Intel has developed an 8-lane parallel-wavelength laser array to tackle the growing challenge of feeding data to integrated circuits (ICs).
Optical input-output (I/O) promises to solve the challenge of getting data into and out of high-end silicon devices.
These ICs include Ethernet switch chips and ‘XPUs’, shorthand for processors (CPUs), graphics processing units (GPUs) and data processor units (DPUs).
The laser array is Intel’s latest addition to its library of silicon photonics devices.
A key challenge facing high-end chip design is the looming ‘power wall’. The electrical I/O power consumption of advanced ICs is rising faster than the power the chip consumes processing data.
James Jaussi, senior principal engineer and director, PHY research lab at Intel Labs, says if this trend continues, all the chip’s power will be used for communications and none will be left for processing, what is known as the power wall.
One way to arrest this trend is to use optical rather than electrical I/O by placing chiplets around the device to send and receive data optically.
Uisng optical I/O simplifies the electrical I/O needed since the chip only sends data a short distance to the adjacent chiplets. Once in the optical domain, the chiplet can send data at terabit-per-second (Tbps) speeds over tens of meters.
However, packaging optics with a chip is a significant design challenge and changes how computing and switching systems are designed and operated.
Intel has been developing silicon photonics technology for two decades. The library of devices includes ring-resonators used for modulation and detection, photo-detectors, lasers, and semiconductor optical amplifiers.
Intel can integrate lasers and gain blocks given its manufacturing process allows for the bonding of III-V materials to a 300mm silicon wafer, what is known as heterogeneous integration.
The company has already shipped over 6 million silicon photonics-based optical modules – the 100-gigabit PSM-4 and 100-gigabit CWDM-4 – since 2016.
Now Intel Labs has demonstrated a laser array that integrates eight distributed feedback (DFB) lasers for wavelength-division multiplexing (WDM) transmissions. In addition, the laser array is compliant with the CW-WDM multi-source agreement.
“This is a much more difficult design,” says Haisheng Rong, senior principal engineer, photonics research at Intel Labs. “The challenge here is that you have a very small channel spacing of 200GHz.”
Each laser’s wavelength is defined by the structure of the silicon waveguide – less than 1 micron wide and tens of microns long – and the periodicity of a Bragg reflector grating.
The lasers in the array are almost identical, says Rong, their difference being defined by the Bragg grating’s period. There is a 0.2nm difference in the grating period of adjacent – 200GHz apart – lasers.
The resulting eight wavelengths have uniform separation. Intel says each wavelength is 200GHz apart with a tolerance of plus or minus 13GHz, while the lasers’ output power varies by plus or minus 0.25dB.
Such performance is well inside the CW-WDM MSA specifications that call for a plus or minus 50GHz tolerance for 200GHz channel spacings and plus or minus 1dB variability in output power.
Rong says that using a 200GHz channel enables a baud rate of 64 gigabaud (GBd) or 128GBd. In future, higher wavelength counts – 16- and 32-channel designs – will be possible, as specified by the CW-WDM MSA.
The laser array’s wavelengths vary with temperature and bias current. For example, the laser array operates at 80oC, but Intel says it can work at 100oC.
The working laser array is the work of Intel Labs, not Intel’s Silicon Photonics Products Division. Intel has yet to say when the laser array will be adopted in products.
But Intel says the technology will enable terabit-per-second (Tbps) transmissions over fibre and reach tens of meters. The laser array also promises 4x greater I/O density and energy efficiency of 0.25 picojoules-per-bit (pJ/b), two-thirds that of the PCI Express 6.0 standard.
Another benefit of optical I/O is low latency, under 10ns plus the signal’s time of flight, determined by the speed of light in the fibre and the fibre’s length.
An electrical IC is needed alongside the optical chiplet to drive the optics and control the ring-resonator modulators and lasers. The chip uses a 28nm CMOS process and Intel is investigating using a 22nm process.