Our waiting was finally over last Monday, July 13. The multi-chip module (MCM) packages each containing four copies of our Glasswing chip arrived in our lab. The hour of truth was upon us: The chip we had painstakingly designed over a period of 5 months and for which we had prepared the grounds for another 3 months while it was being manufactured at TSMC in Taiwan was in our lab. It is a state-of-the art piece of IP: it uses our proprietary Chord Signaling technique to transmit 125 Gbps over 6 correlated wires over 12-24 mm of trace in an MCM. The entire “FedEx”ing of the bits from their reception until their delivery on the other end uses about 1 pico-jule per bit.
That’s right. This is 0.000000000001 jule per bit! In more familiar terms, when it is running at full speed, it consumes about 125 mW of power. To have a comparison: when you access the RAM on your laptop or desktop running a standard DDR3 link, you use 20 to 30 times more energy per delivered bit. Just to make sure that the point is not lost: DDR3 links can use 20 to 30 TIMES more power.
Such an astonishingly low power consumption was the target when we designed and implemented the IP (along with many more technical design targets and specifications which I am not going to bore you with). So, now that the chip was in the lab, would it deliver?
A chip this complex and running at such high speed usually needs time and a lot of TLC to turn on. A lot of parameters have to be adjusted, and parts have to be turned on one-by-one to make sure they work together in unison. We had already prepared for this, and had assigned night and day lab shifts to our test engineers to be able to have the chip turned on within a few weeks.
After a few tests for shorts on the chip, we turned it on to see how it worked. Typically at this stage everything is broken. The incoming bits are squashed by the force of the tremendous speed at which they are transmitted. Test engineers would have to work with the design engineers to find the right parameters for the bits to enter our “wormhole” and leave it intact on the other end. Imagine our surprise: everything was working right out of the box! No special setup was required. The bits were delivered intact, at blazing speed, and at an exceedingly low power. It is very hard to provoke errors during the transmission. Left to its own devices, it doesn’t mis-deliver any bit. It was designed for an error rate of 10 to the negative 15 so it needs to run for days before a single error can be observed, if at all. Our longest run so far has been for 70 hours straight, equivalent to transmitting almost 4 petabytes of data, or about 20 times the amount of information contained in the U.S. Library of Congress, and we have not observed a single error.
I am kind of new to the hardware world, but everyone I talk to finds it staggering that such a complex piece of IP simply turns on and works. It is a witness to the care with which Kandou’s methodical engineers designed and implemented the chip, leaving nothing to chance. I am immensely proud of what they have achieved with so few resources.
So, what can the IP be used for? It was designed for different 2.5D integration applications:
- Slicing large SoC’s to reduce time to market, increase yield, and reduce power consumption
- Constructing a new breed of in-package memory systems with extra high bandwidth, low manufacturing cost, and low power
- Connecting SoC’s to outward SerDes
- Connecting heterogenous die in a package to reduce time to market, development cost, and manufacturing cost
- Interface to Optics Engine
- And many more applications.
Moreover, this IP is a first step in our development program to create DIMM’s that offer a huge amount of memory and very large bandwidth (512 GB of memory at 4 Tbps bandwidth – yes, that is Terabit per second!).
This was a proud week for our young company. Our engineering team delivered in the truest meaning of the word. While it is gratifying to see the mathematical ideas behind our unique signaling method work in hardware, delivering bits with such high reliability and at such low power, it was the dedication, hard work, and creativity of our engineering team that made it work in the end.