AMD: R7xx Speculation

[XMAN26]

http://www.computerbase.de/artikel/hardware/grafikkarten/2008/kurztest_ati_radeon_hd_4850_rv770/



RV770 = new R300

First, Nice job cherry picking on the overall performance graphs, wonder how the figured their numbers.

Seconded, having gotten that out of the way, I want to see how this card does in 4-6m after drivers optimizations and fine tuning as it seems to fall flat in alot of cases in that review. It has very good performance and excellent price, but drivers really need fine tuning.

UT3 show really wierd AA numbers, 4x to 8x shows a 0 to 5% drop on ATI cards but huge drops on Nv stuff. Take that one bench out and the performance for 280 with 8xaa is more like 140-150 vs 4850 100.

 

[Rangers]

Yep, overall 9800GTX holds its own. Now it comes down to price.

4870 is still going to be slotted in a space I dont see Nvidia being able to reach, though. 9800 cant touch it from the below, and 260 cant from above (without Nvidia losing money).

 

[pjbliverpool]

So, RV770's 1200GFLOPs take 104mm2, while GT200's 1000GFLOPs take 153mm2:

• RV770 is 11.5 GFLOPs/mm2 on 55nm
• GT200 is 6.6 GFLOPs/mm2 on 65nm

For double-precision:

• RV770 is 2.3 GFLOPs/mm2 on 55nm
• GT200 is 0.5 GFLOPs/mm2 on 65nm

Jawed

Yeah but if its slower regardless then does it really matter?

GT200 has something like 40% more transistors and 50% more die size on the current process. Will it be 40% faster than the 4870 when it comes to the benchmarks (including GPGPU)?

Probably not but it will be close and thus its comparative efficiency isn't all that poor when it comes to what matters - actual performance. When they move down to 55nm aswell that should brin gthe die size down a lot while simultaneously taking performance up. On a performace/die size comparison at that point GT200 may well win.

 

[Arty]

First indication of RV770 being bandwidth starved in 4850?

http://www.fudzilla.com/index.php?option=com_content&task=view&id=8003&Itemid=1

1920x1200, No AA
4870 - 30.5
4850 - 25
9800GTX - 18.6

1920x1200, 8xAA
4870 - 27.9 (-8.5%)
4850 - 21.7 (-13.2%)
9800GTX - 11.9 (-36.2%)

 

[DegustatoR]

Better die-shot:

http://whatifgaming.com/ati-radeon-hd-4850-review-show-me-the-ruby

Unbelievable.
180 SPs (20 for redundancy?)? 10 clusters with a local memory per cluster? 20 TMU blocks (2 fetches+filtering per block?) Where's the control logic? Where are all the other caches outside of SP clusters?.. Where are ROPs???

P.S. NV-boys are on defensive already? Don't be so fast with writing GT200 down as a mistake =) We'll see how it'll play out futher down the road with new GF drivers and GT2xx lineup.

 

[ChronoReverse]

Well regardless of whether Nvidia strikes back with price cuts or not, it's really nice that ATI has a competitive part that forces such cuts.

 

[Jawed]

Unbelievable.
180 SPs (20 for redundancy?)? 10 clusters with a local memory per cluster? 20 TMU blocks (2 fetches+filtering per block?)

10 SIMDs x 5 ALUs x 1 redundant ALU lane = 50 redundant ALU lanes I reckon. 6%.

10 TUs and there are no longer any TAs and samplers dedicated to vertex data fetch.

Where's the control logic? Where are all the other caches outside of SP clusters?.. Where are ROPs???

~59% of the die is taken up with all that other stuff, including various I/O ports.

Jawed

 

[syl]

Sapphires Vantage scores for the 8800GT and 4850 seem to be lower and higher respectively compared to http://www.pcper.com/article.php?aid=579&type=expert&pid=6 though.

The 8800GT used in the Pcper article was a BFG OCX @ 700/2000MHz.

 

[ChrisRay]

I wonder how do recent customers of 9800GTX cards are feeling now? or even better how does current stock owners of 9800GTX cards are feeling nov?

Well G92 is 330mm² big and I can't see what magic can bring it on 260mm² die size of RV770!?

QFT

Anyone who buys computer hardware IMO knows that price/value drops overtime relatively quickly.

 

[Sunday]

Anyone who buys computer hardware IMO knows that price/value drops overtime relatively quickly.

you would agree that this is the record... hell how long has 9800GTX been present on the market?!

 

[Arty]

Anyone who buys computer hardware IMO knows that price/value drops overtime relatively quickly.

Depends on the timing really. I bet the early R300/G80 adopters got the most out of their money ..

 

[AnarchX]

10 TUs and there are no longer any TAs and samplers dedicated to vertex data fetch.

RV770 is 800SPs and 32 TMUs:
http://www.pcper.com/article.php?aid=579&type=expert&pid=5
~ 19500MTex

Probadly we see here 5x2 half SIMDs? Somewhere are the 8 or 2x4 TUs.

edit:

Up to 160 texture fetches per clock cycle

http://amdzone.com/index.php/reviews/71/9836-ati-radeon-hd-4850

RV670 was 80 and RV63x was 40.

 

[ChrisRay]

you would agree that this is the record... hell how long has 9800GTX been present on the market?!

Not really. Cards move in price. Card has been around 2 and half months. April 1st was its launch date.

Chris

 

[Jawed]

RV770 is 800SPs and 32 TMUs:
http://www.pcper.com/article.php?aid=579&type=expert&pid=5
~ 19500MTex

The leaked diagram shows 40 TUs though. I originally came to the conclusion that RV770 is 32 TUs based on 3DMk06 benchmark, but I prefer to trust the diagram.

Jawed

 

[Mintmaster]

Only with 8x AA.

Indeed. NVidia has 16x CSAA with little perf. hit over 4xAA, so this graph is pretty meaningless.

With 4x or less the scores are much closer. Pretty even with the 9800GTX. Thats certainly a great turn around for AMD given that this is the slower of the two parts but it doesn't make it the new R300. For that, the 48xx series would have to be faster than the competition

You answered your own challenge there. The 4870 will be faster than the competition. I don't know if NVidia can afford to sell a 260 at the same price as the 4870.

I think it's more apt to call the RV770 the new NV43. That's the card that destroyed ATI's R300 momentum and set the stage for NVidia's perf/mm2 dominance in the next gen, so it's a pretty good sign for ATI.

 

[MfA]

10 clusters with a local memory per cluster?

Cache.

 

[simbus82]

Anyone can tell me how it is possible????

i think quarl texture pack is really full of heavy shaders...or not?

 

[Lux_]

Anyone can tell me how it is possible?.

Most definitely bug in NV driver: notice how 9800GTX is 2 times slower than 3870.

 

[mczak]

Up to 160 texture fetches per clock cycle

RV670 was 80 and RV63x was 40.

Yes, I guess I better believe the 40 tu now...
(160 fetches is of course 40 tu with 4 fetches each for bilinear filter - 80 was 4x16 for bilinear + 16 for point sampling etc.)

 

[HawkeyeGnu]

So a future part is announced (9800GTX+) on the day HD4850 debuts.

It seems paper launches are back in vogue and perfectly acceptable to some hardware sites again?