In the final part, Chris Cole, Ranovus' Hojjat Salemi, and Infinera's Harald Bock share their thoughts about the revent OFC show.
Chris Cole, Consultant
OFC and optics were back with a vengeance. The high level of excitement and participation in the technical and exhibit programmes was fueled by artifical intelligence/ machine learning (AI/ML). To moderate this exuberance, a few reality checks are offered.
During the Optica Executive Forum, held on the Monday, one of the panels was with optics industry CEOs. They were asked if AI/ML is a bubble. All five said no. They are right that there is a real, dramatic increase in optics demand driven by AI/ML, with solid projections showing exponential growth.
At the same time, it is a bubble because of the outrageous valuations for anything with an AI/ML label, even on the most mundane products. Many booths in the Exhibit Hall had AI/ML on their panels, for the same product types companies have been showing for years. Some of the start-ups and public companies presenting and exhibiting at OFC have frothy valuations by claiming to solve compute bottlenecks. An example is optically interconnecting memory, which sends investors into a frenzy, as if this has not been considered for decades.
The problem with a bubble is that it misallocates resources to promises of near-term pay off, at the expense of investment into long-term fundamental technology which is the only way to enable a paradigm shift to optical AI/ML interconnect.
I presented a version of the below table at the OFC Executive Forum, pointing out that there have only been two paradigm shifts in optical datacom, and these were enabled by fundamentally new optical components and devices which took decades to develop.
My advice to investors was to be skeptical of any optically-enabled breakthrough claims which simply rearrange or integrate existing components and devices. As with previous bubbles, this one will self-correct, and many of the stratospheric valuations will collapse.
Source: C. Cole.
A second dose of reality was provided by Ashkan Seyedi of Nvidia, in several OFC forums, illustrated by the Today’s Interconnect Details table below (shared with permission).
Source: Ashkan Seyedi, Nvidia
He pointed out that the dominant AI/ML interconnect continues to be copper because it beats optics by integer or decade better metrics of bandwidth density, power, and cost. Existing data centre optical networking technology cannot simply be repackaged as optical compute input-output (I/O), including optical memory interconnect, because that does not beat copper.
A third dose of reality came from Xiang Zhou of Google and Qing Wang of Meta in separate detailed analysis presented at the Future of LPO (Linear Pluggable Optics) Workshop. They showed that not only does linear pluggable optics have no future beyond 112 gigabits per lane, but even at that rate it is highly constrained, making it unsuitable for general data centre deployment.
Yet linear pluggable optics was one of the big stories at OFC 2024, with many highly favourable presentations and more than two dozen booths exhibiting it in some form. This was the culmination of a view that has been advanced for years that optics development is too slow, especially if it involves standards. LPO was moved blazingly fast into prototype hardware without being preceded by extensive analysis. The result was predictable as testing in typical large deployment scenarios found significant problems.
At OFC 2025, there will be few if any linear pluggable optics demos. And it will not be generally deployed in large data centres.
Coincidently, the OIF announced that it started a project to standardise optics with one digital signal processor (DSP) in the link, located in the transmitter. This was preceded by analysis, including by Google and Meta, showing good margin against the types of impairments found in large data centres. The expectation is that many IC vendors will have DSP on transmit-only chips soon, including likely at OFC 2025.
A saving grace of linear pluggable optics may be the leveraging of related OIF work on linear receiver specification methodology. Another benefit may be the reaffirmation that real progress in optics is hard and requires fundamental understanding. Shortcutting of well-established engineering practices leads to wasted effort.
Real advances require large investment and take many years, which is what is necessary for optical AI/ML compute interconnect. Let’s hope investors realise this.
Hojjat Salemi, Chief Business Development Officer, Ranovus
Hyperscalers are increasingly recognising that scaling AI/ML compute demands extensive optical connectivity, and the conventional approach of using pluggable optical modules is proving inadequate.
The network infrastructure plays a pivotal role in the compute architecture, with various optimisation strategies depending on the workload. Both compute scale-up and scale-out scenarios necessitate substantial connectivity, high-density beach-front, cost-effectiveness, and energy efficiency. These requirements underscore the advantages of co-packaged optics (CPO) in meeting the evolving demands of AI/ML compute scaling.
It is great to see prominent tier-1 vendors like Nvidia, AMD, Broadcom, Marvell, GlobalFoundries, and TSMC embracing co-packaged optics. Their endorsement shows a significant step forward, indicating that the supply chain is gearing up for high-volume manufacturing by 2026. The substantial investments being poured into this technology underscore the collective effort to address the pressing challenge of scaling compute efficiently. This momentum bodes well for the future of AI/ML compute infrastructure and its ability to meet the escalating demands of various applications.
What surprise me was how fast low-power pluggable optics fizzled. While initially shown as a great technology, linear pluggable optics ultimately fell short in meeting some critical requirements crucial to Hyperscalers. Although retimed pluggable optical modules have been effective in certain applications and are likely to continue serving those needs for the foreseeable future, the evolving demands of new applications such as scaling compute necessitate innovative solutions like co-packaged optics.
The shift towards co-packaged optics highlights the importance of adapting to emerging technologies that can better address the unique challenges and requirements of rapidly evolving industries like hyperscale computing.
Harald Bock, Vice President Network Architecture, Infinera
I am impressed by the range of topics, excellent scientific work and product innovation each time I attend OFC.
Normally, the show's takeaways differ among the participants that I talk to. This year, most of the attendees I chatted agreed on the main topics. The memorable items this year ranged from artificial intelligence (AI) data centres, 800 gigabit-per-second (Gbps) pluggables, to the Full Spectrum Concert at Infinera’s OFC party that was held on the USS Midway.
AI is becoming the key driver for network capacity. While we are a very technology-driven industry, the interest in different technologies is driven by the business opportunities we expect. This puts AI at the top of the list. It is not the AI use cases in network operations, planning, and analytics, which are all progressing, but rather the impact that deploying AI data centres will have on network capacity and particularly on optical interfaces within and between data centres.
The interest was clearly amplified by the fact that recovery of the telecom networks business is only expected in the year’s second half.
Short term, AI infrastructure creates massive demand for short-reach interconnect within data centres, with longer-reach inter-data centre connectivity also being driven by new buildouts. So, we can expect AI to be the key driver of network bandwidth in the coming years.
It is in this context that linear pluggable optics has become an important candidate technology to provide efficient, low-energy interconnect, and as a result, it generated a huge amount of interest this year, stealing some of the attention that co-packaged optics or similar approaches have received in the past. Overall, AI use cases drove huge interest in 800Gbps pluggable optics products and demonstrations at the show.
Reducing interface and network power consumption have become key industry objectives. In all of these use cases and products, power consumption is now the main optimisation goal in order to drive down overall data centre power or to fit all pluggable optics into the same existing form factors (QSFP-DD and OSFP), even at higher rates such as 1.6Tbps.
I do believe that reducing power consumption, be it per capacity, or per reach x capacity depending on use case, has become our industry’s main objective. Looking at projected capacity growth that will continue at 35 to 40 per cent per year across much of cloud networks, that is what we all should be working on.
Another observation is that power consumption and capacity per duct have replaced spectral efficiency as the figure of merit. You could say that this is starting to replace the objective of increasing fibre capacity that our industry has been working under for many years.
We have all discussed the fact that we are no longer going to be able to easily increase spectral efficiency as we are approaching Shannon’s limit. In order to further increase fibre capacity, we have been talking about additional wavelength bands, with products now achieving beyond 10-terabit transmission bandwidth with Super C- and Super L-band and the option to add the S-, O-, and U- bands, as well as about spatial division multiplexing, which today refers to the use of multiple fibre cores to transmit data.
Before OFC, I was puzzled about the steps we, as an industry, would take since all of these require more than a single product from one company. Indeed it is an ecosystem of related components, amplifiers, wavelength handling, even splicing procedures. After OFC, I am now confident that uncoupled multi-core fibre is a good candidate for a next step, with progress on additional wavelength bands not at all out of the picture.
There is one additional point I learned from looking at this topic. In real-world deployments today, multi-core fibre will accelerate a massive increase in parallel fibres that are being deployed in fibre ducts across the world. To me, that means that while we are going to all focus on power consumption as a key measure for innovation, we should really use capacity per duct as an additional figure of merit.
In terms of technological progress, I would like to call out the area of quantum photonics.
We all saw the results from an impressive research push in this area, with complex photonic integration and interesting use cases being explored. The amount of work done in this area makes it difficult for me to keep up to speed. I continue to be fascinated and excited about the work done.
An entirely different category of innovation was shown in the post-deadline session where Microsoft and University of Southampton presented hollow-core fiber with a record 0.11 dB/km fiber loss. While we have been talking about the great promise of anti-resonant hollow-core fiber for a while as it offers significantly reduced latency, it reduces signal distortion by removing nonlinearity and offering low dispersion. All that has been shown before, but achieving a fibre loss that is considerably lower than that of all other fibre types is excellent news.
It confirms that hollow-core fiber could change the systems and the networks we build, and I will continue to keep close tabs on the progress in this area.
Overall, OFC 2024 was a great show, with my company launching new products and having a packed booth full of visitors, a large number of customer engagements, and meetings with most of our suppliers.
I left San Diego already looking forward to next year's OFC.