[HN Gopher] TSMC bets on unorthodox optical tech
       ___________________________________________________________________
        
       TSMC bets on unorthodox optical tech
        
       Author : Rohitcss
       Score  : 224 points
       Date   : 2025-05-26 17:15 UTC (1 days ago)
        
 (HTM) web link (spectrum.ieee.org)
 (TXT) w3m dump (spectrum.ieee.org)
        
       | Liftyee wrote:
       | As I understand it (from designing high-speed electronics), the
       | major limitations to data/clock rates in copper are signal
       | integrity issues. Unwanted electromagnetic interactions all
       | degrade your signal. Optics is definitely a way around this, but
       | I wonder if/when it will ever hit similar limits.
        
         | moelf wrote:
         | luckily photons are boson (if we ever pushes things to this
         | level of extreme)
        
           | Taek wrote:
           | This comment appears insightful but I have no idea what it
           | means. Can someone elaborate?
        
             | scheme271 wrote:
             | Electrons are fermions which means that two electrons can't
             | occupy the same quantum state (Pauli exclusion principle).
             | Bosons don't have the limit so I believe that implies that
             | you can have stronger signals at the low end since you can
             | have multiple photons conveying or storing the same
             | information.
        
               | xeonmc wrote:
               | Also less chance for external interference.
        
             | cycomanic wrote:
             | What the previous poster is implying is that electrons
             | interact much more strongly than photons. Hence electrons
             | are very good for processing (e.g. building a transistor),
             | while photons are very good for information transfer. This
             | is also a reason why much of the traditional "optical
             | computer" research was fundamentally flawed, just from
             | first principles one could estimate that power requirements
             | are prohibitive.
        
               | nullc wrote:
               | > This is also a reason why much of the traditional
               | "optical computer" research was fundamentally flawed
               | 
               | presumably also because photons at wavelengths we can
               | work with are BIG
        
             | xeonmc wrote:
             | Fermions can "hit each other" whereas bosons "pass through
             | each other".
             | 
             | (Strong emphasis on the looseness of the scare quotes.)
        
         | EA-3167 wrote:
         | The energy densities required for photon-photon interactions
         | are so far beyond anything we need to worry about that it's a
         | non-issue. Photons also aren't going to just ignore local
         | potential barriers and tunnel at the energy levels and scales
         | involved in foreseeable chip designs either.
        
           | wyager wrote:
           | P-p interactions are not an issue but we do have high enough
           | field intensities in high bandwidth fibers to run into p-f
           | nonlinearity issues.
        
         | notepad0x90 wrote:
         | isn't attenuation also an issue with copper? maybe with small
         | electronics it is negligible given the right amps? in other
         | words, with with no interference, electrons will face impedance
         | and start losing information.
        
           | to11mtm wrote:
           | Attenuation is going to be an issue for any signal, but in my
           | experience Fiber can go for many miles without a repeater
           | whereas something like Coax you're going one to two orders of
           | magnitude less. [0]
           | 
           | [0] - Mind you, some of that for Coax is due to other issues
           | around CTB and/or the challenge that in Coax, you've got many
           | frequencies running through alongside each frequency having
           | different attenuation per 100 foot...
        
             | spwa4 wrote:
             | > Coax is due to other issues around CTB and/or the
             | challenge that in Coax, you've got many frequencies running
             | through alongside each frequency having different
             | attenuation per 100 foot
             | 
             | Actually this is true for fibers as well. In DWDM (all
             | internet links are DWDM, including fiber-to-the-home in
             | most places) you have many frequencies running alongside
             | and each frequency has differences in attenuation (though
             | generally measured per kilometer, not 100 foot)
             | 
             | Optical light are standing electromagnetic waves. Which
             | means they don't disrupt each other. Electrical signals
             | aren't standing waves. They affect each other.
             | 
             | The difference can be put like this: how many X (electrical
             | waves, but essentially everything, protons, ...) fit on the
             | tip of a needle? (or in a cable)
             | 
             | 1) electrical waves? Some finite number. Can be large of
             | course, but ...
             | 
             | 2) photons (ie. fiber signals)? ALL OF THEM. Literally
             | every photon that exists in the entire universe would
             | happily join every other photon on the tip of a needle
             | nothing would interfere with anything else
        
           | bgnn wrote:
           | This is the main mechanism of interference anyhow, called
           | inter-symbol-interference.
        
         | lo0dot0 wrote:
         | Optics also have signal integrity issues. In practice OSNR and
         | SNR limit optics. Cutting the fiber still breaks it. Small
         | vibrations also affect the signal's phase.
        
           | cycomanic wrote:
           | Phase variations will not introduce any issues here, they
           | most certainly are talking about intensity modulation. You
           | can't really (easily) do coherent modulation using incoherent
           | light sources like leds.
           | 
           | SNR is obviously an issue for any communication system,
           | however fiber attenuation is orders of magnitude lower than
           | coax.
           | 
           | The bigger issues in this case would be mode-dispersion,
           | considering that they are going through "imaging" fibres,
           | i.e. different spatial components of the light walking off to
           | each other causing temporal spread of the pulses until they
           | overlap and you can't distinguish 1's and 0's.
        
             | abdullahkhalids wrote:
             | Mode dispersion is frequency dependent phase changes.
        
               | cycomanic wrote:
               | That's chromatic dispersion, mode dispersion is spatial
               | "path" dependent phase changes. Vibration is actually
               | somewhat more relevant because if it wasn't for that we
               | could theoretically undo mode dispersion (we would need
               | phase information though).
               | 
               | That said all of that is irrelevant to what the previous
               | speaker said, vibration induced phase variation as an
               | impairment. Thats just not an issue, vibrations are way
               | too slow to impair optical comms signals.
        
               | mycall wrote:
               | How do the gravity wave optical paths solve the vibration
               | issues? Couldn't TSMC do something similar?
        
           | BardiaPezeshki wrote:
           | aha! that is true with lasers that are coherent. But not with
           | LEDs. We don't care about modes, polarization, or phase.
           | Also, no worry about feedback into the lasers, so no
           | isolators. LEDs are way easier!
        
         | wyager wrote:
         | We already regularly run into optical nonlinearity issues in
         | submarine cables. The instantaneous EM fields generated in high
         | bandwidth fiber are sufficiently strong to cause nonlinear
         | interactions with the fiber medium that we have to correct for.
        
           | Salgat wrote:
           | I don't believe this is a factor for the distances that
           | inter-chip transmission has. From what I can find, this is at
           | most an issue for communication spanning tens to hundreds of
           | meters in a datacenter.
        
       | m3kw9 wrote:
       | If each cable is 10gb/s and uses 1 pixel to convert into
       | electrical signals, would that mean they need a 10 giga frame per
       | second sensor?
        
         | cpldcpu wrote:
         | I think that's just a simplifying example. They would most
         | likely not use an image sensor, but a photodetector with a
         | broadband amplifier.
        
         | lo0dot0 wrote:
         | No, not necessarily. If you can distinguish different amplitude
         | levels you can do better. For example four amplitude modulation
         | (4AM) carries two bits per symbol. There is also the option to
         | use coherent optics, which can detect phase, and carry
         | additional information in the phase.
        
       | amelius wrote:
       | > The transmitter acts like a miniature display screen and the
       | detector like a camera.
       | 
       | So if I'm streaming a movie, it could be that the video is
       | actually literally visible inside the datacenter?
        
         | tehjoker wrote:
         | maybe a low res bit packed binary motion picture thats
         | uncompressed
        
         | tails4e wrote:
         | No, this is just an a analogy. The reality is the data is
         | heavily modulated, and also the video is encoded so at no point
         | would something that visually looks like an image be visible in
         | the fibre.
        
         | lo0dot0 wrote:
         | Obviously this is not how video compression and packets work
         | but for the sake of the argument consider the following. The
         | article speaks of a 300 fiber cable. A one bit per pixel square
         | image with approx. 300 pixels is 17x17 in size. Not your
         | typical video resolution.
        
           | fecal_henge wrote:
           | Not your typical frame rate either.
        
         | mystified5016 wrote:
         | Not any more than if you blinked an LED each time a bit comes
         | across your network connection.
        
         | topspin wrote:
         | They story states there are 300 optical lines in the
         | "fiberbundle." Let's assume this is arranged as 20x15 and the
         | wavelength of the led is visible and bright enough to perceive.
         | So if your unencoded, monochrome 20x15 movie was aligned on
         | every frame and rendered at 10E9 FPS, then yes, your movie
         | would be visible at the end of one of these cables through a
         | magnifying glass.
        
       | speedbird wrote:
       | Headline seems misleading. They're buildings detectors for
       | someone, not 'betting' on it.
        
         | ggm wrote:
         | The word is mostly used for finance investment promotion. I'd
         | guess that it's written to attract those bloggers/casters, and
         | make money for somebody.
        
         | walthamstow wrote:
         | The 'bet' is investing time and money into something that may
         | not yield results. It's pretty common business language.
        
       | ls612 wrote:
       | Forgive the noob question but what stops us from making optical
       | transistors?
        
         | qwezxcrty wrote:
         | I think the most fundamental reason is that there is no
         | efficient enough nonlinearity at optical frequencies. So two
         | beams(or frequencies in some implementation) tends not to
         | affect each other in common materials, unless you have a very
         | strong source (>1 W) so the current demonstrations for all-
         | optical switching are mostly using pulsed sources.
        
           | ls612 wrote:
           | I wonder if considerably more engineering and research effort
           | will be applied here when we reach the limit of what silicon
           | and electrons can do.
        
             | cycomanic wrote:
             | No this is not an engineering issue, it's a problem of
             | fundamental physics. Photons don't interact easily. That
             | doesn't mean there are not specialised applications where
             | optical processing can make sense, e.g. a matrix
             | multiplication is really just a more complex lens so it's
             | become very popular to make ML accelerators based on this.
        
             | momoschili wrote:
             | Contrary to the prior commenter, there is definitely
             | significant engineering going toward this, but it's not
             | clear or likely that photonic computing will supplant
             | electronic computing (at least not anytime soon), but
             | rather most seem to think of it as an accelerator for
             | highly parallel tasks. Two major ways people are thinking
             | of achieving this are using lithium niobate devices which
             | mediate nonlinear optical effects via light-matter
             | interaction, and silicon photonic devices with electrically
             | tunable elements. In the past there was a lot of work with
             | III-V semiconductors (GaAs/InAs/GaN/AlN etc) but that seems
             | to have leveled off in favor of lithium niobate.
             | 
             | Photonics has definitely proved itself in communications
             | and linear computing, but still has a way to in terms of
             | general (nonlinear) compute.
        
               | ls612 wrote:
               | Yeah that was sorta what I was thinking is there a clever
               | way to exploit light interacting with matter to turn the
               | gate on/off.
        
           | perlgeek wrote:
           | As somebody who tried to do a PHD in optical communications,
           | this is 100% correct.
           | 
           | I wonder if meta material might provide such nonlinearities
           | in the future.
        
         | elorant wrote:
         | Photons are way more difficult to control than electrons. They
         | don't interact with each other.
        
         | dartharva wrote:
         | Electric signals can be made to operate and switch currents on-
         | demand in CMOS transistors to make logic gates and eventually
         | CPUs. Light signals, however, operate linearly - photons just
         | pass through each other and don't interact.
        
       | qwezxcrty wrote:
       | Not an expert in communications. Would the SerDes be the new
       | bottleneck in the approach? I imagine there is a reason for
       | serial interfaces dominating over the parallel ones, maybe timing
       | skew between lanes, how can this be addressed in this massive
       | parallel optical parallel interface?
        
         | to11mtm wrote:
         | > timing skew between lanes
         | 
         | That's a big part of it. I remember in the Early Pentium 4
         | days, starting to see a lot more visible 'squiggles' on PCB
         | traces on motherboards; the squiggles essentially being a case
         | of 'these lines need more length to be about as long as the
         | other lines and not skew timing'
         | 
         | In the case of what the article is describing, I'm imagining a
         | sort of 'harness cable' that has a connector on each end for
         | all the fibers, and the fibers in the cable itself are all the
         | same length, there wouldn't be a skew timing issue. (Instead,
         | you worry about bend radius limitations.)
         | 
         | > Would the SerDes be the new bottleneck in the approach
         | 
         | I'd think yes, but at the same time in my head I can't really
         | decide whether it's a harder problem than normal mux/demux.
        
         | fecal_henge wrote:
         | SerDes is already frequently parallelised. The difference is
         | you never expect the edges or even the entire bits to arrive at
         | the same time. You design your systems to recover timing per
         | link so the skew doesnt become the constraint on the line rate.
        
           | bgnn wrote:
           | one can implement SerDes at any point of the electro-optical
           | boundary. For example, if we have 1 Tbps incoming NRZ data
           | from the fiber, and the CMOS technology at hand only allows
           | 10 GHz clock speed for the slicers, one can have 100x
           | receivers (photodiode, TIA, slicer), or 1x photodiode, 100x
           | TIA + slicer, or 1x photodiode + TIA and 100x slicers. The
           | most common id the last one, and it spits out 100x parallel
           | data.
           | 
           | Things get interesting if the losses are high and there needs
           | to be a DFE. This limits speed a lot, but then copper
           | solutions moved to sending multi-bit symbols (PAM 3,
           | 4,5,6,8,16.. ) which can also be done in optical domain. One
           | can even send multiple wavelengths in optical domain, so
           | there are ways to boost the baud rate without requiring high
           | clock frequencies.
        
         | waterheater wrote:
         | >serial interfaces dominating over the parallel ones
         | 
         | Semi-accurate. For example, PCIe remains dominant in computing.
         | PCIe is technically a serial protocol, as new versions of PCIe
         | (7.0 is releasing soon) increase the serial transmission rate.
         | However, PCIe is also parallel-wise scalable based on
         | performance needs through "lanes", where one lane is a total of
         | four wires, arranged as two differential pairs, with one pair
         | for receiving (RX) and one for transmitting (TX).
         | 
         | PCIe scales up to 16 lanes, so a PCIe x16 interface will have
         | 64 wires forming 32 differential pairs. When routing PCIe
         | traces, the length of all differential pairs must be within
         | <100 mils of each other (I believe; it's been about 10 years
         | since I last read the spec). That's to address the "timing skew
         | between lanes" you mention, and DRCs in the PCB design software
         | will ensure the trace length skew requirement is respected.
         | 
         | >how can this be addressed in this massive parallel optical
         | parallel interface?
         | 
         | From a hardware perspective, reserve a few "pixels" of the
         | story's MicroLED transmitter array for link control, not for
         | data transfer. Examples might be a clock or a data frame
         | synchronization signal. From the software side, design a
         | communication protocol which negotiates a stable connection
         | between the endpoints and incorporates checksums.
         | 
         | Abstractly, the serial vs. parallel dynamic shifts as
         | technology advances. Raising clock rates to shove more data
         | down the line faster (serial improvement) works to a point, but
         | you'll eventually hit the limits of your current technology.
         | Still need more bandwidth? Just add more lines to meet your
         | needs (parallel improvement). Eventually the technology
         | improves, and the dynamic continues. A perfect example of that
         | is PCIe.
        
         | cycomanic wrote:
         | They are doing 10 Gb/s over each fibre, to get to 10 Gb/s you
         | have already undergone a parallel -> serial conversion in
         | electronics (clock rates of your asics/fpgas are much lower),
         | to increase the serial rate is in fact the bottleneck. Where
         | the actual optimum serial rate is highly depends on the cost of
         | each transceiver, e.g. long haul optical links operate at up to
         | 1 Tb/s serial rates while datacenter interconnects are 10-25G
         | serial AFAIK.
        
       | cycomanic wrote:
       | That article is really low on details and mixes up a lot of
       | things. It compares microleds to traditional WDM fiber
       | transmission systems with edge emitting DFB lasers and ECLs, but
       | in datacentre interconnects there's plenty of optical links
       | already and they use VCSELs (vertical cavity surface emitting
       | lasers), which are much cheaper to manufacture. People also have
       | been putting these into arrays and coupling to multi-core fiber.
       | The difficulty here is almost always packaging, i.e. coupling the
       | laser. I'm not sure why microleds would be better.
       | 
       | Also transmitting 10 Gb/s with a led seems challenging. The
       | bandwidth of an incoherent led is large, so are they doing
       | significant DSP (which costs money and energy and introduces
       | latency) or are they restricting themselves to very short (10s of
       | m) links?
        
         | tehjoker wrote:
         | short links it's in the article
        
           | cycomanic wrote:
           | Ah I missed the 10m reference there. I'm not sure it makes
           | more sense though. Typical intra-datacenter connections are
           | 10s-100s of meters and use VCSELs, so introducing microleds
           | just for the very short links instead of just parallelising
           | the VCSEL connections (which is being done already)? If they
           | could actually replace the VCSEL I would sort of see the
           | point.
        
             | jauntywundrkind wrote:
             | There's been a constant drum-beat that even intra-rack is
             | trying to make its way to optical as fast as it can, that
             | copper is more and more complex and expensive to scale
             | faster. If we have a relatively affordable short range
             | optical system that doesn't require heavy computational
             | work to do, that sounds like a godsend, like a way to
             | increase bits per joule while reducing expensive cabling
             | cost.
             | 
             | Sure yes, optical might use expensive longer range optical
             | today! But using that framing to assess new technologies &
             | what help the could be may be folly.
        
         | qwezxcrty wrote:
         | I guess they are doing direct modulated IMDD for each link so
         | the DSP burden is not related to the coherence of diodes? Also
         | indeed very short reach in the article.
        
           | cycomanic wrote:
           | The problem with both leds and imaging fibres is that modal
           | dispersion is massive and completely destroys your signal
           | after only a few meters of propagation. So unless you do MMSE
           | (which I assume would be cost prohibitive), you really can
           | only go a few meters. IMDD doesn't really make a difference
           | here.
        
             | adgjlsfhk1 wrote:
             | I think this is intended for short distances (e.g. a few
             | cm). cpu to GPU and network card to network card still will
             | be lasers, the question is whether you can do core to core
             | or CPU to ram with optics
        
               | cycomanic wrote:
               | But why are they talking about multicore fibres then? I
               | would have expected ribbons. You might be right though.
        
               | topspin wrote:
               | > I would have expected ribbons.
               | 
               | The cable is just 2D parallel optical bus. With a bundle
               | like this, you can wrap it with a nice, thick PVC (or
               | whatever) jacket and employ a small, square connector
               | that matches the physical scheme of the 2D planar
               | microled array.
               | 
               | It's a brute force, simple minded approach enabled by
               | high speed, low cost microled arrays. Pretty cool I
               | think.
               | 
               | The ribbon concept could be applicable to PCBs though.
        
               | cycomanic wrote:
               | You might be right and they are talking about fibre
               | bundles, but that that's something different to a
               | multicore fibre (and much larger as well, which could
               | pose significant problems especially if we are talking cm
               | links). What isn't addressed is that leds are quite
               | spatially incoherent and beam divergence is strong, so
               | the fibres they must use are pretty large, coupling via
               | just a connector might not be easy especially if we want
               | to avoid crosstalk.
               | 
               | What I'm getting at is, that I don't see any advantage
               | over vcsel arrays. I'm not convinced that the price point
               | is that different.
        
               | topspin wrote:
               | > You might be right and they are talking about fibre
               | bundles
               | 
               | The caption of the image of the cable and connector
               | reads: "CMOS ASIC with microLEDs sending data with blue
               | light into a _fiberbundle_. " So yes, fibre bundles.
               | 
               | > I don't see any advantage over vcsel arrays
               | 
               | They claim the following advantages:                   1.
               | Low energy use         2. Low "computational overhead"
               | 3. Scalability
               | 
               | All of these at least pass the smell test. LEDs are
               | indeed quite efficient relative to lasers. They cite
               | about an order of magnitude "pJ/bit" advantage for the
               | system over laser based optics, and I presume they're
               | privy to vcsels. When you're trying to wheedle nuclear
               | reactor restarts to run your enormous AI clusters, saving
               | power is nice. The system has a parallel "conductor"
               | design that likely employs high speed parallel CMOS
               | latches, so the "computational overhead" claim could make
               | sense: all you're doing is latching bits to/from PCB
               | traces or IC pins so all the SerDes and multiplexing cost
               | is gone. They claim that it can easily be scaled to more
               | pixels/lines. Sure, I guess: low power makes that easier.
               | 
               | There you are. All pretty simple.
               | 
               | I think there is use case for this outside data centers.
               | We're at the point where copper transmission lines are a
               | real problem for consumers. Fiber can solve the signal
               | integrity problem for such use cases, however--despite
               | several famous runs at it (Thunderbolt, Firewire)--the
               | cost has always precluded widespread adoption outside
               | niche, professional, or high-end applications. Maybe LED
               | based optics can make fiber cost competitive with copper
               | for such applications: one imagines a very small, very
               | low power microLED based transceiver costing only
               | slightly more than a USB connector on each end of such a
               | cable with maybe 4-8 parallel fibers. Just spit-balling
               | here
        
               | nsteel wrote:
               | Aren't they also claiming this is more reliable? I'm told
               | laser reliability is a hurdle for CPO.
               | 
               | And given the talk about this as a CPO alternative, I was
               | assuming this was for back plane and connections of a few
               | metres, not components on the same PCB.
        
               | topspin wrote:
               | > Aren't they also claiming this is more reliable?
               | 
               | Indeed they do. I overlooked that.
               | 
               | I know little about microLED arrays and their
               | reliability, so I won't guess about how credible this is:
               | LED reliability has a lot of factors. The cables involved
               | will probably be less reliable than conventional laser
               | fiber optics due to the much larger number of fibers that
               | have to be precision assembled. Likely to be more fragile
               | as well.
               | 
               | On-site fabricating or repairing such cables likely isn't
               | feasible.
        
               | nsteel wrote:
               | I understand that CPO reliability concerns are
               | specifically with the laser drivers. It's very expensive
               | to replace your whole chip when one fails. Even if the
               | cables are a concern (I've no idea), having more reliable
               | drivers would still be preferable to less reliable
               | cables, given how much cheaper/easier replacing cables
               | would be (up to a point, of course).
        
               | topspin wrote:
               | > I understand that CPO reliability concerns are
               | specifically with the laser drivers.
               | 
               | Yes. I've replaced my share of dead transceivers, and I
               | suspect the laser drivers were the failure mode of most
               | of them.
               | 
               | That doesn't fill in the blank for me though: how
               | reliable are high speed, dense microLEDs?
        
               | nsteel wrote:
               | And are they going to work out any better than Linear
               | Drive Optics, the more obvious alternative?
        
               | topspin wrote:
               | LDO is just integration. It certainly has value:
               | integration almost always does. So it's clearly the
               | obvious _optimization_ of conventional serial optical
               | communication.
               | 
               | This new TSMC work with parallel incoherent optics is
               | altogether distinct. No DSP. No SerDes. Apples and
               | oranges.
        
               | nsteel wrote:
               | Ok, but I'm just after solutions to problems I have
               | talking to other chips. I don't mind what's novel and
               | what's optimisation. Whatever is adopted, in either case
               | it's a step-change from the past 20 years of essentially
               | just copper and regular serdes in this space.
               | 
               | And I'm not sure how much of this is actually TSMC's
               | work, the title is misleading.
               | 
               | Edit: actually, they are working on the detector side.
        
               | BardiaPezeshki wrote:
               | we use borosilicate fibers that are used for illumination
               | applications. You might have seen a bundle in a
               | microscope light for example. And they are incredibly
               | robust compared to single mode fibers. Note the very
               | tight bend angle in the picture - that's a 3mm bend
               | radius. Imagine doing that with a single mode fiber!
        
               | topspin wrote:
               | > And they are incredibly robust
               | 
               | See my other comment about non-datacenter applications.
               | There is a serious opportunity here for fixing signal
               | integrity problems with contemporary high bandwidth
               | peripherals. Copper USB et al. are no good and in
               | desperate need of a better medium.
        
               | BardiaPezeshki wrote:
               | The fiber cables we use are basically 2D arrays of 50um
               | thick fibers that match the LED and detector arrays.
               | We've made connectors and demonstrated very low crosstalk
               | between the fibers. Advantage over VCSELs is much lower
               | power consumption overall, much lower cost (LEDs are dirt
               | cheap and extremely high yield), because we are blue
               | light, the detector arrays are much easier and can be
               | modified camera technology, and most importantly, much
               | better reliability. VCSELs are notorious for bad rel.
        
         | rajnathani wrote:
         | The article is about chip interconnects. Think like replacing
         | PCIe, NVLink, or HBM/DDR RAM buses, with optical communication.
        
         | BardiaPezeshki wrote:
         | In datacenters people use optics for longer distances (10m to
         | 2km). Within a rack it is almost always copper. The reason is
         | that for short distances lasers are too expensive, unreliable,
         | and consume too much power. We think microLED based links might
         | replace copper at short distances (sub 10m). MicroLEDs into
         | relatively thick fiber cores (50um) are much easier to package
         | than standard single mode laser based optics.
         | 
         | on the distance - exactly right. The real bottleneck now in AI
         | clusters is the interconnect within a rack or sub 10m. So that
         | is the market we are addressing.
         | 
         | On your second point - exactly! Normally people think LEDs are
         | slow and suck. That is the real innovation. At Avicena, we've
         | figured out how to make LEDs blink on and off at 10Gb/s. This
         | is really surprising and amazing! So with simple on-off
         | modulation, there is no DSP or excess energy use. The article
         | says TSMC is developing arrays of detectors, based on their
         | camera process, that also receive signals at 10Gb/s. Turns out
         | this is pretty easy for a camera with a small number of pixels
         | (~1000). We use blue light, which is easily absorbed in
         | silicon. BTW, feel free to reach out to Avicena, and happy to
         | answer questions.
        
           | throwaway48476 wrote:
           | So, toslink?
        
         | adrian_b wrote:
         | The article mentions lengths of up to 10 m, so this technology
         | is restricted to links inside a cabinet or between closely
         | located cabinets.
         | 
         | The claimed advantage is a very high aggregate throughput and
         | much less energy per bit than with either copper links or
         | traditional laser-based optical links.
         | 
         | For greater distances, lasers cannot be replaced by anything
         | else.
        
       | albertzeyer wrote:
       | There is also optical neuromorphic computing, as an alternative
       | to electronic neuromorphic computing like memristors. It's an
       | fascinating field, where you use optical signals to perform
       | analog computing. For example:
       | 
       | https://www.nature.com/articles/s41566-020-00754-y
       | 
       | https://www.nature.com/articles/s44172-022-00024-5
       | 
       | As far as I understood, you can only compute quite small neural
       | networks until the noise signal gets too large, and also only a
       | very limited set of computations works well in photonics.
        
         | cycomanic wrote:
         | The issue with optical neuromorphic computing is that the field
         | has been doing the easy part, i.e. the matrix multiplication.
         | We have known for decades that imaging/interference networks
         | can do matrix operations in a massively parallel fashion. The
         | problem is the nonlinear activation function between your
         | layers. People have largely been ignoring this, or just
         | converted back to electrical (now you are limited again by the
         | cost/bandwidth of the electronics).
        
           | seventytwo wrote:
           | Seems hard to imagine there's not some non-linear optical
           | property they could take advantage of
        
             | cycomanic wrote:
             | The problem is intensity/power, as discussed previously
             | photon-photon interactions are weak, so you need very high
             | intensities to get a reasonable nonlinear response. The
             | issue is, that optical matrix operations work by spreading
             | out the light over many parallel paths, i.e. reducing the
             | intensity in each path. There might be some clever ways to
             | overcome this, but so far everyone has avoided that
             | problem. They said we did "optical deep learning" what they
             | really did was an optical matrix multiplication, but saying
             | that would not have resulted in a Nature publication.
        
             | programjames wrote:
             | There is, and people have trained purely optical neural
             | networks:
             | 
             | https://arxiv.org/abs/2208.01623
             | 
             | The _real_ issue is trying to backpropagate those nonlinear
             | optics. You need a second nonlinear optical component that
             | matches the derivative of the first nonlinear optical
             | component. In the paper above, they approximate the
             | derivative by slightly changing the parameters, but that
             | means the training time scales linearly with the number of
             | parameters in each layer.
             | 
             | Note: the authors claim it takes O(sqrt N) time, but
             | they're forgetting that the learning rate mu = o(1/sqrt N)
             | if you want to converge to a minimum:
             | Loss(theta + dtheta) = Loss(theta) + dtheta * dLoss(theta)
             | + O(dtheta^2)                              = Loss(theta) +
             | mu * sqrtN * C (assuming Lipschitz continuous)         ==>
             | min(Loss)    = mu * sqrtN * C/2
        
       | smj-edison wrote:
       | With this design, how do they route enough pins from the chip to
       | the optical transceiver? Would it take chiplets to get enough
       | lanes?
        
       | nsteel wrote:
       | There's a link to their press release in the article, it probably
       | answers some questions here: https://avicena.tech/avicena-
       | announces-modular-lightbundle-o...
        
       | rajnathani wrote:
       | Minor: It would be nice if the company TSMC is collaborating with
       | on this, Avicena, is mentioned in the HN title.
        
         | stingraycharles wrote:
         | It's against HN policy to editorialize the titles
        
           | rajnathani wrote:
           | But it is also against HN rules to keep sensationalistic
           | titles, dang (moderator 1) routinely modifies titles wherever
           | necessary.
        
       | sunray2 wrote:
       | Somewhat related: there's a relatively big push for optical
       | interconnects and integrated optics in quantum computing. Maybe
       | this article yields insight onto what may happen in future.
       | 
       | With quantum computing, one is forced to use lasers. Basically,
       | we can't transmit quantum information with the classical light
       | from LEDs (handwaving-ly: LEDs emit a distribution of possible
       | photon numbers, not single photons, so you lose control at the
       | quantum level). Moreover, we often also need the narrow linewidth
       | of lasers, so that we can interact with atoms in the way we want
       | them to. That is, not to excite unwanted atomic energy levels. So
       | you see in trapped ion quantum computing people tripping over
       | themselves to realise integration of laser optics, through fancy
       | engineering that i don't fully understand like diffraction
       | gratings within the chip that diffract light onto the ions. It's
       | an absolutely crucial challenge to overcome if you want to make
       | trapped ion quantum computers with more than several tens of
       | ions.
       | 
       | Networking multiple computers via said optical interconnects is
       | an alternative, and also similarly difficult.
       | 
       | What insight do i gleam from this IEEE article, then? I believe
       | if this approach with the LEDs works out for this use case, then
       | I'd see it as a partial admission of failure for laser-integrated
       | optics at scale. It is, after all, the claim in the article that
       | integrating lasers is too difficult. And then I'd expect to see
       | quantum computing struggle severely to overcome this problem.
       | It's still research at this stage, so let's see if Nature's cards
       | fall fortuitously.
        
         | avsteele wrote:
         | TSMC's approach here sounds sensible but I don't think it
         | speaks much to QC. It is a pretty different problem domain. The
         | trapped-ion QCs can use much more expensive / less practical
         | lasers and optics and still be useful.
        
         | mmmBacon wrote:
         | Quantum computing is still a technology of the future. When we
         | are still talking about 12 qubits as a breakthrough, there's a
         | long way to go. Optical interconnects are the least of quantum
         | computing's problems.
         | 
         | However, it's not correct to say lasers are unreliable. It's
         | fundamentally false and it's not supported by field data from
         | today's pluggable modules. 10's of millions of lasers are
         | deployed in data centers today in pluggable modules.
         | 
         | It's also useful to remember that an LED is essentially the
         | gain region of a laser without the reflectors. When lasers fail
         | in the field, they fail for the same reasons an LED will fail;
         | moisture or contamination penetration of the semiconductor
         | material.
         | 
         | An LED is not useful for quantum computing. To create a Bell
         | pair (2qubits) you need a coherent light source to create
         | correlated photons. The photons produced by an incoherent light
         | source like an LED are fundamentally uncorrelated.
        
           | ziofill wrote:
           | Actually optical interconnects are the _biggest_ of
           | (photonic) quantum computing problems. If we had good enough
           | optical interconnects (i.e. with low enough optical loss) we
           | would already have a fault-tolerant quantum computer. See
           | https://www.nature.com/articles/s41586-024-08406-9 (also note
           | that Aurora produces 12 physical qubit modes _at each clock
           | cycle_ )
        
         | dgfl wrote:
         | Trapped-ion and neutral-atom QC require lasers because the
         | light signal needs to be coherent. That's the main feature of a
         | classical laser, really. The explanation with the number of
         | photons doesn't really cut it, because even a perfect laser
         | does not have a definite photon number: coherent states are
         | inherently uncertain in both photon number and phase. But LEDs
         | are even worse, because the light signal is truly incoherent.
         | It's not even a good quantum state, it's a classical
         | superposition of incoherent photons that you can't really use
         | for any quantum control.
         | 
         | But even more than that, this seems to me like a purely on-chip
         | solution. For trapped ions and neutral atoms you really need to
         | translate to free-space optics at some point.
        
           | sunray2 wrote:
           | Indeed, it is nuanced, as you point out. For example, you
           | can't just attenuate a laser and use that as a single photon
           | source (instead you'd get a coherent state). To realise a
           | true single photon source you need an additional (quantum)
           | process, like controlled stimulated emission from single
           | atoms, or driving some nonlinear crystal to generate photon
           | pairs (that's using spontaneous parametric down conversion, i
           | think). And that's where the coherence properties of the
           | laser are essential.
           | 
           | As for fully integrated optics, it's where quantum computers
           | eventually want to be, and there's no physical limitations
           | currently. But perhaps it's too early to say whether we would
           | absolutely require free space optics because it's impossible
           | to do some optics thing another way.
        
       | dartharva wrote:
       | I wonder if I will ever see a photonic CPU in my lifetime.
       | Probably not, you'll have to invent a completely new material
       | never seen before that somehow enables nonlinear interactions
       | with light signals. It'd be nothing short of magic.
        
       | cubefox wrote:
       | This article is misleading. TSMC doesn't "bet" on the tech by
       | Avicena (the startup in question). Instead, Avicena appears to
       | simply pay TSMC to help them with manufacturing. Here is the
       | linked press release by Avicena:
       | 
       | https://www.businesswire.com/news/home/20250422988144/en/Avi...
       | 
       | Noting also that there have been multiple articles on IEEE
       | Spectrum about this startup in the past, I really hope the
       | journalists don't own the stock or are otherwise biased.
        
         | BardiaPezeshki wrote:
         | TSMC is developing custom detectors for Avicena on their own
         | dime. They almost never do anything like for start-ups. That is
         | why it is a bet on Avicena. They are nurturing this technology
         | because they think it has real potential. See the TSMC quotes
         | in the article.
        
       ___________________________________________________________________
       (page generated 2025-05-27 23:01 UTC)