THE SAGA of Nvidia’s GT300 chip is a sad one that just took a turn for the painful when we heard about first silicon yields. Nvidia’s execution has gone from bad to absent with low single digit yields.
A few weeks ago, we said that Nvidia was expecting first silicon back at the end of the week, the exact date was supposed to be Friday the 4th plus or minus a bit. The first bit of external evidence we saw that it happened was on the Northwood blog (translated here) and it was a day early, so props to NV for that. That lined up exactly with what we are told, but the number of good parts was off.
The translation, as we read it, says there were nine good samples that came back from TSMC from the first hot lot. That is below what several experts told us to expect, but in the ballpark. When we dug further, we got similar numbers, but they were so abysmal that we didn’t believe it. Further digging confirmed the numbers again and again.
Before we go there though, lets talk about what a good die is in this case. When you get first silicon back, it almost always has bugs and problems. First silicon is meant to find those bugs and problems, so they can be fixed in succeeding steppings.
By ‘good’, we mean chips that have no process induced errors, and function as the engineers hoped they would. In other words not bug free, but no more errors than there were in the design. ‘Good’ in this sense might never power on, just that the things that came out of the oven were what was expected, no more, no less.
Several experts in semiconductor engineering, some who have overseen similar chips, were asked a couple of loaded questions: What is good yield for first silicon? What is good yield for a complex chip on a relatively new process? The answers ranged from a high of 50% to a low of 20% with a bunch of others clustered in the 30% range. Let’s just call it one-third, plus or minus some.
The first hot lot of GT300s have 104 die candidates per wafer, with four wafers in the pod Nvidia got back a week and a half ago. There is another pod of four due back any day now, and that’s it for the hot lots.
How many worked out of the (4 x 104) 416 candidates? Try 7. Yes, Northwood was hopelessly optimistic – Nvidia got only 7 chips back. Let me repeat that, out of 416 tries, it got 7 ‘good’ chips back from the fab. Oh how it must yearn for the low estimate of 20%, talk about botched execution. To save you from having to find a calculator, that is (7 / 416 = .01682), rounded up, 1.7% yield.
Nvidia couldn’t even hit 2%, an order of magnitude worse than the most pessimistic estimate. Ouch. No, just sad. So sad that Nvidia doesn’t deserve mocking, things have gone from funny to pathetic.
At this point, unless there’s a massive gain in yields on the second hot lot, there might not be enough chips to do a proper bring up and debug. This stunningly bad yield could delay the introduction of the chip, adding to the current pain and bleak roadmap. If there aren’t enough ‘good’ parts from the second hot lot, that might require running another set, adding weeks to the total. Q1? Maybe not.
It is going to be very interesting to see what Nvidia shows off at ‘Not Nvision’ in a couple of weeks. Will it give the parts to the engineers to work on, or show them off as a PR stunt? We will know soon enough. In any case, the yields as they stand are sub-2%, and the status of the GT300 is far worse than we had ever imagined.S|A
Charlie Demerjian
Latest posts by Charlie Demerjian (see all)
- What platform is Intel’s Diamond Rapids on? - Nov 7, 2024
- What Is Diamond Rapids +1 Called? - Oct 31, 2024
- ARM ‘cancels’ Qualcomm’s IP license - Oct 23, 2024
- Microsoft Hobbles Intel Once Again - Sep 20, 2024
- What is really going on with Intel’s 18a process? - Sep 9, 2024