R700 has too many shaders and not enough ROPs

[lemonadesoda]

OK, so here is a quick analysis of the data at http://www.hexus.net/content/item.php?item=16300



And now for a quick analysis: look at the Excel table below. The numbers in black are data from Hexus. The numbers in red are my calculations.


Clock standardised FPS takes the FPS and scales it as though all clocks were at 4850 speeds, ie 625Mhz. It shows (approx) what the FPS would be if we OC the 4830 and underclock the 4870.

FPS per shader takes this standardised FPS and divides by the number of shaders on the GPU.

LOOK! The 4830 is the most efficient card (100%), with the crossfire (2x) 4830 (92%) coming next and the 4850/4870 somewhere behind (78%/81%). In COD4, the 4830 is about 25% more effective per clock per shader than the 4850/4870! Why is the 4870 better than 4850 if they have the same shaders and we standardised the clocks? Because of the faster memory.

BUT even with faster memory, is still underperforms the 4830 clock for clock, shader for shader. The 4830 is 25% faster.

So what gives? Easy, there are only 640 shaders on the 4830 but 40 texture units and 16 ROPs just like 4850/70.

THEREFORE, the 4830 is a more optimal combination of shaders, texture units and ROPs. Since texture units are 40 compared to 16 on the previous gen cards 3xxx, but the ROPs remain at 16, then I conclude that the ROPs is the primary bottleneck on the card, followed by the texture units. Why because adding more shaders on the 4850/70 doesnt increase the relative performance.

We need more ROPs and more texture units in R800! Not more shaders! (Relatively speaking).

[[I.R.A]_FBi]

good find

[phanbuey]

very nice

[Duffman]

there is a 4830x2? or is that two of them in Xfire?

[spearman914]

Originally Posted by Duffman

there is a 4830x2? or is that two of them in Xfire?

It isn't released so i think it's either future info or just 2 x xfire

[Duffman]

i think my eyes were playing tricks on me. I swore it read 4830x2 not x24830

nice info !

[lemonadesoda]

I guess another comment is that crossfire really has come on a long way. Remember when it was launched? Somewhere between 0% and 40% performance improvement in the best cases. Now we are often seeing 80% FPS improvement on a crossfire setup. ATI have done a good job getting the hardware and drivers to work.

[swaaye]

ATI has been running lopsided like this since R580/X1900 (well, X1600 actually). They always bank on software needing more shader ALU power than texturing/fill-rate. RV770 didn't change that even though it may look like it did on the surface.

Believe it or not but RV770 outperforms even GTX 280 by a long shot on most game shader code. Games just aren't that heavy with it for it to matter and other bottlenecks take over.

More info:
http://forum.beyond3d.com/showthread.php?t=49327

[AddSub]

Excellent post lemonadesoda. I've been saying this for years (link below) but it is hard to be heard through all the fanboyism on a forum like this. Raw pixel fillrate and throughput is still very important and it will stay as such until all the functions of a GPU are truly streamlined into a some sort of a unified array architecture. Unified shader architecture was one of the steps in that direction, however, plenty of operative functions on-GPU are still done on the side. Many of them through ROP's, as well totally separate subsections dedicated to certain types of texturing, and similar. All outside of shader realm.

As for why ATI/AMD GPU's are ROP/pixled fillrate starved and have been stuck on 16 ROP's since 2004? I answered that question in the post/link below, months ago. (In short, cost.)

Heck, I answered it many times over, here and on other forums but fanboysim noise is hard to evade. As soon as you start putting out names of corporations (AMD/ATI/nVdia, etc) you attract a certain type of "enthusiast" into the conversation and all intelligent exchange is pretty much finished.

http://forums.techpowerup.com/showpo...&postcount=169

[MKmods]

we are gonna be hearing great things about the 4830..Nice post lemonadesoda

[LAN_deRf_HA]

So what is the actual performance difference between a 4830 and a 4850 when you clock them both at 625/1986?

Edit*
Just checked the reviews, at the same speeds it doesn't really close the gap much, you need 700 MHz on the core for it to match the 4850, and a number of 4830s can't clock that high... so I'd rather pay $20 more for a 4850.

[a111087]

this was the same for HD2xxx series as well

[lemonadesoda]

I wonder what the results would look like running the 560 shader 4830 (http://forums.techpowerup.com/showth...highlight=4830).

If we similar game performance as the 640 shader 4830 (not just the synthetic shader only benchmark that W1z did) then we know that the RV770 is very seriously ROP handicapped.

I think R8xx needs to be up there at 32 ROPs. Any less and there is going to be that nasty performance wall at 1920x1200 and especially at 2560x1600. (Oh, and that NEW samsung resolution, 2048x1156 http://forums.techpowerup.com/showth...t=wide+monitor)

I was always surprised at how hard that wall was. It didnt make sense. Now it does. It's not a shader limit, is the ROPs and texture units. They just can't cope with ultra-resolutions. (And why crossfire is so effective at these resolutions... it's NOT the extra shaders, it's the extra ROPs and texture units).

[christof139]

Great idea and charts showing the equalized ratio/percent!!!!

I hope some manufacturer decide to put 512MB GDDR4 on a 4830 (if they have any GDDR4 left in stock, and if so simply to use it up) to make a limited edition model. Even though the GDDR4 won't make a big difference it will give a small gain and it the limited editon model is slightly OC'd from the factory that would just add a bit more performance, maybe 5% overall performance gain for $5 -$15 additional cost.

Anyone seen ATI's roadmap for the next gen. cards?? 12 (model 5000), 16(5200), 20(5400), 24(5600), 30(5800) and 32(5850) ROP's would be the way to go as you all say for a low to high end line of models. 36 (5900) could be possible. That would just be great as long as the price stayed down, and with the world economy as it is the price may not stay down, although it just may not.

A Special Edition GDDR4 HD 4830 or 48xx would also be great now, maybe even with 4 more ROPS added, and/or a 4830x2. Heck, a 4670x2 would be great. A 4670 with 12 ROPs would also be great. There was a 1GB 2600XT x2 card made by Saphire and Visiontech and ATI.

Why doesn't/didn't ATI/AMD add even a mere 32 or 64 or or 96 or 128 bits to the memory interface?? With the present 4000 series cards that would make a decent difference I bet. I guess it is in the die size production problem area; how to size and cut up the wafers most efficiently.

Same for memory. 576MB or or 608MB or 640MB would be interesting. Just another cost-efficiency per wafer cutting deal I guess.

NVIDI and partner manufacturers did and does this so maybe ATI/AMD and/or one or more of their partners may try something like this.

More diversity might be OK. Limited Editions of several models in a number series might also be OK and economically feasible. Beats me.

Chris

[btarunr]

$250 for two HD 4830 cards is better than $400 for one GTX 280. The argument is just like that of two GTX 260 being better than one HD 4870 X2.

http://www.hexus.net/content/item.php?item=16300&page=1

640 SPs is the right combination for the ROPs. 8 ALU clusters.

[ShogoXT]

Im not so sure.

[EastCoasthandle]

^^Agreed, 2 4830 (with MIR) are $200. A GTX280 would cost $359.99 after MIR. Although some may prefer another brand at $419.99 after MIR. A difference of $160 or $220 if a person prefers another brand.

A 4870 X2 costs $479.99 (after MIR). 2 GTX 260 216s cost $229.99 each totaling $459.98 after MIR. Another brand can cost $10 more a piece after MIR. A difference of $20 or costing similar to the 4870 X2 depending on what people prefer.

So by using a popular online retailer in the states the difference is quite substantial. Therefore, IMO I don't think that buying 2 4830s are the same as buying 2 260 216s when compared to it's competitor high end part.

Keep in mind that prices may change.

[ShogoXT]

I mean I understand the logic in the arguement. The thing is though alot of tech website forums (not so much this one) were speculating on the specs of the 4800 series and saying why it would fail against nvidia. They said because of the ROPs it would suck and never stand a chance to the Nvidia 9k series. And would be stomped by the GTX 200s in value right when they came out....

Oh wait a minute.

[EastCoasthandle]

Originally Posted by ShogoXT

I mean I understand the logic in the arguement. The thing is though alot of tech website forums (not so much this one) were speculating on the specs of the 4800 series and saying why it would fail against nvidia. They said because of the ROPs it would suck and never stand a chance to the Nvidia 9k series. And would be stomped by the GTX 200s in value right when they came out....

Oh wait a minute.

Yup, this is what happened back then. Turned out to be a fabrication.

[douglatins]

R700 is the HD4870X2 only!

[SiliconSlick]

Originally Posted by AddSub

Excellent post lemonadesoda. I've been saying this for years (link below) but it is hard to be heard through all the fanboyism on a forum like this. Raw pixel fillrate and throughput is still very important and it will stay as such until all the functions of a GPU are truly streamlined into a some sort of a unified array architecture. Unified shader architecture was one of the steps in that direction, however, plenty of operative functions on-GPU are still done on the side. Many of them through ROP's, as well totally separate subsections dedicated to certain types of texturing, and similar. All outside of shader realm.

As for why ATI/AMD GPU's are ROP/pixled fillrate starved and have been stuck on 16 ROP's since 2004? I answered that question in the post/link below, months ago. (In short, cost.)

Heck, I answered it many times over, here and on other forums but fanboysim noise is hard to evade. As soon as you start putting out names of corporations (AMD/ATI/nVdia, etc) you attract a certain type of "enthusiast" into the conversation and all intelligent exchange is pretty much finished.

http://forums.techpowerup.com/showpo...&postcount=169

Thanks for telling the truth - and good luck not getting your head chopped off.
(things seem rather nice here actually compared to most spots{all spots I've ever seen} I hope it stays that way)

[W1zzard]

Originally Posted by AddSub

As for why ATI/AMD GPU's are ROP/pixled fillrate starved and have been stuck on 16 ROP's since 2004?

16 rops, 2 pixels per clock = 32 rops effective

[christof139]

A few days ago I read an article that stated ATI shaders are less complex than NVIDIA ones, therefore ATI simply adds more shaders to make up for this, and it worked fine in the 4000 series.

Chris

[tigger]

If anyone has a 4830 we could do our own test with it against a 4850 downclocked to the same clocks as the 30.

Is that the same for the 4830 w1zzard,32 effective?