Intel’s Integrated Photonics Solutions Group attended the 2024 Optical Fiber Communication Conference where it debuted its OCI. The fully integrated optical compute interconnect utilized an Intel CPU to demonstrate the new technology’s ability to transmit 32 Gbps of data across 64 channels. Intel’s Integrated Photonics Solutions Group has said the OCI is capable of spanning up to 100 meters of fiber optics and designed to use lower power. OCI is being developed to aid in AI applications where the scalability of CPU/GPU clusters, data transmission rates, and low power consumption are a focus.
“The ever-increasing movement of data from server to server is straining the capabilities of today’s data center infrastructure, and current solutions are rapidly approaching the practical limits of electrical I/O performance. However, Intel’s groundbreaking achievement empowers customers to seamlessly integrate co-packaged silicon photonics interconnect solutions into next-generation compute systems. Our OCI chiplet boosts bandwidth, reduces power consumption and increases reach, enabling ML workload acceleration that promises to revolutionize high-performance AI infrastructure.”
–Thomas Liljeberg, senior director, Product Management and Strategy, Integrated Photonics Solutions Group
Press Release Excerpts:
“AI-based applications are increasingly deployed globally, and recent developments in large language models (LLM) and generative AI are accelerating that trend. Larger and more efficient machine learning (ML) models will play a key role in addressing the emerging requirements of AI acceleration workloads. The need to scale future computing platforms for AI is driving exponential growth in I/O bandwidth and longer reach to support larger processing unit (CPU/GPU/IPU) clusters and architectures with more efficient resource utilization, such as xPU disaggregation and memory pooling.”
“Electrical I/O (i.e., copper trace connectivity) supports high bandwidth density and low power, but only offers short reaches of about one meter or less. Pluggable optical transceiver modules used in data centers and early AI clusters can increase reach at cost and power levels that are not sustainable with the scaling requirements of AI workloads. A co-packaged xPU optical I/O solution can support higher bandwidths with improved power efficiency, low latency and longer reach – exactly what AI/ML infrastructure scaling requires.”
As explained by Intel’s Integrated Photonics Solutions Group, the first OCI implementation achieved a staggering 4 terabits per second (Tbps) bidirectional data transfer rate. That version was compatible with PCIe Gen5.
Press Release Excerpt:
“The live optical link demonstration showcases a transmitter (Tx) and receiver (Rx) connection between two CPU platforms over a single-mode fiber (SMF) patch cord. The CPUs generated and measured the optical Bit Error Rate (BER), and the demo showcases the Tx optical spectrum with 8 wavelengths at 200 gigahertz (GHz) spacing on a single fiber, along with a 32 Gbps Tx eye diagram illustrating strong signal quality.”
OCI’s next version was adapted to support 64 channels capable of up to 32 Gbps of data transmission in each direction with a maximum length of 100 meters. Intel states that depending on application design the transmission lengths may be limited to tens of meters. Intel adds that OCI can operate at a third of the power needed for pluggable optical transceiver modules.
Press Release Excerpt:
“The current chiplet supports 64 channels of 32 Gbps data in each direction up to 100 meters (though practical applications may be limited to tens of meters due to time-of-flight latency), utilizing eight fiber pairs, each carrying eight dense wavelength division multiplexing (DWDM) wavelengths. The co-packaged solution is also remarkably energy efficient, consuming only 5 pico-Joules (pJ) per bit compared to pluggable optical transceiver modules at about 15 pJ/bit. This level of hyper-efficiency is critical for data centers and high-performance computing environments and could help address AI’s unsustainable power requirements.”
Intel’s Integrated Photonics Solutions Group further explains how using a hybrid laser-on-wafer technology combined with direct integration has allowed higher yields at lower costs. Intel is already at work in developing the next generation of OCI which will support 200G/lane PICs capable of 800 Gbps and up to 1.6 Tbps.
Press Release Excerpts:
“This unique approach enables Intel to deliver superior performance while maintaining efficiency. Intel’s robust, high-volume platform boasts shipping over 8 million PICs with over 32 million integrated on-chip lasers, showing a laser failures-in-time (FIT) rate of less than 0.1, a widely utilized measure of reliability that represents failure rates and how many failures occur.”
“These PICs were packaged in pluggable transceiver modules, deployed in large data center networks at major hyperscale cloud service providers for 100, 200, and 400 Gbps applications. Next generation, 200G/lane PICs to support emerging 800 Gbps and 1.6 Tbps applications are under development.”
Discussion (6 replies)
Join Discussion →Very nice. Optical interconnect.
That is really cool.
The first application here is likely in the server / datacenter space, but if I am understanding this correctly (and it is very possible I am not), if they can make this work on a mass produced scale (which is always the difficult part compared to getting it to work once or twice in a lab environment) it could solve many if not all of the latency issues with chiplet interconnect.
You could get to the point where you could have completely modular core units, scalable as high as a package integrator wants to go, connecting to memory controllers and other typical on package functions, and absolutely kill the yield problems with large monolithic dies.
It could also alleviate many signal pathway issues for motherboard manufacturers
Inahve to wonder at the actual power requirements as compared to the silicone pathway interconnect power needs and the logic processors converting the optical data to digital and back. Changing mediums has to have a cost in power and potentially latency. Until we get full optical chips and pathways of course.
My gut tells me the same thing, but this appears to not be the case with current generation optical network transceivers that go into switches. (SFP+, QSFP+, SFP28, QSFP28, etc.) These are quite significantly lower power and lower latency than their copper equivalents, both in theory and in practice.
I actually saw pretty significant performance gains - both in latency and in bandwidth - by switching from 4gbit copper QSFP+ DAC cables to Optical transceivers when I was setting up my 40gig network...
...and that was comparing the direct attach copper cable directly linking two NIC's to four sets of 40 Gbit QSFP+ transceivers going throug a switch!
In other words:
NIC -> QSFP+ Optical transceiver -> Fiber Optic -> QSFP+ Optical transceiver -> Switch -> QSFP+ Optical Transceiver -> Fiber Optic -> QSFP+ Optical Transceiver -> NIC
was much lower latency and not insignificantly higher bandwidth than:
NIC -> Direct Attach Copper Cable -> NIC
When compared side by side between the same two machines.
How that is possible, I don't fully understand, but the numbers from my testing (which I sadly didn't record for posterity) don't like.
And they pretty much line up with the expectations of all of the datacenter type folks I have spoken to about the issue.
Maybe it's magic? :p
Anyway, move all of this on die, and its possibly even faster?
This leads me to believe that the DACs are doing something weird / inefficient. I assume you have a similar conclusion, as adding two media transitions (electrical -> optical, then optical back to electrical) has to add latency. Could just be lower quality by design.
What we do know is that at higher bandwidths, pushing signals through copper is getting much, much harder over any distance, in terms of signal integrity as well as power.
That's the idea.
Basically along with the above, and as we've started to see with longer TB4 / HDMI / DP / ethernet cabling, higher-bandwidth over any non-trivial distance becomes nearly impractical over copper. To the point that silicon photonics become not just the more performant option but also the more economical option as well. Probably :).
Soo, they going cheap and available or expensive and exclusive?
Thinking about this a bit more. Most every type of transmission at least at the enterprise level involves Fiber switches and HBA's for the fiber. So there are performance improving options there that simply don't exist at the consumer grade.
Mainly some intelligent processing and caching on the controllers.
Will be interesting to see how the implementation makes it down to the SOC level.