What is going on with AMD’s (NYSE:AMD) Bulldozer and why did it only ship last week? That is a long story, and a complex one, filled with enough minor and major problems to write a medium length book about. We will try and do it in a slightly condensed version.
The short story is that Bulldozer is delayed from where it was in June, and a little bit, but not all that much, from what we said a few weeks ago. We said within two weeks on August 4th, and it was just over a week later than that. Why the mess? Well, to paraphrase one pundit, “The nice thing about having 10 fingers is that you can point in a lot of directions”. Bulldozer delays come down to a lot of little things, and three big ones.
The first one is that there is a new stepping coming, SemiAccurate is hearing mid- to late Q1/2012 for the next rev. That rev is said to bump performance, specifically integer performance, up by quite a bit, and possibly improve clocks too. Either way, it looks like that stepping is the one to keep an eye out for. It isn’t a Barcelona type fiasco, but it isn’t an HD4870 launch either.
Next on the list is yield, or lack thereof. The same problems that affect Llano affect Bulldozer, and we took a look at those earlier. The short version is that GloFo has problems with the 32nm SHP process, but they are still best in class. No other foundry can do a 32nm HKMG + strain process, so they are currently the only choice. AMD pays only per good die, so they are financially insulated from the yields.
More problematic for AMD is that they can’t get enough chips. Most observers agree that if GloFo could make more chips, AMD could easily sell them. Coupled with bad dies, this means a lot of money not earned for both companies. The situation would be much better for AMD if Intel used a foundry, they would basically be in the same boat. Fortunately for Intel, and unfortunately for AMD, they don’t, and 32nm HKMG + strain is not a big deal for them.
What we are hearing is that currently, the delays are due to availability of product. Bulldozer is built on the same process as Llano, 32SHP, so the yield problems are going to hit that chip at least as hard as Llano. Llano has a large portion of it’s area taken up by GPU shaders, so you can salvage a lot of lower binned dies because of that, and more by going to 3 or 2 cores. Assuming any single module can be fused off in Bulldozer, it could have a very high resistance to defects as well.
That said, Bulldozer is a larger die, about 50% larger than Llano, so yield should go down noticeably. Llano’s yield, to use the technical term, sucks, so Bulldozer should be at least as bad. Time will tell, but the constant slipping by a week or two here or there is a pretty good sign that this is the case. When Bulldozer is released in volume, what SKUs are offered, and their availability should give a lot of clues about yields and bin splits.
The last thing that could delay Bulldozer’s release is simple, money. Global Foundries can only run so many 32nm wafers at the moment, and that is not going to go up dramatically in the near future. Assuming AMD is the only customer for that process, it is up to AMD to allocate the wafers to one chip or the other. Yield, die size, and ASPs all come in to play on this one, as does expected process improvements.
If you want to look at it the simple way, on average, each wafer will yield X Llanos or Y Bulldozers. On average, each Llano will sell for $A, and each Bulldozer will sell for $B. Assuming that AMD will be able to sell everything they make, likely at this point, so the math is simple. If X * A is bigger than Y * B, make more Llanos. If X * A is smaller, make more Bulldozers. Yield improvements, customer bitching, and many other things come in to play to influence the split, but the gross calculation is simple. You run the one that makes the most money in the end.
So Bulldozer keeps slipping. Several hard deadlines have passed, but each time, the difference between the current date and the expected one keeps shrinking. When will it ship? If we had a hard date, we would not be able to say what it is. If we don’t, we don’t, so that is about what you will get, nothing. The whispers in the background are still saying that it isn’t long though, and the whispers still say that it is not September 19th.S|A
Latest posts by Charlie Demerjian (see all)
- Spin Transfer Technologies talks about their ST-MRAMs - Sep 27, 2016
- ARM adds CMN-600 interconnect and DMC-620 memory controllers - Sep 27, 2016
- Globalfoundries 7nm process isn’t even close to the name - Sep 26, 2016
- ARM upgrades realtime offerings to v8-R and adds Cortex-R52 - Sep 21, 2016
- Everspin and Globalfoundries team up for embedded ST-MRAM - Sep 15, 2016