WORD JUST IN from the far east says that TSMC may have canceled its 32nm process node. On the surface this might seem very bad, but it is much more nuanced than that.
Update: The knifing of 32 has been confirmed by multiple sources.
SemiAccurate has not heard it directly from TSMC yet, but given the circumstances, we believe this report to be accurate. As we pointed out earlier in the week, some of the 32nm processes at TSMC are delayed and somewhat underperforming. Given how well 40nm has shaped up, or not, the earlier news is likely only the tip of the iceberg. As far as we can tell, this will be the first time TSMC has canceled a node entirely.
Three scenarios come to mind as to why this happened, a good, a bad, and a sarcastic set of possibilities. Lets start with the good. The 32nm node was delayed, but delays are pretty normal for TSMC lately, and is likely more indicative of overly eager schedule makers than any serious technical problem. Marchitecture meets graphing software, all in the hands of salespeople, a dangerous combo.
So, the 32nm node slipped a month or two, and was almost assuredly going to slip a little more. At that point, the rather small delay between 32nm and 28nm became vanishingly small, a few months. 32nm is the unloved process variant of late anyway, Global Foundries dropped it a long time ago to focus on making a 40nm process along with one at 28nm. It said that was because of low customer demand, and from what I understand, uptake on TSMC’s 32nm is about as strong.
Given the lack of customers, and the need for engineers on 40nm and 28nm, it only made sense for TSMC to stop beating its head against the wall, and focus on where the money is, or will be. Since the delay will only be a few more months from when 32nm is scheduled, now, customers won’t mind that much. They can easily be placated with a well timed discount.
The good version is that it is a minor delay, customers don’t really mind, and will lead to better 28nm and 40nm ramps. All is happy in Hsinchu, time to hit the night market for some celebratory stinky tofu.
If you are a bit more jaded, the bad version might be more to your liking. That one says that TSMC can’t make 40nm wafers if their company depends on it, and it does. It is a year late, yields are in the toilet, and there is no light at the end of the tunnel yet.
The 32nm process was flawed, badly, late and power hungry. High-K and Metal Gate (HKMG) technology wasn’t working out, and there was no way things would be ready by the time TSMC promised they would be. Canning 32nm gave TSMC a few more months of running around in circles trying to fix the mounting problems that it had. Hopefully, this will be enough time to figure out why things are broken before all the customers jump ship. It is breathing room.
All is lost, TSMC is clueless, and in general, it doesn’t have a plan. The Magic 8-ball of process tech and material science can only be shaken so fast, and dropping 32nm gives it that 8-ball, plus the one allocated for 28nm, to shake and hope that it comes up right before the company has to deliver wafers.
In short, the ‘bad’ interpretation of events is that there is a huge technical problem, with no solution in sight. No night market trip in Hsinchu tonight, no stinky tofu either.
The cynical and sarcastic version is that Global Foundries doesn’t seem to have any problems with making 28nm parts, and it doesn’t seem to have 40nm problems either. It gave up on 32nm a while ago to focus on 40nm, and that worked for it.
TSMC is simply trying to be more like GloFo, and the first step there is to can 32nm and ramp 40nm. Since this is exactly what TSMC needs to do, it is just following a proven path. Expect it to change its logo from red to a more orangy color any day now. The night market trip will be accompanied by enough Gao Liang booze to float the stinky tofu, and insight will flow from this.
In the end, the truth is probably closer to the happy version. 32nm has weak uptake everywhere other than the CPU world, and TSMC is not the only one with that problem. 28nm is very close behind, and that is where the customers and volume are, again, industry wide.
Since 32nm slipped a bit, the six months or so between the process nodes became four months or less, and 32nm seemed like a pointless exercise in spending money. If it slipped any more, there was no point to it at all, and that slip was probably plan B if TSMC didn’t stop it entirely.
While knifing 32nm does make a lot of sense, it still doesn’t boost our confidence in TSMC. The company really needs some good news after the 40nm debacle(s) of late, and today’s killing doesn’t bring anyone nearer to that point.
The only scary thing that is looming now is that 28nm is a half-step from 32nm. Any problems that 32nm encountered are going to be the same or worse at 28nm, as it is basically the same thing with even smaller tolerances. If TSMC doesn’t have an ‘aha’ moment soon, the second half of 2010 could be hard times for the stinky tofu vendors at the night markets in and around Hsinchu.S|A
Latest posts by Charlie Demerjian (see all)
- Thing go bump(gate) in the night for Nvidia’s GP100 Pascal GPU - May 3, 2016
- Cavium’s Octeon TX blends compute and packet moving - May 2, 2016
- Qualcomm releases Zeroth API to developers - May 2, 2016
- Another detail about Qualcomm server SoCs revealed - Apr 27, 2016
- AMD finally really honestly launches the dual Fury - Apr 26, 2016