Consumes 1/3 the power of optical, but costs 1/3 more than optical:
Scale-up connectivity is crucial for the performance of rack-scale AI systems, but achieving high bandwidth and low latency for such interconnections using copper wires is becoming increasingly complicated with each generation. Using optical interconnections for scale-up connectivity is a possibility, but it may be an overkill, so start-ups Point2 and AttoTude propose to use radio-based interconnections operating at millimeter-wave and terahertz frequencies over waveguides that connect to systems using standard pluggable connectors, reports IEEE Spectrum.
Point2's implementation uses what it calls an 'active radio cable' built from eight 'e-Tube' waveguides. Each waveguide carries data using two frequencies — 90 GHz and 225 GHz — and plug-in modules at both ends convert digital signals directly into modulated millimeter-wave radio and back again. A full cable delivers 1.6 Tb/s, occupies 8.1mm, or about a half the volume of a comparable active copper cable, and can reach up to seven meters, more than enough for scale-up connectivity. Point2 says the design consumes roughly one-third the power of optical links, costs about one-third as much, and adds as little as one-thousandth the latency.
A notable aspect of Point2's approach is the relative maturity of its technology. The radio transceivers can be fabricated at standard semiconductor production facilities using well-known fabrication processes — the company has already demonstrated this approach using a 28nm chip with the Korea Advanced Institute of Science and Technology (KAIST). Also, its partners Molex and Foxconn Interconnect Technology have shown that the specialized cables can be produced on existing lines without major retooling.
AttoTude is pursuing a similar concept, but at even higher frequencies. Its system combines a digital interface, a terahertz signal generator, and a mixer that encodes data onto carriers between 300 and 3,000 GHz that feeds the signal into a narrow dielectric waveguide. Early versions used hollow copper tubes, while later generations rely on fibers measuring approximately 200 micrometers across with losses as low as 0.3 dB per meter (considerably lower than copper). The company has demonstrated 224 Gb/s transmission over four meters at 970 GHz and projects viable reaches of around 20 meters.
Both companies use waveguides instead of cables because, at millimeter-wave and terahertz frequencies cables fail. While at very high data rates copper cables can pass signals, they do so by becoming thicker, shorter, and more power-hungry. Furthermore, their losses and jitter rise so fast that the link budget collapses and breaks, so cables cannot be used for such applications. Meanwhile, waveguides are not an exotic choice, they are among a few viable option for interconnects with terabit/s-class bandwidth.
(Score: 2) by DadaDoofy on Saturday January 03, @02:52PM
"costs 1/3 more"..."costs about one-third as much"
So which is it? Words matter.
(Score: 5, Interesting) by VLM on Saturday January 03, @03:25PM (3 children)
https://spectrum.ieee.org/rf-over-fiber [ieee.org]
Something I don't like about IEEE Spectrum is sometimes it assumes its audience is very popular science non-EE focused barely graduated high school. But those people are already on 1234567 other non-EE focused sites, so essentially the IEEE has abandoned their legacy audience of EEs, which always seems wild to me. If you want "EE news" don't go to IEEE Spectrum they don't do that anymore, which is awful.
The article is essentially technical content free. "Its half the size" because they're running twice as fast so have half the (waveguide) conductors. We'll see how long that lasts as they'll likely want to use the allocated crosssectional space to the max so install more. POssibly its making lemonade out of lemons if individual bridge yields are low then using half the number of bridges will result in more working cables shipped... The "ends" terminating the individual waveguide strands are also quite expensive so using half for half speed might be acceptable.
Theres a discussion of skin effect and classic coax has a tiny little non-resonant center conductor so the skin effect disruption is worse for coax than bulkier resonant waveguide. So, yeah, make expensive waveguides instead of tiny-wire-in-a-tube.
The "standard pluggable connector" is stretching the truth a bit. Its more like an active powered waveguide bridge at each end, with, yes, technically standard ish connectors at each end. Not different conceptually from installing a fiber link instead of a normal copper ethernet cable for ethernet, just using tiny plastic plated waveguides instead of a fiber.
The summary of the IEEE summary of the manufacturer's whitepaper breathlessly all but claims its new tech, which it isn't. They're shipping "active cables" like this in the "hundreds of GHz" if you're wealthy enough, and the whitepaper claims they'll soon ship in the very low THz range, so its like 4+ times faster than whats shipped but other than being smaller and faster its the same. Like claiming a 74ALS00 chip is fundamentally new form of computation using never before seen logical processing, although technically a 74ALS000 just a faster 7400 chip that remains a plain old NAND gate...
Given that IEEE has given up on supporting a technical audience, I think a better technical summary would be via analogy that they are shipping slow active cables that are bridges on each end to waveguide. And if you're new to the tech the waveguide is made kind of like old fashioned 1980s plastic optical fiber where there's a plastic tube being covered in other stuff (conductive layer) to make a circular waveguide. Its an impressive display of miniaturization and the thermal challenges to keep this dude's innards under 150C must be the most impressive part of the entire story.
Its 2026 so they "have to" AI fad the heck out of the coverage although IRL I'm sure in the long run you'll have something like this doing more useful things like connecting between ethernet switches. Instead of a SFP port connecting two switches using optical you'll have like SFP+++++ running 1000 times faster using these VERY expensive active waveguide cables.
If you ever want to meet some optimists talk to the SFP peeps, they think they're going to ship "SFP-DD" modules running at 1600 g/s but I think they'll have trouble not melting down. Going to need pipes for liquid nitrogen to chill those dudes. Essentially the linked product is "we gonna make a competitor to SFP-DD, except we'll use waveguide instead of standard fiber because custom proprietary things are always cheaper than commodity standards (sarcasm, sorry) also unlike SFP-DD we'll actually ... ship not talk about it"). I mean yeah you can buy COTS 100 gig SFP-DD modules today for "hundreds of bucks" which is CHEAP but the SFP-DD bros think they're shipping 1600 GBPS modules ah no, not any time soon. Of course the waveguide guys aren't shipping either just suggesting they will.... Thats the "big picture" story will interconnects be moderately expensive SFP hardware or this new stuff that they claim will be cheaper but I think will end up more expensive (although maybe at the cutting edge, temporarily faster...)
(Score: 5, Interesting) by JoeMerchant on Saturday January 03, @03:58PM (2 children)
I feel like 20 years old we were pricing some commodity waveguide bus interconnect for a 100 node "supercomputer" made up of Mac Pros....
Then we ran the profiler and did a code review and sped up the algorithm implementation by 100x and the whole need for the super computer evaporated...
🌻🌻🌻🌻 [google.com]
(Score: 2) by VLM on Saturday January 03, @07:27PM (1 child)
Infiniband? Would fit the "big metallic rectangles" criteria but its not actually waveguide. I will eventually have InfiniBand in my basement. Its only a matter of time. If things were quieter at work&home I'd already have it...
Those infiniband cable ends are crazy little piece of plumbing things so I could see someone glancing and going "Oh thats about WR28 size waveguide, yeah yeah" although its not.
(Score: 2) by JoeMerchant on Saturday January 03, @08:09PM
Infiniband sounds like one of the candidates - that wasn't my perspective on the problem, I just heard the guy on the other side of the room rambling about how this one's so much faster than that one, etc. etc. but at the end of the day they were all just shoving around a bunch of bytes they didn't need to move in the first place - but they were getting lots of price lists for how to shove those bytes as quickly as possible...
🌻🌻🌻🌻 [google.com]
(Score: 2) by Frosty Piss on Saturday January 03, @06:03PM
...Startup suggest you give them all your money, because Science Fiction is now!