Re-evaluating Strategy, Creating the RV770 in 2005

Remembering what lead to success rather than the success itself is generally more important, and around 2005 ATI had another R300 moment. What made R300 successful was ATI re-evaluating the way it made GPUs and deciding on something that made sense. At the time, it made sense for ATI to work towards building the biggest chip possible, win at the high end and translate that into better, more competitive products at all other price points. It is worth mentioning that part of R300’s continued success was due to the fact that NVIDIA slipped up royally with the GeForce FX, it was the perfect storm and ATI capitalized. ATI built the biggest chip and NVIDIA had no real answer for some time to come.

Just as Intel bumped into the power wall when it tried to ramp Pentium 4 clock speeds to ridiculous levels, ATI was noticing that GPUs were getting a little too big for comfort. Design cycles were getting longer, die size was growing and financially there were some things that didn’t quite make sense when it came to making these large GPUs.

ATI asserted that a very large GPU, without the use of repair structures or harvesting (two techniques I’ll describe in a bit) may only have a 30% yield. That means for every 100 GPUs produced on a single wafer, only 30% of them would be fully functional and could be sold as advertised. Manufacturing these complex circuits is particularly expensive, those fab plants easily cost a couple of billion dollars to build and thus having most of your wafer go to waste isn’t really a good way of running a business.

Thankfully there are ways to improve that crappy 30% yield to a value north of 90%. The first is a technique generally known as repairability. The idea behind repairability is simple: build redundancy into your design. If you have a unit with 10 shader processors, actually build 11 but use one as a spare should there be a defect in any of the remaining processors. The same technique is often used for on-die memory, include additional cells so that a defect in one part of a cache won’t render the entire chip unusable. There are tradeoffs here however, if you build in too much repairability then you run the risk of having a die that’s large but offers no additional performance benefit. If you don’t build in enough repairability then your yield suffers.

Harvesting is another technique that all of you are quite familiar with, although its street name is feature binning. The idea here is that you assume that only a small percentage of the die on a wafer will be completely functional, but instead of throwing away the ones that aren’t fully functional, you disable features and sell them as a lower end part. For example, if your target architecture has 10 shader processors but only 30% of your production will have 10 working shader processors but 50% of the production will have 8 working units you’d sell a model with 10 SPs and a model with 8 SPs, the latter being at a reduced cost. This way instead of throwing away 70% of the wafer, you only throw away 20% - and if possible you could make a third SKU with fewer number of functional SPs and use close to 100% of the wafer.

Yields do improve over time, and what eventually happens if you build a product relying on harvesting is you hurt yourself financially. In the example above where you have two configurations of the same die, yields will eventually improve to the point where your 8 SP die could just as easily have 10 functional SPs but you have to sell them as reduced functionality parts because you’ve built up market demand for them. In a sense, you’re selling something worth $1 for $0.80. You could always create a new version of the chip with a smaller die, but that takes additional engineering resources, time and money.

The most recent public example of die harvesting was actually with NVIDIA’s GT200. The GeForce GTX 260 was simply a harvested version of the GTX 280 with 192 SPs vs. 240. As yields improved, NVIDIA introduced the GeForce GTX 260 Core 216 with 216 SPs. Note that NVIDIA didn’t use any repairability for the shader processors on GT200 there are no more than 240 functional SPs on a GT200 die, but that’s because NVIDIA would rely on harvesting anyway - if any of the 240 SPs didn’t work the chip would be branded as a GTX 260 or Core 216.

This is in stark contrast to RV770 where both versions of the GPU were functionally identical, the only differences were clock speeds and power requirements. But from a die standpoint, ATI didn’t have to sell any reduced versions of the chip to keep yields high.

Note that whether or not a company has to rely on harvesting doesn’t really impact the end user. In this case, the GeForce GTX 260 Core 216 actually benefitted the end user as it manages to deliver the same sort of value that AMD does with the Radeon HD 4870. It isn’t ideal for NVIDIA financially and if a company has to rely on harvesting for too long it will eventually take its toll, but I suspect that in the GT200 case NVIDIA will bail itself out with a 55nm revision of GT200 before things get too bad.

ATI didn’t want to build a GPU that would rely on excessive repair and harvesting to keep yields high. And then came the killer argument: building such a GPU was no longer in the best interests of its customers.

Sidenote: Building a GPU

Before we get too far down the rabbit hole it’s worth adding some perspective. ATI’s RV770 design took three years of work, that means ATI started in 2005. It takes about a year for manufacturability from tapeout to final product ship, another 1 - 1.5 years for design and another year for architecture.

The team that I met with in Santa Clara was talking about designs that would be out in 2012, we’re talking 4 TFLOPS of performance here, 4x the speed of RV770.

Decisions being made today are going to impact those products and in a sense it’s a big guessing game, ATI has no idea what NVIDIA will have in 2012 and NVIDIA has no idea what ATI will have. Each company simply tries to do the best it can for the customers and hopes they’ve got the right mix of performance, price and power.

To give you an idea of the intensity of some of these debates, Carrell debated, every evening, for three hours, for three weeks, the amount of repairability that would go into the RV770. Building redundant structures in the RV770 was an even bigger deal than before because ATI was going to be working within serious die constraints, simply making the chip bigger wasn’t an option. Repair structures in the RV770 wouldn’t add any performance but would increase the die size, on an already large chip that’s fine, but on a chip that’s supposed to target the sweet spot and can’t grow without limit it’s a big deal.

The Beginning: The Shot Heard Around the World Building a GPU for the Masses
POST A COMMENT

116 Comments

View All Comments

  • pastyface - Wednesday, December 03, 2008 - link

    Great job on the article! Generally today's reviews consist of me quickly going to the benchmarks portion and seeing if a new game was used or if any screwy results came out. This article however was much different. You had my attention from the get go and I didn't take a break in my reading until the whole article was finished.

    It is a real shame that so much of the work in reviews are overlooked in favor of simple graphs but this article was different and I thank you for that.
    Reply
  • MarchTheMonth - Wednesday, December 03, 2008 - link

    I really enjoyed the read, and it really gives me an appreciation for the card i just happen to get (hd 4850).

    I may not speak for others, but these are the kind of articles I like to read, the kind that really explain in detail what's really happening. Anand, you did an excellent job of giving perspective (be in ATI's shoes in 2007 when nvidia was doing this...etc) to the article that gave definition between the "so obvious" hindsight we have now to the "this is suicide!" view that it must have seemed like to be there in 2005.

    Now, for my own counter-perspective, I can understand why AMD, Intel, and nVidia may not do this very often. On the flip side of the coin, I'm not a mainstream user, and I don't exactly build 1000s of computers that ATI can sell. Bottom line speaking, a story that's interesting to me, I don't bring them $$$$. And on top of that, this story is also giving a lot of info to the competition, which can be at best a double edged sword, and at worst too much information to be self-destructive.
    Reply
  • belladog - Wednesday, December 03, 2008 - link

    Excellent article, we love this stuff. Benchmarks gets a bit boring after a while.

    I wonder what affect if any, the revelations about price fixing(between ATI and Nvidia) had on the pricing of the RV770 GPU's??

    I mean if the lawyers hadnt broken up the party maybe the 4870 could have been $80-$100 dearer?

    Anyway, interesting article.
    Reply
  • Griswold - Wednesday, December 03, 2008 - link

    The price fixing took part before AMD bought ATI. And it would be safe to assume that it stopped at the latest at that time, but it probably did stop well before that (the earliest evidence is an e-mail from 2002). AMD should know better than to point the finger at Intel and do something that is equally wrong in another segment of their business. Reply
  • feelingshorter - Wednesday, December 03, 2008 - link


    Keep up the good work and never let the haters get you down! There's always people b!tching when they don't know how hard it is to write well (any moron can "write"). But its good stories like this that has been the bread and butter of Anandtech.

    The pressure of deadlines, writer's block, or not having enough to write. I appreciate what you do and I know its stressful at times. Others can sympathize but I can empathize having been an amateur journalist myself (in high school and at the university newspaper).
    Reply
  • san1s - Wednesday, December 03, 2008 - link

    If this article was about an nvidia GPU then the ati fanboys would proclaim it reeks of bias.
    good article though anamdtech
    Reply
  • Adul - Wednesday, December 03, 2008 - link

    That was one of the best article I read in a while. It was very enjoyable to get an idea of how things are worked out years in advanced of when the product launches.

    This was a huge gamble on AMD/ATI part. My hats off to them for having the balls to do something different.
    Reply
  • dani31 - Wednesday, December 03, 2008 - link

    Speaking of AMD, it would have been nice to have more insight on how the acquisition of ATI fitted in the design process.

    But this topic seems to have been deliberately omitted in this article.
    Reply
  • JonnyDough - Wednesday, December 03, 2008 - link

    Maybe that's because the interviewed the ATI chip designers and not the AMD head haunchos? Just a thought. Reply
  • lifeobry - Wednesday, December 03, 2008 - link

    Really fascinating article. The amount of work put into creating these cards and the competition between the two companies is compelling stuff. Reply

Log in

Don't have an account? Sign up now