The RV770 Story: Documenting ATI's Road to Success

Depression Sets in but the Team Goes On

The entire RV770 design took around three years, which means that while we were beating ATI up over the failure that was R600, those very engineers had to go into work and be positive about RV770. And it was tough to, after all ATI had just completely lost the crown with R600 and Carrell, Rick Bergman and others were asking the team to ignore what happened with R600, ignore the fact that they lost the halo, and try to build a GPU that aimed at a lower market segment.

Through all of my interviews, the one thing that kept coming up was how impressed ATI was with the 770 team - never once did the team fall apart, despite disagreements, despite a shaky direction, the team powered through.

The decision not to go for the king of the hill part was a decision that made a lot of sense with ATI, but there was so much history about what would happen if you didn’t get the halo part; it took a very strong discipline to cast history aside and do what the leads felt was right, but the team did it without question.

The discipline required wasn’t just to ignore history, but to also fight the natural tendency for chips to grow without limits during their design phase. What ATI achieved with RV770 reminded me a lot of Intel’s Atom design team, each member of that team had strict limits on how big their blocks could be and those limits didn’t waver.

Adversity tends to bring the best out of people. The best stories I’ve been told in this industry, the Intel folks who made Banias and the ATIers that were responsible for RV770 put their hearts and souls into their work, despite being beat down. Passion has a funny way of being a person’s strongest ally.

The Power Paradigm

We were all guilty for partaking in the free lunch. Intel designed nearly five years of processors without any concern for power consumption and the GPU guys were no different.

In the R300 and R420 days ATI was almost entirely ignoring power, since estimating how much power the parts would use was so off from the final product that they just didn’t care. It was such a non-issue in those days that ATI didn’t even have a good way to estimate power even if it wanted to, it was impossible to design for a specific TDP. Today ATI’s tools are a lot better, now targeting a specific TDP is no different than aiming for a specific clock speed or die size, it’s another variable that can now be controlled.

These days power doesn’t change much, the thermal envelopes that were carved out over the past couple of years are pretty much stationary (ever wonder why the high end CPUs always fall around 130W?). Everyone designs up to their power envelope and stays there. What matters now is every year or two increasing performance while staying within the same power budget. Our processors, both CPUs and GPUs, are getting more athletic, rather than just putting on pounds to be able to lift more weight.

One of the more interesting things about architecting for power is that simply moving data around these ~1 billion transistor chips takes up a lot of power. Carrell told me that by the time ATI is at 45nm and 32nm, it will take as much power to move the data to the FPU as it does to do the multiply.

Given that data movement is an increasingly power hungry task a big focus going forward is going to be keeping data local when possible, minimizing moving to registers and on-chip caches. We may see more local register files and more multi-tiered memory hierarchies. As chips get more complex, keeping the register file in one central location becomes a problem.

ATI admitted to making a key manufacturing mistake with R600. The transistor technology selected for R600 was performance focused, designed to reach high clock speeds and yielded a part that didn’t have good performance per watt - something we noticed in our review. ATI has since refocused somewhat away from the bleeding edge and now opts for more power efficiency within a given transistor node. With leakage a growing problem as you go to smaller transistors it’s not worth it to be super leaky to gain a few picoseconds. If you’ve got a 100W GPU, do you want to waste 40W of that budget on leakage? Or would you rather do 80W of real work and only waste 20W? It’s the same realization that Intel recognized during the Pentium 4’s term and it’s the mentality that gave us the Core microarchitecture. It’s an approach that just makes sense.