Pentium 6 / Early 1990s large-format chip bonding

Two-chip tab tape (1 cm² processor + ~0.6 cm² memory) circa 2000, used to show the intermediate state before everything moved on-chip.

Back at one point I told you the Pentium 6 had two chips on a tab tape instead of just one. One was about a centimeter square, the other was about six-tenths of a centimeter square, and I think the other was basically a bunch of memory. This was back around 2000, or late '90s. Nowadays they can get so many transistors on a single chip — we're up to almost 100 million transistors on a chip now — they can cram everything onto one chip. But in the old days, which is about the time you were born, they had a problem because the distances the electrons had to go were getting into a range that was limiting the speed.

Strain-energy scaling argument made concrete: chips cannot exceed ~1 cm on a side because the 30–40°C thermal gradient required to extract heat exceeds bond capacity. Intel's solution was two chips rather than one larger chip.

It turns out computer chips haven't been made larger than about a centimeter on a side. In general, integration sizes are such that you can put 10 million transistors on a single chip a centimeter on a side. Ten or twelve years ago they couldn't get that packing density, and when they wanted, I think it was the Pentium 6 or whatever they called it, they actually had two chips, because they couldn't go larger than about a centimeter. If they tried a centimeter and a half on a side, they couldn't bond it to its substrate without it cracking, just from about a 30° centigrade temperature differential that you have to carry the heat out of the chip. If you don't have a temperature gradient, Fourier doesn't help you get the heat out — Fourier says the heat won't flow out if you don't have a temperature gradient. So with a 30 or 40° gradient on the dimensions you have, you really can't bond chips more than about 1 cm on a side. That's just dissimilar-metal bonding.

Tom's tour of an Intel plant in the late 1990s revealed that the Pentium 2, unlike the original Pentium, was actually two chips in one package — because you couldn't join something more than about one centimeter on a side without cracking from residual soldering/brazing stresses. Used to undermine the redefinition of Moore's Law.

Somewhere here between the Pentium and the Pentium 2, what they didn't really tell you is that the Pentium was all on one chip, but the Pentium 2 was actually two chips in one package — because you couldn't join something more than about one centimeter on a side without it cracking due to residual stresses from the soldering or brazing operation. I remember walking through the plant and they showed me that back in the late '90s. So here's the Pentium 2 in the mid to late '90s.

Tom's recollection from Intel ~2000 during Pentium 3 design. Pentium 3 = two adjacent chips; Pentium 4 reintegrated. Beginning with Pentium 2, redundancy designed in so a partially-bad die can be downgraded rather than scrapped (e.g., Pentium 4 → Pentium 3).

Getting back to your question, starting around Pentium 2 — I was at Intel when they were designing the Pentium 3, probably around 2000. We'll get into this — the Pentium 3 was a chip this size plus another chip right next to it at about half its size. With Pentium 4 they were able to put everything back into one chip as they miniaturized it. As they got to that density, they started making things redundant, so that when you made the chip, if this area didn't work, you could reprogram it so you had some redundancy on the chip. Otherwise you're trying to make 10 million transistors and every one of them has to be good, and one being bad means you haven't got a good chip. This isn't a bonding problem, this is just a processing problem. Now they're up above half a billion transistors on a thing. About 10 years ago the big thing was to design redundancy in, so you can have something that — it's like, I've got two kidneys, right? If one dies, I can still live. They design chips like that.

Physical artifact: 1995-era Pentium chip used to introduce TAB soldering for the soldering module. ## Figures referenced (not cases)

[Tom holds up a section of plastic gas pipe.] Gas pipe. They used to make gas pipes that run down the street out of steel, but after 30 or 40 years it corrodes, and now you have gas leaking into the soil and you can blow up homes. Now they use plastic — they weld plastic pipe. Very corrosion resistant, but it also fails, and it can fail at the joints. We'll talk about some of that. We talked last time about soldering. I passed around one of the semiconductor chips — an old Pentium chip from the 1995 era. I don't have the newest ones, but the technology of tape-automated bonding is the same, and we'll go through the soldering techniques.