Horrible ati yeilds (3%!!)

[turtle]

Word, so did I. Here we were expecting 90nm to bring super high clocks from ATi, but 32pp and a high clock (the best of both worlds) may have been asking too much.

In theory, if it was MADE to be 16pp, the room to clock to 700mhz would be there, but with a chip made to have 32pp, couldn't that hamper it's ability to have a high clock? This makes sense to me as when nvidia moved to 110nm you'd think the chip would clock a good deal better than the 130nm gf6, but it doesn't, just a little. It makes me think this is hampered by the addition of pipes. I'm not an engineer though, so don't take that as fact.

Also, obviously ATi is having yield problems. Some chips probably do clock really well, but they have to set a standard somewhere so they can get an acceptable amount of chips out to sell. Maybe it's 3% that can do 700mhz with 24pipes? Maybe 700mhz @ 16pp was possible with 16 quads not working at an acceptable, but they figured 500 @ 24pp was a optimal design that also had an acceptable yield? That would make sense then they'd release it at 500mhz if a greater yield could operate at that clock with 24pipes, and still beat G70. Who knows what they did and why they did it, maybe those 3% will become the xtpesupermegadeluxeultra's to beat the G70 ultra. That would also give you a clue how high the R520's will clock (If 3% hit target speeds...say 600/700/800...we'll obviously hit a wall in our overclocking much below that under normal cooling...but who knows what that target was...

It still looks to me though as the refresh will look more exciting.

ATi R580 - 32pp, clock prolly similar to R520 but really anyone's guess, more TMUs? (500x32?)
Nvidia G71 - 90nm G70...so 24pp, prolly a super high clock. (24x650?)
^^anyone's guess on clocks

That'll be a fun one...

[Cybercat]

Yeah, more pipes would definitely limit the clockspeed. Think of it this way, more pipes means more transistors. More transistors means more heat, and power consumption, plus the signal integrity suffers (communication via wires between transistors). I mean if you've ever overclocked before you know that heat can be a real obstacle when you're trying to get up there in clockspeed.

What I find funny is that people who buy the RV530 are gonna have a chip with 2/3 of the pipelines that the R520 has, less transistors, while being built on the same process. Now see, last gen the X700 series had only half as many pipes as the X8x0, but they were built on a less mature process, 110nm. As the X800XL proved, this 110nm wasn't as efficient in clocking as the refined 130nm low-k process. This time around mainstream users are going to get all the clocking benefits of the highend card, without the added heat. I mean the R520 doesn't just have 16 or 24 pipes creating heat, it has all 32 contributing thermal waste, whether they're all enabled or not. People who buy the RV530 are going to be able to clock MUCH higher than the R520 can achieve, and get very similar (though not quite as much, as you observed turtle) performance, for like half the price.

NVIDIA will probably make the G72 (7600 series) on a 90nm process. I'm eager to see if this line has 16 pipes or not. Most likely 12, I'm guessing.

BTW, what are TMUs?

[alexio]

What he said ^^^
I was doing the calculation on fillrates though, and it looks like 24/500 would be better than 16/700 by a good chunk (12,000 as opposed to 11,200). So let's hope that's true.

Well I've done some testing on my Asus 6800NU and I have to say that the card running 8 pipes @ 400 mhz is faster then 16 pipes @ 200mhz. 3dmark gives higher scores @ 8 pipes. 16 pipes is like a dual-core 8 pipeline card, it can handle more threads at the same time but it completes a thread in a slower time.

What I try to say is that the fastest card would just have 1 pipeline at an xtreme speed. That's why you can't say that 16 pipes at 700mhz is slower than 24 pipes at 500mhz. The fillrate is lower, that's for sure. But just look at the R420 vs NV40. The R420 has a much higher fillrate but isn't faster in particular.

[turtle]

Yeah, more pipes would definitely limit the clockspeed. Think of it this way, more pipes means more transistors. More transistors means more heat, and power consumption, plus the signal integrity suffers (communication via wires between transistors). I mean if you've ever overclocked before you know that heat can be a real obstacle when you're trying to get up there in clockspeed.

What I find funny is that people who buy the RV530 are gonna have a chip with 2/3 of the pipelines that the R520 has, less transistors, while being built on the same process. Now see, last gen the X700 series had only half as many pipes as the X8x0, but they were built on a less mature process, 110nm. As the X800XL proved, this 110nm wasn't as efficient in clocking as the refined 130nm low-k process. This time around mainstream users are going to get all the clocking benefits of the highend card, without the added heat. I mean the R520 doesn't just have 16 or 24 pipes creating heat, it has all 32 contributing thermal waste, whether they're all enabled or not. People who buy the RV530 are going to be able to clock MUCH higher than the R520 can achieve, and get very similar (though not quite as much, as you observed turtle) performance, for like half the price.

NVIDIA will probably make the G72 (7600 series) on a 90nm process. I'm eager to see if this line has 16 pipes or not. Most likely 12, I'm guessing.

BTW, what are TMUs?

1. Your thing on transistors, heat, ect...Exactly what I was thinking. Makes perfect sense.

2. 2/3 the pipes, 16 broken pipes, 16 working pipes, same amount of transistors (RV530 sounds like it'll be a R520 with 4 broken quads, 2 more broken quads than R520, which already has 2 because it was meant to be 32pp...that's what i've gathered anyway...IF THIS WHOLE THING IS TRUE) who knows how well they'll clock. True, they may clock much better because only 4 quads have to hit that number instead of 6...who knows? Yes, it being on 90nm, just like 520 helps it, but there still will be all the "waste", at least the way I understand it at the moment...there will actually be more, % wise. My guess is RV530 will be a great mainstream part and compete against 7800gt well, and you'll pay more than the last gen high-end mainstream card. It'll prolly be $100 less than R520 or something.

3. No idea about 7600gt. Hope it kicks ass. Everyone that only had $150-$200 to spend on a gfx card last gen really got a good deal on the 6600gt, hope they keep up that tradition.

4. Texture Mapping Units.

[Cybercat]

Eh, I really have my doubts about the rumors for the RV530. Maybe it's just R530, as the "RV" series is reserved for ~$200 parts usually. A codename like "RV510" would be more in line with what's been done in the past. Then again, with so few details on the big flagship itself, you can't fault the details on the mainstream parts for being a bit shakey at best.

[turtle]

Alexio, you're right. In SOME instances one will be faster, and others it'll be reversed, won't it? I would think in some instances a greater fillrate would be to much an advantage, as some benchies/games R420 beats up on the NV40 rather well...like HL2 and Far Cry for example. I imagine those use the higher fill rate? Granted there are times it doesn't help it, and nv40 beats up on R420.

Admittedly, I don't know a lot about greater mhz vs more pipes. I just know the fillrate would be higher, and it seemed logical as a possible reason why they'd do it. Couldn't ATi's pipe architecture be more efficient than nvidia's when doing this comparison, especially in this upcoming generation? That would be one reason to do with 24pp@500 rather than 16@700 right?

Also 200x16 = 3200, 400x8 = 3200. They're equal. The later is more efficient, but would it be enough to make up the differance in fillrate assuming the pipe architectures are similar (11,200 as opposed to 12,000). Feel free to enlighten me...I'm intrigued.

[turtle]

Well, I assume RV530 is the pro model...and it's supposedly confirmed that it's 16pp...who knows if it's a different chip or just just a failed R520. Considering the yields though, it'd make sense they'd use failed R520's to make the lower-end chips, as otherwise (according to the yield rumours) they'd be wasting a WHOOOLE lot of money on those wafers.

And remember, 7800GT's are prolly just gtx's with a failed quad and they're using them to use the failed yields of the gtx, and they will probably overclock pretty well...I mean, look at how close the core clockspeed is to the GTX...and even if the mem is rated at 1100, it probably uses the same stuff as GTX, so you can expect a similar clock there as well. I truely only think the 20p vs 24 pipes is going to make a 2-3% differance anyway...(based on a 5% differance of a GTX vs 6800GT/U at same clocks..16pp vs 24pp) so assuming the GT can hit similar clocks to gtx, it'll be a good deal just like the 6800gt was. Who's to say RV530 couldn't do the same, it'd especially be true under Alexio's reasoning. I do think that ATi's pipes are better than nvidia's though, and it might make a bigger differance than 5%. I imagine RV530 will clock overall just slightly better than a x850xt or something, with the same amount of pipes, but probably better pipes then that previous gen. I expect RV530 and 7800gt to fight tooth and nail, but who knows...There will obviously have to be a winner. We'll know when we see some benchies for it...We've already seen 7800gt benchies, and they're smack between 68U and 78GTX, I expect RV530 to be the same with x850xt and R520...so again, it'll be close, might come down to price and availability.

But yes, things are shakey. We should know in the next month or so though...

[alexio]

Alexio, you're right. In SOME instances one will be faster, and others it'll be reversed, won't it? I would think in some instances a greater fillrate would be to much an advantage, as some benchies/games R420 beats up on the NV40 rather well...like HL2 and Far Cry for example. I imagine those use the higher fill rate? Granted there are times it doesn't help it, and nv40 beats up on R420.

Admittedly, I don't know a lot about greater mhz vs more pipes. I just know the fillrate would be higher, and it seemed logical as a possible reason why they'd do it. Couldn't ATi's pipe architecture be more efficient than nvidia's when doing this comparison, especially in this upcoming generation? That would be one reason to do with 24pp@500 rather than 16@700 right?

Also 200x16 = 3200, 400x8 = 3200. They're equal. The later is more efficient, but would it be enough to make up the differance in fillrate assuming the pipe architectures are similar (11,200 as opposed to 12,000). Feel free to enlighten me...I'm intrigued.

It's all about optimizing really. You can optimize a game to be the fastest at 1 pipeline but also for 24 pipelines. Todays games are made to also run on slower cards with less pipelines. This is the part where the unified shader model becomes interesting for high-end cards.

On a normal fixed 16 pipes 6 vertexes card there are one or more pipes and/or vertexes doing nothing, waithing for a new task. With the unified shadermodel every task can be handled by all the pipes and vertexes; that way there will never be a scenario where for example the vertexshaders are working there asses of while the pixel pipelines are hardly doing anything, or the opposite ofcourse.

A combination of different game executables for different types of hardware (graphicscards, cpu's, amount of RAM) and the unified shadermodel will make games run much better. For example in a game all physics are being handled by the cpu, this is fixed. If you can make an executable that gives this task to the graphicscard when the cpu has to much tasks to handle while the gpu has to little than you can achieve much better performance in the game. The first step will ofcourse be making games multithreaded for use with multi core cpu's. In the beginning this might be fixed, than for example the A.I in the game is being handled by the second core (only). Further on in the development of games this process can be made dynamic, just like the unified shadermodel.

[Cybercat]

Alexio got me interested in doing some runs in 3DMark05 to see how the NV40 architecture scales with clockspeed. He said he scored higher at 400x8 than at 200x16. Well I've got a 6600 so I don't have the luxury of of 16 pipes but I can go lower with higher clockspeeds.

So at 200MHz with all 8 pipes enabled, this is what I got:


At 560MHz (my max core OC) with 4 pipes, this is what I got:


I was pretty surprised with the results. Having higher clockspeeds versus more pipes gives about a 19% overall improvement. Here's how they do in each game.

Return to Proxycon: 7%
Firefly Forest: 28%
Canyon Flight: 23%

It could have something to do with the architecture, or it could have something to do with software optimizations like Alexio said. Can't know for sure. But this is very interesting, and it could mean that lower end cards with 16 pipes at higher clockspeeds could perform the same or better than a 24-pipe card at lower clockspeeds (not taking into account other differences and dis/advantages).

[HARDCORECLOCKER]

Grrrrrrrr........., me wants a X900XT PE with 32 pipes 0.09...

[alexio]

Can you run at 200mhz 8 pipes and at 400mhz 4 pipes for a better comparation? The results a got when I tested 200mhz 16 pipes vs 400mhz 8 pipes was with a card having some problems so I'm not 100% sure that I got the right scores. I don't have 3dmark on my hdd right now so can you please test it for me?

[Cybercat]

Why would that yield different results? 200 is half of 400, just like 280 is to 560. The scores would be lower, but the % difference should be the same.

[alexio]

When testing with 4 pipes you should enable only 1 vertex shader and testing 8 pipes you should use 2. This is the best comparation possible. Your card normally has 3 vertexes but you can't divide 3 into whole numbers. You need to use double the vertexes at 8 pipes because the vertexes run at the same speed as the pipelines, in other words: you need double the amount at half the speed.

EDIT: I just wants to know if 3dmark can use the extra clockspeed better or the extra pipes. Like said above at half the speed you also need double the vertexes.

[Cybercat]

Well that's something entirely different. I'll be back with the results soon.

[7he]-[0rr0r]

As far as iv'e seen it's been rumored to be above 600mhz in fact the latest inq well rumor:

http://theinquirer.net/?article=24698
but then again its all a crap shoot till we see em

[WiCKeD]

Well I've done some testing on my Asus 6800NU and I have to say that the card running 8 pipes @ 400 mhz is faster then 16 pipes @ 200mhz. 3dmark gives higher scores @ 8 pipes. 16 pipes is like a dual-core 8 pipeline card, it can handle more threads at the same time but it completes a thread in a slower time.

What I try to say is that the fastest card would just have 1 pipeline at an xtreme speed. That's why you can't say that 16 pipes at 700mhz is slower than 24 pipes at 500mhz. The fillrate is lower, that's for sure. But just look at the R420 vs NV40. The R420 has a much higher fillrate but isn't faster in particular.

You're forgetting that if the clockspeed isn't up there, the extra bandwidth is not being utilized. That is the main advantage of having more pipes. 200MHz is probably not enough to push 16-pipes worth of data.

Thats bad and this is after they retaped it about 3 times

Where does it say that? They don't even list a source or when they got the info.

whew!!! And to think, I almost bought some ATI stock (atyt) this week... will wait til it hits 11 now.

C

Ya don't. They won't be turning a real profit until next year anyway.

[Cybercat]

200x8 (2 VS)

400x4 (1 VS)

As you can see, the dividiing the vertex shaders made a very large difference to how the performance scales. Overall difference is only 3%.

Return to Proxycon: 2%
Firefly Forest: 6%
Canyon Flight: 3%

[WeStSiDePLaYa]

looks like ati might be screwed now. no matter how fast their card is it doesnt mean if they cant get it to the consumer.

[perkam]

fanboyism is the key to ati success

Nods Enthusiastically lol

Well, the nVidiots have vanished because more often than not nowadays, Nvidia is the sensible choice...where as ATI is still the choice of the ATI Loyalist (aka fanboy) .. but at the end of the day, ppl will chose what gives the best performance.

ATI's problem right now isnt its yields, its the multiple dillemmas between 16/700 and 24/500 and what not. Once it makes a decision, it'll be ready to release the chips.

BTW Kudos to Turtle for an R520 thread that is more informative than these usually are.

Perkam

[crodan85]

You're forgetting that if the clockspeed isn't up there, the extra bandwidth is not being utilized. That is the main advantage of having more pipes. 200MHz is probably not enough to push 16-pipes worth of data.

Where does it say that? They don't even list a source or when they got the info.

Ya don't. They won't be turning a real profit until next year anyway.

http://theinquirer.net/?article=24540 Only going on what Inq said and as far as im concerned they know more than me and you.

[DilTech]

Crodan, that article was written by fuad, aka Fraud, we all know we can't trust anything he has to say...

[turtle]

First of all, thanks for the compliment Perk.

I figure there's some extremely intelligent minds floating around this forum with some free time, and since we're all eagarly awaiting this thing called the R520, so me might as well hash out what we know, think we know, or hear and try to make sense of it all, using that extra brain cycles of said people. I find it xtremely interesting...even if it's not what we think, at least I learned something, and perhaps helps me make more informed guesses and choices in the future.

Thanks for running those tests. As you can see, with one less quad (on the gf6 gen) it only makes a 3% differance at the same clocks. That's half the quads, 1/2 the vertex shaders, and 7800gt will have 4/5 the quads of gtx, with assuming the same number of rops and vertex shaders, but on a newer architecture that's pipes are prolly slightly more efficient than gf6, so again, 7800gt probably will be pretty close to 7800gtx if it clocks well, probably VERY close. The less pipes might make more of a differance then they did before, but it still probably won't be substantial, especially since the amount of pipes (in percentage) is more, and all vertex shaders supposedly will still be there from the gtx on the gt.

When making an assumption on R520, wouldn't it be best to run similar tests using an X800 card? Granted, supposedly R520's pipes are supposed to be more efficient than x800, but wouldn't it give a closer idea? Is there a way to do this, as obviously there's no nvstrap for ATi?

As for the Inq...Well, the last article I read is that it could be 16-32 pipes and 600-700+ mhz. Um...okay. They're right, it probably will be somewhere in there. Even if R520 is 500mhz, it may be able to overclock that well in some cases (3%... ) scaling that high, who knows? While anything could happen, it seems likely they couldn't get 32 pipes to work at an optimal clockspeed to feed the pipes in a large enough quantity, and 500 might be the spot where they can get out enough cards that will perform at that clock with 6 quads, while still being able to feed the pipes. Granted, maybe they can get 700/16 to work, just like maybe they could get 300/32 to work (following the same logic) but this was the most optimal. Who knows...Maybe the pro at 16 pipes in more cases than the xt will clock at 600-700mhz. If that's true, the pro would be an overclocker's darling indeed...especially if all the vertex shaders were intact on the lesser card. In that case, it would all boil down to the efficiency of the pipes...which is again why the xt may be 24 at a lower clockspeed, if they're 24 very efficient fatty pipes.

[crodan85]

[QUOTE=DilTech]Crodan, that article was written by fuad, aka Fraud, we all know we can't trust anything he has to say...[/QUOTE

Ok but better to have any news than no news even if it is eventually confirmed to be wrong.

[turtle]

I wouldn't put the Inq's sources as any more reliable than the three or four people i've seen claiming to know ATi partners/fab peeps (that all posted the same specs) Surely, they could all be wrong, without a doubt. Like I said, it's all speculation and/or news, some which may be outdated.

[.sentinel]

I would expect them to release it 24/500 rather than 16/700 because normally people will only look at whats higher in numbers for pipelines and say it is better. On the other hand my friend said get this card it is 100 dollars cheaper and is only 50 mhz slower when it was 5 generations old a 9600Xt rather than a 6800.
This is going to be one interesting SpecWars