The death of Moore's Law has been punted back and forth between engineers and pundits any number of times at this point. And as silicon-based transistors become smaller and smaller, manufacturers have had to grapple with increased temperature densities (more transistors in a smaller area, generating more heat), not to mention other issues that naturally arise from closely packing smaller, faster transistors.
And even as chiplet technologies such as TSMC's InFO_LI and Intel's Foveros 3D technology have enabled increased functionality and the ability to pair multiple chips in the same substrate, connecting those chips to each other still requires electrical wires carrying electrons around. Flying electrons means both increased temperatures (from traveling through semiconductor's resistance) and increased power consumption. As covered by The Register, startup Lightmatter has another idea: connect chips without electrical wiring altogether. The company took to HotChips with its alternative: photonics.
"Arrays of electrically interconnected chiplets suffer fundamentally from issues, including concatenating power consumption,” Nicholas Harris, founder and CEO of Lightmatter said in the company's HotChips presentation.
The issue is clear and has already been well-identified: The more chiplets connected in a single package, the more interconnections those chiplets must have with each other in order to trade the data required for computation. While electricity is a fast medium, it's not the fastest available--that prize is reserved for light. The company's Passage technology thus aims to bring photonics to the chiplet era, by allowing different chips to be interconnected through nano-photonic wave guides. These essentially use photons (instead of the more ubiquitous electrons) to ferry information, with extremely low signal loss and much increased bandwidth.
Interestingly, AMD itself has been exploring photonics designs that could allow for information transfer for its architectures as well. For its part, Intel has a whole research center dedicated to it.
“Passage is diced from a 300mm Silicon Photonics wafer that includes lasers, optical modulators, photo detectors, and transistors all side-by-side integrated in the platform,” Harris continued.
Chiplets to be interconnected (such as ASICs, CPUs, GPUs or memory chips) are then laid on top of this photonics-powered 'sandwich.'
“Because Passage has integrated lasers and transistors, the co-packaged chips don’t have to deal with any of the complexity of the transmit, receive, or circuit switching photonics elements,” Harris said. “Each Passage tile can house an array of heterogeneous chips. For example, a tile might contain two different types of ASICs and maybe two HBM stacks.”
The company claims its approach brings sub-2 nanosecond hop times between the information's exit and entry point, irrespective of distance between points (so the farthest chiplets will communicate as fast as the closest ones). The nano-photonic waveguides used by Lightmatter have advantages over traditional fiber optic interconnects in that they're much smaller: The company says it can fit as many as 40 waveguides in the space of a single optical fiber.
According to Lightmatter, this allows them to provide 96 TBps of bandwidth to each die. Compare that to AMD's Infinity Fabric maximum theoretical bandwidth of 800 Gbps. Off-chip communication--from Passage through to other systems via fiber arrays--peaks at around 16 TBps.
Furthermore, since Passage is a fully customizable interconnect-on-a-package, manufacturers no longer have to design their own interconnect designs (such as AMD's Infinity Fabric or Intel's EMIB). They can simply drop their devices into a photonics-powered Passage that can accommodate up to 48 full-reticle chips (full reticle meaning that these chips can occupy as much area as manufacturing processes allow them to), and provides an already-existing interconnect between them.
Despite this being photonics through and through, chips planted on Passage will be of the more traditional, silicon-based transistor type, which means they still require electrical communication. This is also supported by using Through Silicon Vias (TSVs), which also deliver power to the dies and support the PCIe and CXL standards.
Moore's Law isn't dead, but only because of chip designer's ingenuity. Lightmatter's approach is just another in a series of steps that aim to sustain computing's acceleration today and tomorrow. The only question is, how willing will major chip companies be to adopt Lightmatter's tech when they are also spending large amounts of money and engineering resources to develop their own?
