WHEN WE FIRST started talking about TSMC 40nm delays a year ago, people laughed, but the ramifications of those, and subsequent slips, are becoming clear. The lack of availability of ATI’s RV740/HD4770 chip is tied to this, but not in the obvious way.
Promised for early December
If you look closely at the slide above, it says that the RV740 will be out in late 2008, about a month after Nvidia’s first 40nm parts were scheduled for release. The first cards bearing that chip came out in late April 2009, 5 months late. The first Nvidia 40nm parts came out almost two months later. Actually, they don’t seem to be available anywhere, so lets just call it announced 8 months late.
The problem has been known for a while, leakage. Leaping into action, TSMC repeatedly announced that they know what the problem is, something we don’t doubt. Unfortunately, identifying the problem is the easy part, fixing it is much harder.
Intel spends billions in order to avoid this class of problems precisely because once you get to the point where production doesn’t work, it is quite possibly too late to fix. Customers get annoyed, parts are delayed, and penalties rack up. Bad bad bad. Worse yet, customers end up doing some of the R&D, and attitudes get a bit testy.
Back to the RV740. It was supposed to be released in December 2008, working back about 5 months, the minimum for test silicon, debug, and production, it was a done design about this time last year. We hear the delays are due almost entirely to TSMC’s process problems.
There have been a lot of claims about low yields for RV740, some as low as 20%. Having heard much more exact figures, lets just say those numbers aren’t even close, yields are much higher than that. ATI ran a full production run of wafers, and got tens of thousands of working parts back. The reasoning behind running the wafers was simple, TSMC 40nm has many problems, some of which aren’t seen until you are in a volume run, was to figure all out the problems out.
So, ATI ran RV740 knowing that yields would be low, and that they would have problems. If the problems were solved during the run, more could be put in the oven. If not, the Plan B was to use an R770 variant fused off to fill that market segment. There wasn’t a real down side to it other than perception.
The result of the RV740 test is that TSMC 40nm did indeed have problems, and yields were indeed low. The places that problems lurked were illuminated, and that knowledge was rolled into the whole Evergreen family. Given the lead times between the two families of parts, and the long gestation periods, we are told it gave Evergreen quite a nice leg up. If you look at the simplified version of yield, basically a fixed distribution of defects on a wafer, a 50% yield on a 135mm^2 chip means almost zero on a chip double the size, and essentially zero on something 3x as large. Reality doesn’t mirror that, but it is a good starting point for arguments.
This production run was essential for ATI, and it makes you wonder about Nvidia and how they are going to cope with the same problems given the curious disappearance of the GT215. If they are doing the same thing with the 216 and 218, that is smart, but a 500mm^2 part is still going to be a long stretch.
In the end, there are basically no more R740 based parts for a bit. We hear production will start up again in late summer or early fall. TSMC is promising tolerable yields for July, likely wafers in for July, not out, so September seems like a good bet. Until then, there are a lot of R770 chips to bleed off before the Evergreen family obsoletes them.
The other interesting wrinkle is a squeeze on price. You can now buy HD4850s on Newegg for $89! Think about that, less than two months ago, the HD4770 debuted at $109, and now you can get a faster HD4850 for $20 less. given that, is there a reason to restart 4770 production at all?S|A
Latest posts by Charlie Demerjian (see all)
- You can now buy bare Snapdragon SoCs with 410E and 600E - Sep 28, 2016
- 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