unfortunately.

There is actualy a half visible RV770 behind a penryn die where u can see the 10*4 ALU blocks in the midle of the chip with some imagination I think they could fit 3200 sp-s
with the planed 40nm 480mm2 die and wider memory controller if they would go the nvidia way. That would be quite shocking 800->3200

 

[Kaotik]

There is actualy a half visible RV770 behind a penryn die where u can see the 10*4 ALU blocks in the midle of the chip with some imagination I think they could fit 3200 sp-s
with the planed 40nm 480mm2 die and wider memory controller if they would go the nvidia way. That would be quite shocking 800->3200

There's plenty of dieshots of RV770, I have one @ 3091x2992 (with 1 chip in middle and parts of other RV770's on the sides)
It tells us absolutely nothing about Cypress, though