Designed to address growing bandwidth demands and high-level applications that rely on artificial intelligence and machine learning, Marvell Technology’s latest cloud-optimized 400G DR4 silicon photonics platform is production ready.
The company says its transceiver can provide lower costs per bit while also accelerating time to market thanks to silicon photonics.
“What silicon allows you to do is leverage the toolset within the standard silicon industry, so we can build these things out of 200-millimeter wafers, which give you scale, which is important,” said Radha Nagarajan, senior VP and CTO of the optical and copper connectivity group at Marvell. “Secondly, it gives you speed. You can build both photodetectors and modulators at speeds higher than what you can do with standalone lasers. And it also gives a path to a lower overall cost, but that depends on performance. Cost is all relative to performance.”
Marvell’s 400G DR4 silicon photonics platform offers four low-loss transmit Mach-Zehnder modulators as well as four high-responsivity, receive photodiodes. On top of that, the product boasts four transimpedance amplifiers (TIAs), an analog controller IC and the company’s own Porrima 7-nm low-power PAM4 DSP with an integrated 56Gbaud laser drive — all of which Marvell claims can help customers scale cloud data centers beyond current means.
“Traditional optical interconnects would just have a laser and you would modulate that laser. These traditional optical interconnects run out of steam at certain data rates. As you go to a higher and higher data rate, you need a different kind of modulator,” Nagarajan said. “What happens is, we usually have an external modulator. To drive the modulator you have a driver, and on the receiver end you have a photodetector. The silicon photonics integrated circuit combines the modulator and the detector, and its backend to a driver and a transimpedance amplifier.”
The key here is to develop the modulator within silicon; not on top of it.
“We have developed modulators in silicon; that part is easy. The modulator is silicon, the detector is also silicon. We’re not trying to integrate it onto silicon. They are monolithic,” Nagarajan said. “What silicon actually allows you to do so much better than other platforms is I can use a silicon photonics chip as a silicon interposer. What we can achieve is a flip-chip on top of the silicon and complete a 2.5D integration. 2.5D and 3D integration is the frontier.”
With its 400G DR4 silicon photonics platform, Marvell’s mission is to provide cloud-optimized silicon solutions capable of meeting the compute, networking and optical interconnect demands of a variety of cloud instances. The chip designer also said its 400G DR4 module enables customers to fully grasp the cost advantages of “high-volume silicon wafer-scale manufacturing.”
“The world is becoming more and more cloud-centric. All of our data is going into the cloud and coming back, but each of these clouds are pretty unique,” said Nigel Alvares, VP of solutions marketing at Marvell. “These cloud instances need silicon solutions that are optimized for what they’re servicing. And these different workloads require different types of compute, networking, optical interconnects, scale, AI-accelerators and computer accelerators that need to connect to the switch before they go across the whole data center and out of the data center.”
Marvell’s 400G DR4 platform can enable cloud data centers to transport electrical signals over intra-center distances up to two kilometers. But its COLORZ II 400ZR solution, which is now in volume production, can reach even farther distances, allowing cloud data centers to connect within a region that spans thousands of kilometers.
Based on similar silicon photonics technology, the chip designer’s latest data center interconnect product is a QSFP-DD pluggable module, an area Marvell expects to continue to expand on.
“The next trend that will be evident is can we move these modules into a smaller form factor and move them onto the cards themselves. Closer and closer to the processor, closer and closer to where the work gets done. That’s a whole new trend called co-packaged optics, which lends itself really well to silicon photonics,” Nagarajan said.
The advantage here, according to the chip designer, is that its 400ZR module is capable of providing comparable levels of performance and power to that of a transport box, but in a significantly smaller package. Thereby placing data centers closer to end users thanks to the footprint of silicon photonics technology.
“Silicon photonics allows you to integrate some complex functionality into a much smaller space for better cost, better power and then you can optimize it for such short reach; that’s co-packaging silicon photonics on top of the same substrate as a CPU or a switch,” said Josef Berger, associate VP of marketing for the optical and copper connectivity group at Marvell. “That all the way to thousands of kilometers connectivity over fiber, that you can also do based on similar silicon photonics technology and components, obviously built up a little more complex and optimized for that.”