Did anyone every figure out why R600 has a 512-bit bus?

[hoom]

You're asserting that both GDDR3 wouldn't continue scaling and GDDR4 would be late

I'm suggesting that around the time ATI was deciding the fundamental specs of R600 (I'm guessing about R420 launch time?) that ATI didn't expect to see GDDR3 go as high as it did by the time R600 launched, nor for GDDR4 to be around until the (smaller process, smaller MC) refresh.
There were posts here at B3D from ATI guys complaining about how core processing power was vastly outstripping the scaling of memory around then.

 

[Jawed]

I'm suggesting that around the time ATI was deciding the fundamental specs of R600 (I'm guessing about R420 launch time?) that ATI didn't expect to see GDDR3 go as high as it did by the time R600 launched, nor for GDDR4 to be around until the (smaller process, smaller MC) refresh.
There were posts here at B3D from ATI guys complaining about how core processing power was vastly outstripping the scaling of memory around then.

It appears they were out by a factor of 2. It's such a ludicrous amount that "didn't understand the memory market" doesn't explain it all.

Particularly as little of R600 is 2x the performance of its predecessor - it really is a feature inflection GPU (D3D10) not a performance inflection.

Though fp16-based data is a major focus of the architecture, and that does like bandwidth. HDR+AA render targets chew 2x the bandwidth of non-HDR+AA. And fp16 texture formats (which aren't compressed like their int8 brethren) consume way way more bandwidth per texel (but are barely used - e.g. only for render target tone-mapping, a fraction of the total frame-rendering workload). So there are spikes in R600 usage that could well chew through all its instantaneous bandwidth.

I agree that memory has been surprisingly kind in its capabilities these last 2-3 years. RV530 had a hell of a lot of extra bandwidth, too (I guesstimate X1600XT would have performed the same with about 16GB/s, instead of 22.4GB/s, i.e. it had "40% surplus").

HD2600XT should provide a control point in this argument. Its design is barely younger than R600, yet with GDDR4 it's not got "100% surplus bandwidth". I haven't tried to work out what kind of surplus it appears to have, but at a rough guess, 20-40%?

Jawed

 

[swaaye]

So they keep giving their cards extra bandwidth to prove a design concept? But, if they keep doing it with each generation, how can we really say it's "forward looking"? I think I see them more as over-achieving, unbalanced designs that were wasteful and costly. Heh.

 

[rwolf]

Speculation: I suspect they expected the chip to have much greater performance than it does.

 

[Bouncing Zabaglione Bros.]

Speculation: I suspect they expected the chip to have much greater performance than it does.

As I said upstream, I think they were speccing the memory based on much higher GPU clockspeeds, but R600's power leakage put paid to that.

 

[fellix]

As for the GPU speed, it was already clarified that the initial design was never expected to launch at anything above 800MHz. I think ATi engineers are pretty well aware of the process technology limits and trends, at the time of for the tape out.

 

[ShaidarHaran]

As for the GPU speed, it was already clarified that the initial design was never expected to launch at anything above 800MHz. I think ATi engineers are pretty well aware of the process technology limits and trends, at the time of for the tape out.

"B" as in B, "S" as in S.

 

[Megadrive1988]

haven't read the whole thread yet.

couple of quick thoughts....


remember when the Matrox Parhelia-512 came out in 2002, it was the first consumer GPU to have a 256-bit bus.

I'm not trying to draw an exact parallel to R600 with its 512-bit bus, because the Parhelia had no bandwidth-saving features, whereas ATI/AMD GPUs do.

though clearly the Parhelia and R600 being the first to offer 256-bit and 512-bit bus respectively, have not reaped a huge benefit compared to pre-Parhelia and pre-R600 GPUs with narrower busses.


certainly for midrange to upper-mid range, 256-bit is going to be the norm for some time. but you don't think 512-bit will go away do you?

is there even a possibility that future architectures in the next 3-5 years will go beyond 512-bit? is a 768-bit or 1024-bit bus even conceivable?

 

[dizietsma]

Err, which "previous" card?

The X1950XTX of course which had a 512bit bus. In the shape of a ring no less. Like I said, big numbers sell cards no matter

 

[fellix]

The internal wiring doesn't count here (for the comparison purpose), only the DRAM interface.

 

[swaaye]

I have to say that, with its massive texturing hardware, I wonder how much faster G80 would be with a 512-bit bus and its accompanying bandwidth.

 

[fellix]

The ROP hardware will eat it before TMUs would see any significant advantage.

 

[Megadrive1988]

The X1950XTX of course which had a 512bit bus. In the shape of a ring no less. Like I said, big numbers sell cards no matter

oh hehe, then why not count the 2560-bit bus in PS2's Graphics Synthesizer

 

[Dave Baumann]

Did anyone ever figure out why R600 has a 512-bit bus?

Yes.

 

[Geo]

The question is compared to what? That it didn't stomp a 384-bit G80 doesn't mean it doesn't stomp a 256-bit R600 with the same speed GDDR3. Nvidia clearly revamped their design (from the previous generation) to tie bit-ness to other functional units. Those decisions would have been made quite aways in advance. I think I would agree that the "Legend of Bandwidth as King" has taken some knocks of late, and may take some more.

 

[digitalwanderer]

Did anyone ever figure out why R600 has a 512-bit bus?

Yes.

Classic.

 

[Jawed]

My reference point is R580...

Jawed

 

[swaaye]

The ROP hardware will eat it before TMUs would see any significant advantage.

Well I was trying to refer to everything except the shader core really, cuz I think it's been accepted that G80's texturing/ROP : shader ratio is a little lopsided.

 

[Pantagruel's Friend]

I bet the R600 based FireGL card loves that bandwidth