Making AI Silicon Smart with PVT Monitoring

PVT – depending on what field you are in those three letters may mean totally different things. In my undergraduate field of study, chemistry, PVT meant Pressure, Volume & Temperature. Many of you probably remember PV=nRT, the dreaded ideal gas law. However, anybody working in semiconductors knows that PVT stands for Process, Voltage and Temperature. Well, at least the T still stands for temperature. One out of three isn’t bad.

Chip operation is completely dependent on these three parameters. It would be hard enough to make some of the most advanced chip produced today if PVT were constant across the chip or across time. However, no such luck for chip designers, because each of these varies from chip to chip and across each chip as well. So, guard-bands, margins and binning were all created to deal with this reality.

All of the above techniques still leave a lot of performance, yield and reliability on the table. With the rapid growth of AI and the demand for dedicated silicon to address this market, the need to manage all three of P, V, T has grown too. AI chips are on leading nodes, very large and often have low power or high reliability requirements – such as ISO 26262 when used in automotive systems. Well, drawing on the chemistry reference above, wouldn’t it be nice if there was a Maxwell’s Daemon for regulating chip performance at a microscopic level?

While Maxwell’s Daemon is an imaginary inspector of molecules that controls an imaginary gate to seemingly impossibly controvert the laws of thermodynamics, there are practical solutions to monitoring and appropriately controlling fine grain circuit operation. Think of it as intelligence for artificial intelligence silicon. This is what Moortec’s PVT monitoring IP does across state of the art AI chips.

With AI chips there are often hundreds or more processing elements working in parallel, needing to be fully synchronized. Processing elements that are running too fast are wasting power. Ones that are running too slow, hold up the entire chip. Moortec PVT monitoring blocks can be dispersed throughout a design to help tune operation to ensure that the timing on all elements is similar.

Another area where PVT monitoring can help ensure reliable chip operation is to detect and compensate for aging. As chips age their performance characteristics change. PVT monitors can assess the actual performance of silicon in real time and permit the adjustment of operating conditions to maintain proper operation.

Lastly, due to the possibility of thermal runaway, PVT monitoring can monitor chip operations, and help detect a hot spot before thermal runaway can begin. The real-time element is important because joule heating is load dependent and can change rapidly and may be localized. In the worst case if adjustment is not adequate, device shutdown might be necessary. PVT monitoring allows silicon operation with the highest performance, with safeguards in place to ensure operation within safe operating range.

Moortec has IP for voltage, process and temperature sensing that are connected to create an on chip network or fabric for monitoring and managing performance. Their sophisticated PVT controller supports multiple monitor instances, statistics gathering as well as other compelling features. Moortec offer these IPs on TSMC 40nm, 28nm, 16nm, 12nm and 7nm and have a webpage dedicated to AI, among their other application areas.