GlobalFoundries Details Plans For 2011 And Beyond
Ties That Bind: GlobalFoundries, Bulldozer, (And Bobcat?)
Now that we've discussed GlobalFoundries' plans in general, let's take a gander at what we can expect from the company's 32nm SOI and 28nm bulk silicon processes. Both processes will use what's known as high-k metal gate technology. High-k can be built either through the use of a gate-first or a gate-last approach. Intel—the only company to date that's shipped high-performance / high volume products using high-k—favors the gate-last approach. GlobalFoundries, IBM, and Samsung are championing gate-first technology.
Here's how the nerdfight breaks down. Proponents of gate-first technology argue that it allows for less-restrictive designs, smaller dies, and is easier to build. It's also cheaper—GF expects to save ~$75 million over four years by opting for gate-first rather than gate-last. Gate-last proponents (Intel, TSMC) claim that gate-first technology isn't a long-term solution and can't deliver all the benefits of the gate-last approach. We won't actually know which approach makes more sense until we see shipping silicon.
What we do know is that the shift to high-k metal gate manufacturing has significant implications for Bulldozer and, by extension, Bobcat. If you remember Intel's Prescott, you should also remember that the CPU's high temperatures and significant power consumption made it a lose/lose proposition for enthusiasts almost across the board. One of the reasons for this was the amount of current that 'leaked' out of the CPU—at just 1.2nm thick on a 90nm process, the gate dielectric material was unable to prevent electron tunneling.
Prior to 90nm, both Intel and AMD shrank gate thickness by ~0.7x per process, right up to the point where gate leakage, which was already present at 130nm, skyrocketed. Intel didn't change gate thickness at 65nm, but was able to do so again at 45nm because switching from Poly/SiON to high-k reduced the amount of gate leakage exponentially. AMD's processors up to and including the Phenom II, have yet to make this switch. Here's GlobalFoundries own prediction for gate scaling at 32nm and below.
AMD's curve looks much the same as Intel's, with the exception that AMD held gate lengths constant through 90nm, 65nm, and 45nm. At 32/28nm, however, power leakage will drop while gates again scale downwards. Over the past few weeks, we've wondered how Bulldozer's power consumption and performance will compare to Sandy Bridge. We're nowhere near filling in that particular puzzle, but we do know, at least, that Bulldozer will be fighting with an advantage previous generations of AMD chips haven't had.
When Will Bobcat Make The Leap?
AMD's first generation of Ontario processors (dual-core Bobcat + GPU) will be built by TSMC on that company's 40G (high performance) process. Left unanswered is the question of whether or not AMD will stick to TSMC at the 28nm node or move to GlobalFoundries. The peculiarities of 28nm technology make this a more important question than it might have been at previous nodes. TSMC currently plans to launch two 28nm processes. One will use high-k metal gate and a gate-last approach while the other is built around conventional Poly/SiON. There are questions regarding the suitability of Poly/SiON at 28nm and TSMC has, as we've noted, opted for the more difficult gate-last approach.
Even if we assume that TSMC executes both approaches perfectly, the inherent differences between gate-first and gate-last mean that there's no simple way to jump from one to the other. This difference will also apply to Radeon chips; switching from one foundry to the other at or below 28nm will necessitate significant and time consuming design/material changes.
AMD may still be keeping its options open and watching to see how the two approaches play out in the real world, but we'd be surprised if the company doesn't opt for one or the other before the end of the year.
Here's how the nerdfight breaks down. Proponents of gate-first technology argue that it allows for less-restrictive designs, smaller dies, and is easier to build. It's also cheaper—GF expects to save ~$75 million over four years by opting for gate-first rather than gate-last. Gate-last proponents (Intel, TSMC) claim that gate-first technology isn't a long-term solution and can't deliver all the benefits of the gate-last approach. We won't actually know which approach makes more sense until we see shipping silicon.
What we do know is that the shift to high-k metal gate manufacturing has significant implications for Bulldozer and, by extension, Bobcat. If you remember Intel's Prescott, you should also remember that the CPU's high temperatures and significant power consumption made it a lose/lose proposition for enthusiasts almost across the board. One of the reasons for this was the amount of current that 'leaked' out of the CPU—at just 1.2nm thick on a 90nm process, the gate dielectric material was unable to prevent electron tunneling.
Prior to 90nm, both Intel and AMD shrank gate thickness by ~0.7x per process, right up to the point where gate leakage, which was already present at 130nm, skyrocketed. Intel didn't change gate thickness at 65nm, but was able to do so again at 45nm because switching from Poly/SiON to high-k reduced the amount of gate leakage exponentially. AMD's processors up to and including the Phenom II, have yet to make this switch. Here's GlobalFoundries own prediction for gate scaling at 32nm and below.
AMD's curve looks much the same as Intel's, with the exception that AMD held gate lengths constant through 90nm, 65nm, and 45nm. At 32/28nm, however, power leakage will drop while gates again scale downwards. Over the past few weeks, we've wondered how Bulldozer's power consumption and performance will compare to Sandy Bridge. We're nowhere near filling in that particular puzzle, but we do know, at least, that Bulldozer will be fighting with an advantage previous generations of AMD chips haven't had.
When Will Bobcat Make The Leap?
AMD's first generation of Ontario processors (dual-core Bobcat + GPU) will be built by TSMC on that company's 40G (high performance) process. Left unanswered is the question of whether or not AMD will stick to TSMC at the 28nm node or move to GlobalFoundries. The peculiarities of 28nm technology make this a more important question than it might have been at previous nodes. TSMC currently plans to launch two 28nm processes. One will use high-k metal gate and a gate-last approach while the other is built around conventional Poly/SiON. There are questions regarding the suitability of Poly/SiON at 28nm and TSMC has, as we've noted, opted for the more difficult gate-last approach.
Even if we assume that TSMC executes both approaches perfectly, the inherent differences between gate-first and gate-last mean that there's no simple way to jump from one to the other. This difference will also apply to Radeon chips; switching from one foundry to the other at or below 28nm will necessitate significant and time consuming design/material changes.
AMD may still be keeping its options open and watching to see how the two approaches play out in the real world, but we'd be surprised if the company doesn't opt for one or the other before the end of the year.
