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.

If it Ain’t Broke... Just One Small Problem: We Need a New Memory Technology
POST A COMMENT

115 Comments

View All Comments

  • Sahrin - Monday, January 25, 2010 - link

    Anand,

    I love this piece. Not sure if you'll get notified, but while doing some research on the performance of Hybrid Crossfire, I came back - it was interesting to see the tone of the piece, and hear about the guys at ATI talking vageuly about what would become the 5870. Fascinating stuff, I've got to put a bookmark in my calendar to remind me to come back to this next year when RV970 is released (pending no further difficulties).

    Any chance of a follow-up piece with the guys in SC?
    Reply
  • caldran - Wednesday, December 24, 2008 - link

    the gpu industry is squeezing more and more transistors(SP s or what ever) .it would be energy efficient if it could disable some cores when there is less load than reducing clock frequency and 2D mode.just like in the latest AMD processor.a HD 4350 would consume power less than HD 4850 in IDLE right. Reply
  • bupkus - Wednesday, December 10, 2008 - link

    I couldn't put it down until I had finished.

    Extremely enjoyable write!
    Reply
  • yacoub - Tuesday, December 09, 2008 - link

    "that R580 would be similar in vain"

    You want vein. Not vain, not vane. Vein. =)
    Reply
  • CEO Ballmer - Sunday, December 07, 2008 - link

    You people don't mention their alliance with MS!

    http://fakesteveballmer.blogspot.com">http://fakesteveballmer.blogspot.com
    Reply
  • BoFox - Sunday, December 07, 2008 - link

    LOL!!!!! Reply
  • BoFox - Sunday, December 07, 2008 - link

    Great article--a nice read!

    However...

    From how I remember history:

    In 2006, when the legendary X1900XTX took the world by surprise, actually beating the scarce and coveted 7800GTX-512, I bought it. It was king of the hill from January 2006 until the 7950GX2 stole the crown back for the fastest "single-slot" solution about 6 months later around June 2006, only a few months after the smaller 90nm 7900GTX was *finally* released in April 2006. Everybody started hailing Nvidia again although it was really an SLI dual-gpu solution sandwiched into one PCI-E slot. Perhaps it was the quad-gpu thingy that sounded so cool. It was obviously over-hyped but really took the attention away from ATI.

    GDDR4 on the X1950XTX hardly did any good, since it was a bit late (Sept 2006) with only like 3-4 performance increase over the X1900. Well then the 8800GTX came in Nov 2006 and had a similar impact that the 9700Pro had.

    As everybody wanted to see how the R600 would do, it was delayed, and disappointed hugely in June 2007. The 8800GTX/Ultra kept on selling for around $600 for nearly 12 months straight, making history. 80nm just did not cut it for the R600, so ATI wanted to have its dual-GPU single card REVENGE against Nvidia. And it would be even better this time since it's done on a single PCB, not a sandwiched solution like Nvidia's 7900GX2. Hence the tiny RV770 chips made on unexpected 55nm process! The 3870X2 did beat the 8800GTX in most reviews, but had to use Crossfire just like with SLI. Also, the 3870X2 only used GDDR3, unlike the single 3870 with fast GDDR4.

    But Nvidia still took the attention away from the 3870 series by tossing an 8800GT up for grabs. When the 3870X2 came out in Jan 2008, Nvidia touted its upcoming 9800GX2 (to be released one month afterwards). So, Nvidia stopped ATI with an ace up its sleeve.

    Round 2 for ATI's revenge: The 4870X2. And it worked this time! There was no way that Nvidia could expect the 4870 to be *that much* better than the 3870. Everybody was saying the 4870 would be 50% faster, and Nvidia yawned at that, thinking that the 4870 still couldnt touch the 9800GTX or 9800GX2 when crossfired. Plus Nvidia expected the 4870 to still have the "AA bug" since the 3870 did not fix it from the 2900XT, and the 4870 had a similar architecture. Boy, Nvidia was all wrong there! The 4870 actually ended up being *50%* faster than the 9800GTX in some games.

    So, now ATI has earned its vengeance with its single-slot dual-GPU solution that Nvidia had with its 7900GX2 and 9800GX2 a while ago. With the 4870X2 destroying the GTX 280, ATI does indeed have its crown or "halo".

    Unfortunately, Quad-crossfire hardly does well against the GTX 280 in SLI. We now know that quad-GPU solutions give a far lower "bang-per-GPU" due to poor driver optimizations, etc.. So most enthusiast gamers with the money and a 2560x1600 monitor are running two GTX 280's right now instead of two 4870X2's.. oh well!

    One thing not mentioned about GDDR5 is that it eats power like mad! The memory alone consumes 40W, even at idle, and that is one of the reasons why the 4870 does not idle so well. If ATI reduces the speed low enough, it messes up the Aero graphics in Vista. It would have been nice if ATI released an intermediate 4860 version with GDDR4 memory at 2800+MHz effective.

    Now, I cannot even start to expect what the RV870 will be like. I think Nvidia is going to really want its own revenge this time around, being so financially hurt with the whole 9800 - GTX 200 range plus being unable to release a 55nm version of G200 to this day. Nvidia just cannot beat the 4870X2 with a dual G200 on 55nm, and this is the reason for the re-spins (delays) with an attempt to reduce the power consumption while maintaining the necessary clock speed. Pardon me for pointing out the obvious...

    Hope my mini-article was a nice supplement to the main article! :)
    Reply
  • CarrellK - Sunday, December 07, 2008 - link

    Not bad at all.

    BTW, 55nm has less to do with how good the RV770 is than the re-architecture & re-design our engineers did post-RV670.

    To illustrate, scale the RV770 from 55nm to 65nm (only core scales, not pads & analog) and see how big it is. Now compare that to anything else in 65nm.

    Pretty darned good engineers I'd say.

    Reply
  • BoFox - Sunday, December 07, 2008 - link

    True, and nowhere in the article was it pointed out that since the AA algorithm relied on the shaders, simply upping the shader units from 320 to a whopping 800 completely solved the weak AA performance that plagued 2900's and 3870's. It did not cost too much chip die size or power consumption either. ATI certainly did design the R600 with the future in mind (by moving AA to the shader units, with future expansion). Now the 4870 does amazing well with 8x FSAA, even beating the GTX 280 in some games.

    I wanted to edit my above post by saying that the dual G200 needed to have low enough power consumption so that it could still be cooled effectively in a single-slot sandwich cooling solution. The 4870X2 has a dual-slot cooler, but Nvidia just cannot engineer the G200 on a single PCB with the architecture that they are currently using (monster chip die size, and 16 memory chips that scales with 448-bit to 512-bit bandwidth instead of using 8 memory chips with 512-bit bandwidth). That is why Nvidia must make the move to GDDR5 memory, or else re-design the memory architecture to a greater degree. Just my thoughts... I still have no idea what we'll be seeing in 2009!

    Reply
  • papapapapapapapababy - Saturday, December 06, 2008 - link

    more like uber mega retards, right? if they are so smart... why do they keep making such terrible, horrible, shitty drivers?

    why?

    i really really, really want to buy a 4850, i really do. but im not going to do it. im going to go and buy the 9800gt. And i know is just a re branded 8800gt. And i know nvidia is making shitty @ explosive hardware ( my 8600gt just died) And i know that gpu is slower, older, oced 65nm tech. And that nvidia is pushing gimmicky tricks " physics" and buying devs. but guess what? NVIDIA = good, clean drivers. New game? New drivers there. Fast. UN-Bloated drivers, that work, is is that hard ati? Really. or maybe you guys just suck?

    Im going to pick all tech@ because of that. Thats how much i fkn hate your bloated and retarded drivers ATI. Install ms, framework for a broken control panel? stupid. And whats up with all those unnecessary services eating my memory and cpu cycles? ATI Hotkey poller?, ATI Smart?, ATI2EVXX.exe, ATI2EVXX.exe,NET 2.0 ? always there and the damm thing takes forever to load? Nvidia dsnt use any bloated crap, so why do you feel entitled to polute my pc with your bloated drivers?

    AGAIN HORRIBLE DRIVERS ATI! I DONT WANT A SINGLE EXTRA SERVICE! i just build a pc for a friend. I choose the hd4670, beautiful card, really cool, fast, efficient. I love it. I want one for myself. But the drivers? ARg, i ended up using just the display driver and still the memory consumption was utterly retarded compared to my nvidia card.

    so geniuses? move your asses and fix your drivers.

    thanks, and good job with your hard.
    Reply

Log in

Don't have an account? Sign up now