AMD: R8xx Speculation

+
r100->rv200 (r7200->r7500)
r350->r360 (r9800pro->xt)
r580->r580+ (x1900xtx->x1950xtx)

 

How difficult would it be for them to increase triangle setup rate if it is indeed one of the performance bottlenecks of the system. After all the R600 -> Rv670 "refresh" saw some pretty big changes also.

Regards,
SB

 

I don't think any of those really count though, as there's more than just a simple clock bump involved in all of those transitions, all of which have been re-hashed on these forums already.

 

R420 was AGP part and R480 PCIe, so even this change wasn't targeted on clock-speed only (I think the main reason was some manufacturing problem with R420, which negatively affected yields, not 5% higher clock-speed)

 

Call it R423 then, R420 was the base part from which R423 was derived, and the only change was PCI-e support.

 

Marketing, its easier to make it with a 1ghz card than a 600mhz one.
I rather own a 1ghz one at simliar speed than a 600mhz even if I know its not logical

the feeling of a 1ghz card sitting there, rocks.
the refresh if going 1ghz seems to be a good way to spin the cards

 

Clockspeed does sell, but then NV can always counter by advertising their shader domain clock and say "look 1.xGHz vs. a paltry 1GHz!"

 

Redwood's going to clock nicely.

 

will it matter though ? for a sub $99 part limited to 128bit and half juniper I would think performance would be very very limited to begin with that even very high clock rates (900-1Ghz) wouldn't move it up from value to mainstream performance segment (though I could be pleasantly surprised)

 

IF GT240 got introduced and judged as not having enough performance for it's price, would you just need more performance at that pricepoint then? or a lower price?

 

Well at the rumored clocks of 775MHz for the GPU and 1000MHz for the memory, it just might compete with the 8800 GT/9800 GT/GTS 240/whatever it's going to be renamed to now.

 

With 20 TMUs and 8 ROPs, it should be more in GT 240 range.

 

Judging by the rumored specs (400SP, 20TU, 8 ROPs), it should have a little less that 30% more shading & texturing power than a 4670, the GDDR5 version will have a lot more bandwidth and pixel fill rate is almost the same as RV730. I would expect almost 20% more performance in average than the 4670, that should be in the 9600GT league.

 

It will have 40% or so less texturing power than 4670.
Could indeed be competitive with GT240/9600GT, though hopefully it will be cheaper?

 

ALU performance +29%
texturing perf. -35%
Z / blending +3%
triangle setup +3% (or more?)

Thats zero at the average. But RV830 has 2 significant advantages - double bandwidth and it isn't limited by interpolators. RV730 had 32 TMUs, but only 16 interpolators. RV830 has only 24 TMUs, but they can be utilized much better (more interpolation power, more bandwidth, higher ALU:TEX ratio).

GF9600GT league - yes (~RV730 x 1.3)... I think Redwood is pretty interesting GPU - only 20% transistors over RV730, but performance boost is around 30% and it supports DX11 in addition.

In some teoretical tests performance will be double when compared to RV730...

 

it's more about the clocks I guess.

 

Lower price, definitely. The 64bit-addled x3xx series seems to have become almost completely obsolete with the last round of IGPs, so targeting a lower price point would make good sense (even if it meant going with just GDDR3).

 

I think on average ALUs have much more impact on performance than TMUs nowadays. Also, the 4670 is quite bandwidth-limited in several games, so twice the bandwidth should already make quite a difference on its own, when looking at this example: http://www.guru3d.com/article/ati-radeon-hd-4670-review/13

Besides, I expect Nvidia will fade out production of the 9600 GT soon and replace it completely with the slightly-slower-on-average GT240-GDDR5. The GT240 has its 32 TMUs and 8 ROPs clocked at only 550 MHz, so the 5670 comes quite close in theoretical texturing power, clearly wins in pixel fillrate, and has more raw ALU power and memory bandwidth. I expect the 5670 to be faster than GT240 GDDR5 and at least on par with the 9600 GT, and possibly very close to the 9800 GT in some games.

This assumption applies to average numbers though, actually I expect it to compete with 9800 GT when ALU-power is required and/or 4xAA but low AF or no AF at all are applied, and to lose against 9800 GT and sometimes struggle even against 9600 GT when only 16xAF but no AA is enabled and shader-load is relatively low.
4xAA/8xAF is probably the "sweet spot" for the 5670 (though it might already be too slow for such settings at resolutions above 1280x1024 with highest details).

 

Yes, sorry. I remembered that RV730 had only 16 TU but I was wrong. Hmm.. I remembered that RV730 had the same architecture of RV770, only with fewer SIMD, but now I gave another look at the RV730 diagram and I see that the SIMD are also half-lenght with respect to RV770 (half the ALU:TEX ratio). So 8 SIMD with 40 ALU each instead of 4 SIMD with 80 ALU. Interesting (and I have the proof that my memory is getting worse).

 

TKK: According to my experience HD4670 wasn't primarily limited by BW - 10% OCed memory modules = 2-3% performance (at the average). 32 texturing units were quite overkill (HD4770 has 32 TMUs, too, but scales with bandwidth much better). The TMUs were never fully utilized - low interpolators/TMUs ratio, low ALU:TEX ratio, low BW... I think HD4670/RV730 was the most unbalanced of the entire RV7xx serie.