AMD’s Llano has a breakthrough

Tale twistier than the river it is named after

Fusion logoA LOT OF people have been wondering what is going on with AMD’s Llano chip, tipped to arrive before Ontario, now delayed until Q2/2011. The story is a bit complex, but after weeks of digging, we can bring it to you with a new surprise ending.

Earlier this year, Llano was slated to be the first ‘fusion’ chip out the door, beating it’s TSMC fabbed brothers, Ontario/Zacate by a fair margin. Things went well for Ontario/Zacate, and not so well for Llano. The reasons behind that were far less clear than mud, and changed quite a bit depending on who you asked.

Most were quick to blame either AMD or GloFo, sometimes both. Depending on which vantage point the commenter was seeing the action playing out from, the story changed quite a bit. While snide comments, “All I will say is that it taped out on time….”, were rife, details were not. The chip went from Q4/2010 to Q1/2011, then Q2, Q3, and eventually back to Q2.

Where it stands now is production starting in late Q1 with volume ramping hard in Q2. That is likely to change though, for reasons we will get into later, but don’t expect to buy a system before late spring. How it got there is convoluted and rather painful to watch.

The first outward signs of problems were in the Q2 AMD conference call last July. If you recall, Dirk Meyer punted on when Llano would be out, and most people took that as a really bad sign. It was, but only for a very short period of time. The reason for it was simple, at that point in time, AMD really didn’t know for sure what was going on.

That may be a bit of an overexaggeration, but the problems AMD were staring at did not lend themselves to precise prognostication, prestidigitation was more in order. If you recall, Llano is the lead off product for several technologies, most notably AMD’s integrated GPU. It also is the first High-K Metal Gate (HKMG) process for GloFo, the first time they are going under 40nm node, and the first new process on their own. To top it off, it is AMD’s first SOI based GPU, and the first time they are using the really long power gating transistors. There are a lot of places for things to go wrong, and they did. AMD went from, “Yay, we are happy” to “WTF just hit us” in very short order. Sources tell us, and AMD denies, that the first batch of chips out of the fab had a yield of precisely zero.

For a chip with this many ‘firsts’ as Llano, you expect low yields on first silicon, there is a world of difference between low and zero. If you have low yield, you can work your way up. This is common for new processes and technologies, there are a lot of knobs to turn. If you have zero yield, you have a hugely expensive chunk of glass with pretty pictures drawn on it, and not much more. Testing opportunities are limited, and the data you can get off of it is even more limited.

To make the timing more exciting, this all happened a few days before the Q2 CC. Instead of getting up in front of the analysts with an exciting tale of adventure, “We shipped Ontarios this morning”, Dirk had to punt. In public. While smiling. Until the next batch of silicon came back, there wasn’t much to do or say. If that one came back DOA, things would get very ugly. If they didn’t, chips could come out only a little late, hence the rather vague punt.

Luckily, things got much better very quickly. Sources tell SemiAccurate that the next batch of chips came back days after the CC, and yield was infinitely better. By infinitely, we mean that in the mathematical sense, not that yields were production ready, they were not, but progress was steady. Think of it as having the normal, or slightly better than normal, ramp curve, just that it was pushed out further. Mathematically, it can be thought of as (X^2 + 3 months) instead of (X^1.5).

From there, things progressed as they do in the semiconductor world, slowly and methodically. Knobs are turned, and weeks later, results are measured. Rinse and repeat and repeat and repeat until you are confident enough to ramp the volume. Then keep rinsing and repeating, never stop.

Checking in with some people wandering the halls of a certain town north of Bannewitz showed that progress was moving forward, and things were tracking well for the Q2 release. Some tool monkeys said that things may be pulled in a few weeks, but still Q2. Optimism was the key word.

In the last few days, one specific vendor, don’t read anything into that, there are hundreds on site right now, called with some very interesting news. Without getting into specifics, it was cited as a major breakthrough that everyone thinks changed the game. People went from smiles and nods to really excited about the prospects.

As things stand today, there are a lot of people dancing in bunny suits, then going to the local beer hall and getting blitzed. Bad pun. It would not surprise me at all to hear that Llano was pulled in a great deal soon, and late Q2 becomes early Q2, maybe even sooner. Again, in math terms, (X^2 + 3 months) became (X^2 + 2 months), or more likely (X^2.5 + 2 months).

First production silicon may not happen much sooner, but the ramp undoubtedly will. This is better news for AMD shareholders than it is for consumers, but both will benefit. Keep an eye on AMD PR and how they message. If they get very aggressive in tone and posture between now and CES, you know things got pulled in. Either way, I think the budget for AMD’s CES party was just increased.S|A

The following two tabs change content below.

Charlie Demerjian

Roving engine of chaos and snide remarks at SemiAccurate
Charlie Demerjian is the founder of Stone Arch Networking Services and is a technology news site; addressing hardware design, software selection, customization, securing and maintenance, with over one million views per month. He is a technologist and analyst specializing in semiconductors, system and network architecture. As head writer of, he regularly advises writers, analysts, and industry executives on technical matters and long lead industry trends. Charlie is also available through Guidepoint and Mosaic. FullyAccurate