4 ms·

Are the physics and manufacturing of analog vs digital components really so different that there's no applicable knowledge transfer? I would be surprised to lea

Are the physics and manufacturing of analog vs digital components really so different that there's no applicable knowledge transfer? I would be surprised to learn that.

jjoonathan 4y ago
Well, there's more transfer than nothing, but the hard parts of digital (sophisticated logic, absurd per-gate reliability, low feature sizes) just don't overlap much with the hard parts of analog, which is all about device physics and materials. Sure, the people who design logic transistors would probably adapt quickly to designing power transistors, but they are a very small part of a very large machine. For example: a digital organization may have learned through harsh and repeated lessons 100 subtle ways to be more conservative about device changes. FinFETs took 15 years to go from R&D to prod, for example. While this level of risk aversion might be critical for shipping billion-transistor devices, it may be maladaptive for shipping single transistors, where a 15 year deployment leaves you 10 years behind your competition. Likewise, the processes (patterning, etching, implantation, etc) are largely similar, but the digital side has made colossal investments optimizing for things that just don't matter as much for analog design. The famed EUV machines? What does a power transistor gain from tiny features? I'd be surprised if they were useful, let alone useful enough to justify the cost, let alone useful enough to justify the cost and the risk of repurposing a machine for process innovation. Like using a F1 car to commute to work -- even if cost were no object, extreme optimization for one purpose makes it worse at others.
- R0b0t1 4y ago
  Power electronics gain surface area from small feature sizes, allowing the electrons to diffuse more efficiently seen as lower loss and higher switching frequencies. Same with gate design.
PhaseLockk 4y ago
It's not that there is zero applicable knowledge transfer, but there are definitely some fundamental differences between digital components and analog power devices, in particular. For digital, the transistor are CMOS, optimized to be as tiny as possible, running at < 1V, and with individual devices mostly driving microamps, going up to milliamps in some particular areas of the chip. For power, we're talking about IGBTs, optimized to handle as much power as possible with maximum efficiency, at hundreds of volts. They are both semiconductor devices, but they are about as similar as the engine for a sports car vs. for a panamax cargo ship.
akiselev 4y ago
The feature size of power electronics is fundamentally limited by the dielectric breakdown voltage of materials. Under a high enough voltage potential, even the best insulator [1] will "break down" and become a conductor The dielectric breakdown voltage of SiO used on wafers is about 2.7x10^9 V/m (2.7 V/nm) so a 600V part that relies on SiO for insulation has to have a minimum gap of at least 222 nm [2] between any conductors that carry that voltage. Cutting edges nodes manufacturing CPUs and stuff generally deal with single digit voltages so while a lot of knowledge can carry over, the meat of the designs are very different. [1] Even a perfect vacuum would breakdown due to the Schwinger effect, but at an astronomical voltage potential like 10^18 V/m. [2] It's actually a lot more complicated at that scale, but it's a useful ballpark value.

Are the physics and manufacturing of analog vs digital components really so different that there's no applicable knowledge transfer? I would be surprised to learn that.

jjoonathan 4y ago

Well, there's more transfer than nothing, but the hard parts of digital (sophisticated logic, absurd per-gate reliability, low feature sizes) just don't overlap much with the hard parts of analog, which is all about device physics and materials. Sure, the people who design logic transistors would probably adapt quickly to designing power transistors, but they are a very small part of a very large machine. For example: a digital organization may have learned through harsh and repeated lessons 100 subtle ways to be more conservative about device changes. FinFETs took 15 years to go from R&D to prod, for example. While this level of risk aversion might be critical for shipping billion-transistor devices, it may be maladaptive for shipping single transistors, where a 15 year deployment leaves you 10 years behind your competition. Likewise, the processes (patterning, etching, implantation, etc) are largely similar, but the digital side has made colossal investments optimizing for things that just don't matter as much for analog design. The famed EUV machines? What does a power transistor gain from tiny features? I'd be surprised if they were useful, let alone useful enough to justify the cost, let alone useful enough to justify the cost and the risk of repurposing a machine for process innovation. Like using a F1 car to commute to work -- even if cost were no object, extreme optimization for one purpose makes it worse at others.

R0b0t1 4y ago
Power electronics gain surface area from small feature sizes, allowing the electrons to diffuse more efficiently seen as lower loss and higher switching frequencies. Same with gate design.

PhaseLockk 4y ago

It's not that there is zero applicable knowledge transfer, but there are definitely some fundamental differences between digital components and analog power devices, in particular. For digital, the transistor are CMOS, optimized to be as tiny as possible, running at < 1V, and with individual devices mostly driving microamps, going up to milliamps in some particular areas of the chip. For power, we're talking about IGBTs, optimized to handle as much power as possible with maximum efficiency, at hundreds of volts. They are both semiconductor devices, but they are about as similar as the engine for a sports car vs. for a panamax cargo ship.

akiselev 4y ago

The feature size of power electronics is fundamentally limited by the dielectric breakdown voltage of materials. Under a high enough voltage potential, even the best insulator [1] will "break down" and become a conductor The dielectric breakdown voltage of SiO used on wafers is about 2.7x10^9 V/m (2.7 V/nm) so a 600V part that relies on SiO for insulation has to have a minimum gap of at least 222 nm [2] between any conductors that carry that voltage. Cutting edges nodes manufacturing CPUs and stuff generally deal with single digit voltages so while a lot of knowledge can carry over, the meat of the designs are very different. [1] Even a perfect vacuum would breakdown due to the Schwinger effect, but at an astronomical voltage potential like 10^18 V/m. [2] It's actually a lot more complicated at that scale, but it's a useful ballpark value.