Honest to god question about the R520

[roguerower]

I was thinking recently, what was the point of ATI going to 90nm? Why would they go through all the problems of overclocking a card that high, and then only giving it 16 pipes. Why not go with 24 or at the least, 20? Wouldn't that cause benchmarks to soar? So main questions are:

1) Why did ATI go 90nm if there are no obvious changes except for the huge clock and core speeds
2) Why is ATI still at 16 pipes while NVidia is at 24?

If any of my logic is wrong, please point it out.

 

[EightySix Four]

Cost to manufacture the chip, 90nm allows it to be smaller, more pipes = bigger... so they make bigger profit margins on each chip sold.

 

[johnnqq]

90nm is cheaper and it's supposed to yield a much lower rate of wasted cards.

 

[Deception]

ATI cards typically rely on clock speed to bring about their performance. In recent generations, NVIDIA cores usually run at slower clock speeds when paired up against comparable ATI cards (like 6800 Ultra - > X800XT as an example).

Anyhow, ATI probably went 90nm so that they could continue to be competitive and ramp up their clock speeds further and further. The 512MB of ram and new ring-bus do play a factor in performance, but the die shrink will probably be realized once (1) enthusiasts are able to get their hands on the XT's and see if they o/c as well as believed and (2) the r580 and later generations should take full advantage of the 90nm process.

Simply put, ATI was due to make a big leap, and they did with this generation. Granted it put them very behind NVIDIA when it comes to this launch, but only time will tell whether or not moving to 90nm was a good decision for them. NVIDIA, on the other hand, most likely did not move to 90nm because (a) they saw no need given their pipe-heavy cores and (b) I *believe* that the mass hard-launch of their 7800 lineup would have been more difficult if they had chosen to go with a die shrink.

Also, try not to compare pipes from one card to the next. Simply put, pipelines are no longer the indicator of GPU performance that they once wear. Comparing ATI's 16p to NVIDIA's 24p is like comparing a 3.0 GHZ P4 to a 3.0 GHZ AMD.

Deception

 

[Goi]

I wouldn't say "traditionally, NVIDIA cores usually run at slower clock speeds when paired up against comparable ATI cards (like 6800 Ultra - > X800XT as an example)." It was only in recent generations that this became the case. The 9800XT had a lower clock than the 5900Ultra for example, and before that, nvidia and ATI were basically neck in neck for clockspeed for every generation. A 90nm process technology means a smaller die size, or more dies per wafer, which lowers cost somewhat. Of course, there's the yield problem, which is always gonna be lower when you're using a new process technology, so it's a tradeoff.

 

[Deception]

Originally posted by: Goi
I wouldn't say "traditionally, NVIDIA cores usually run at slower clock speeds when paired up against comparable ATI cards (like 6800 Ultra - > X800XT as an example)." It was only in recent generations that this became the case. The 9800XT had a lower clock than the 5900Ultra for example, and before that, nvidia and ATI were basically neck in neck for clockspeed for every generation. A 90nm process technology means a smaller die size, or more dies per wafer, which lowers cost somewhat. Of course, there's the yield problem, which is always gonna be lower when you're using a new process technology, so it's a tradeoff.

You are correct. When I said that, I was referring to the most recent gens, but technically you are correct.

I will edit my post accordingly.

Deception

 

[Munky]

Once and for all let me explain the situation about the pipes - forget everything you previously knew about em. The traditional pipeline has a pixel shader and a texture unit (TMU), but they both count as one pipe because they work on one batch of pixels at a time. The x1k pipelines are designed in such a way that the texture units can work independently on a different batch of pixels than the shader units. Moreover, no one texture unit is tied to any certain shader, they're all managed by the scheduling unit, so you could have the texture units working on one group of pixels while the shader units can work on a different group of pixels, and shader #1 doesnt have to pass it's data to TMU #1, it could use another texture unit. This is no lousy 16 pipes, and saying the r520 has 16 pipes is like saying "Bah, the Lancer Evolution only has 4 cylinders, it must be slow as hell".

 

[Pete]

According to them, ATI saved R520 for 90nm because it would only have been (economically? technically?) feasible starting at that manufacturing process. Keep in mind that the new "ring bus" memory architecture takes up a lot of the die space, and I'm sure the efficient conditional branching performance (courtesy of the scheduler and small batch sizes) and separated texture units didn't come cheap in terms of transistors (considering the R520 with 16 "pipes" uses more transistors than the G70 with 24 [slightly more traditional] pipes).

We'll see who positioned themselves better when R580 and nV's 90nm respin of G70 compete in the spring. Me, I think ATI may still be slightly ahead of the game programming curve, even with their four-month delay, considering how the X1600XT struggles to outclass a 6600GT and how the X1800XL merely ties a 7800GT overall.

 

[Polish3d]

Also, try not to compare pipes from one card to the next. Simply put, pipelines are no longer the indicator of GPU performance that they once wear. Comparing ATI's 16p to NVIDIA's 24p is like comparing a 3.0 GHZ P4 to a 3.0 GHZ AMD.

Deception

Not according to the efficiency test done on DH if you're implying that 16pp ATI = 24pp nvidia

 

[Deception]

Those efficiency tests mean absolutely nothing. Who cares which one is faster pipe for pipe (and etc)? The bottom line is that each card comes with different features & specs, so underclocking cards to compare 'apples to apples' performance is entirely moot.

Deception

 

[Munky]

If they really wanted to test efficiency, the'd have to run a shader test like shadermark, or similar, to test only the pixel shader pipes of a gpu. Regular games also depend on other parts of the gpu like vertex shaders, video memory latency, driver efficiency, and dozens of other factors. Not to mention all the app-specific optimizations that might be built in to the driver.