Comment by mleonhard | Hacker News Reader

Comment by mleonhard | Hacker News Reader

This is important work. Communications bandwidth limits the efficiency of applications on multi-core computers. On-chip lasers can be used for core-core and core-RAM optical interconnects. Such optical interconnects have advantages over electrical interconnects:

- Optical lines can run tangent to the chip surface. This frees up precious chip space and layers. It also changes processor geometry from 2D to 3D.

- Longer electrical interconnects use more power than short ones. This means that electrical interconnect power consumption is more than linear in the number of cores. Optical lines use nearly the same power regardless of length. This means that optical interconnect power consumption is linear in the number of cores.

- Light travels twice as fast as electrical signals in wires. Therefore, optical lines have half the latency of electrical ones. Latency is especially important with large processors because the distance between any two cores is larger.

Taken together, these mean that we should eventually be able to use optical interconnects to create processors with 1,000 times more cores (than processors with electrical interconnects) and fixed power consumption per core regardless of processor size.

The state-of-the-art in massively-multi-core processors:

- Cerebras WSE-2 chip [0] has 850,000 cores and uses 15,000 W of power [1] to achieve 220 PB/s of inter-chip bandwidth.

- Esperanto ET-SoC-1 chip has 1,088 cores and uses 20 W of power [0]. I could not find out its interconnect bandwidth.

[0] https://cerebras.net/chip/

[1] https://www.tomshardware.com/news/cerebras-wafer-scale-engin...

[2] https://www.esperanto.ai

[3] https://www.esperanto.ai/wp-content/uploads/2021/08/Esperant...

Very interesting points. Thanks. Maybe in the future we will have high bandwidth chip to chip standards (like PCIe, and network chip to chip standards like XFI) with optical communication for better latency, especially in large switches with latency constraints.

There are several misconceptions (as pointed by others) relating the trade-off of electrical vs optical for interconnects. Optical interconnect are promising, and in fact, at some point they will become almost unavoidable: bandwidth vs path length. If you want to dive into the fundamental reasons, the following paper might help: https://www.researchgate.net/publication/224504016_Miller_D_...

Thank you for sharing that article.

Miller p.4, says:

> Typical results have energies per bit of 2 – 30 pJ/bit in recent demonstrations. (1 pJ/bit is the same as 1 mW/(Gb/s); the latter is a more common way of stating the unit in the electrical interconnect literature, though the former relates more obviously to the physics of the interconnect devices.) The best current results for transceivers are ~ 2.8 – 6.5 pJ/bit for board or backplane interconnects [17], and ~ 2 pJ/bit [16] for moderate length chip-to-chip interconnects with a relatively ideal electrical channel.

Miller p.12:

> Hence, from the above discussion of systems energy targets, we have optical device energy targets of 10 – 20 fJ/bit for off-chip interconnects, and ~ 2 – 10 fJ/bit for on-chip interconnects.

Kiani's on-chip laser begins operating at 16mW of power. If the laser is part of a chip running at 5GHz, switching the laser at the same clock yields 5Gbit/s. This is 3.2 mW/(Gb/s) = 3.2 pJ/bit. This is still 3 orders of magnitude more than the 10 fJ/bit target.

Light does not, in fact, travel faster in glass than electrical signals in wires. It is a bit slower, instead. Even in vacuum or air, light is only half-again as fast, not twice as fast.

The appeal of optical interconnect is not speed, which is less, but isolation and noise resistance.

Good point about isolation and noise.

I got the propagation numbers from Wikipedia:

> the refractive index of glass is typically around 1.5, meaning that light in glass travels at c / 1.5 ≈ 200000 km/s [0]

> the refractive index of air for visible light is about 1.0003 [0]

> Wires have an approximate propagation delay of 1 ns for every 6 inches (15 cm) of length. [1]

The speed of light in air is c/1.0003, or 299,700 km/s. In glass, it is c/1.5, or 200,000 km/s. If Wikipedia is correct, wire propagation delay is about 15 cm/ns, or 150,000 km/s, which is still slower than in glass. If that number is off, would you please update the Wikipedia article with the correct number? It would be good to include a better reference than the university textbook it currently cites.

Also, I don't think we should assume that a future processor optical mesh will carry signals in glass. I expect it will use micro-randomized mirrors and protect the optics with inert gas. The randomized mirrors will form random optical lines between cores, on the same chip or between chips. The lines will spontaneously reconfigure due to mechanical shock or changes in temperature. The OS will detect such changes and automatically reconfigure the software to optimally utilize the current connections. This should allow hot-swapping CPU cores, and upgrading or downgrading processors on live computers.

There is a scene in the film 2001 Space Odyssey where (MOVIE SPOILER) the protagonist removes pieces of a computer and the AI running on it becomes slower and slower. Such a thing may actually happen.

[0] https://en.wikipedia.org/wiki/Speed_of_light

[1] https://en.wikipedia.org/wiki/Signal_propagation_delay#Elect...

Propagation delay of electrical signals depends on properties of the insulation. In coaxial cable, the insulation is foam, mostly air, so faster than light in glass. On a chip, the wire often has glass and silicon on one side, organic passivation goo on the other. So, propagation there will be slower than in coaxial cable, and could even be a little slower than in fiber.

Optical interconnect via straight-line paths through inert gas, steered by mirrors, could indeed be strictly faster than through fiber. At the same time, propagation across a wafer cannot take less than a nanosecond, at best, so 1.0ns, 1.5ns. The total delay will be a product of message marshalling and routing, typically dozens or more times that.

Real-time adaptation may be necessary for managing flaky quantum bits, but the more of it that happens, the less total speed you get.

Reply on news.ycombinator.com