AMD: R7xx Speculation

Whereas I concur WRT to HD2900, I am quite certain this time will be better.

 

Additionally, I don't think anyone right now is actually thinking Rv770 is going to beat GT200 head to head as was the case when leading up to the launch of R600.

I think pretty much anyone that's been paying attention expects GT200 to be quite a bit faster than Rv770, however with a much larger power consumption and larger cost of production.

The only thing people seem to be all over the place is how competitive R700 (2xRv770) will be with GT200 both in performance and power.

Considering how little is known for sure about Rv770, there's even less known about R700.

Regards,
SB

 

No, thank YOU. And i dont need any help to promote such a great product as the HD4800 series. Like always, ATI is the best and if i could i would marry u




This is important, casue even the AIBs are only people and we wouldnt like to see a HD 4870 XTX-PE, would we?




So much about the NDA ending, its never gonna end

 

At this rate the HD4850X2 will arrive before the HD4870.

I presume that the tales of HD4850 sitting on shelves are true, and that it's HD4870 (GDDR5?) causing delays.

Jawed

 

... skip a heartbeat ...

 

It's OBR, not ORB by the way. On a Czech forum he added that AMD/ATi never lent him anything for review and that he always gets the hardware by himself.

 

I doubt he will release results.

 

http://my.ocworkbench.com/2008/amd/AMD_Radeon_HD4850_HD4870_Info/AMD_HD4850_HD4870-1.htm

So it will be...

 

I think we can just about all accept the 800 SPs, ATi's always had relatively cheap SPs/ALUs compared to NV. Now we need to know TMU count, and fillrate numbers (Zfill, AA samples/cycle).

 

Slightly lower bandwidth on the HD4870 than previously reported, 115.2GB/s.

Jawed

 

With G200 probably devoting around 10% of its area to ALUs I still find amazing that AMD can pack 800 ALUs in such a small chip (again, compared to G200)

 

Which would place the GDDR5 clock at 1.8GHz (3.6GHz "effective DDR"). I thought this was known all along? ATi always goes for the lowest possible speed-bin GDDR to reach the desired bandwidth "range".

 

Forget that. I'm still in shock

Efficiency concerns aside it's downright insane that AMD built a chip less than half the size of Nvidia's best and totally creamed them in peak processing power. And apparently it gets only worse with DP!

One nice surprise would be if AMD used all that power to get rid of fixed function texturing completely. One thing's for certain - games won't be making use of all those flops.

 

I think he will. After he vented in on at least two forums, plus it got here, I pretty much think he will release it. By the way, a proof he really has the cards:
http://i27.tinypic.com/2m6n5gz.jpg (the coin is a 1 Czech Crown shot from the backside)

 

It's precisely because of ATi's relatively inefficient architecture that they must cram so many SPs into their chips to extract even a reasonable amount of performance from them (real-world apps, compared to competition).

FF hardware still has its place, but the trend toward homogenous cores in GP/GPU is clear.

 

How many gflops then ? Just 1 teraflop for HD 4850 ? if that´s the case, wow!. These things will be polys crushing beasts!.

 

We are not in shock cause they must, we are because they can!

 

625Mhz * 800Units * 2flop/unit = 1 teraflop. So the 4850 delivers exactly 1 teraflop peak performance. The question is how it translates into real world performancer.

 

I've gotten used to it. I really hope there'll be a decent die shot this time around...

With 1300MHz shaders on GTX 280 presumably a new process (40nm?) will send that soaring. When G92b appears perhaps we'll get a clue about how well power/clocks work out for NVidia going from 65nm->55nm, though I suspect the sheer size of GT200 will make that a limited increment.

I think that's a few years off - there'd be almost no effective ALU capability left even with 800 lanes.

I guess it's time to start bandying around some ludicrous ALU counts for 40nm... 2000?

Jawed

 

I'm curious to see what kind of crazy AA levels the card can pull off. AA probably appears somewhat free because they use it to pad the scheduling. Heck, if a significant number of ALU ops were non-dependent AA instructions they could be padding that efficiency quite nicely.