DSP vs LPO Technologies
High-performance AI and data centers face critical power constraints due to network interconnects between GPUs, CPUs, and storage.
New optical solutions – known as LRO and LPO – are leveraging silicon photonics (SiPh) technology to offer a solution by removing the Digital Signal Processor (DSP), reducing power consumption by up to 40%.
What is LPO, for starters?
LPO stands for “Linear-drive Pluggable Optic”. LRO stands for “Linear Receiver Optic” and are sometimes called “half-retimed pluggable optics”.
Both are attempts to reduce power consumption requirements by removing components that account for much of the power usage (and heat generation) in high-speed transceivers.
So, what exactly makes these pluggable optics different from existing transceivers? To answer this, we must first look at the technology that LRO- and LPO-based transceivers have been developed to improve upon – the DSP-Based transceiver.
DSP-Based Transceiver Background
Traditional transceivers have operated the same way for a very long time.
The transmitter end of the link is performing a digital-to-analog conversion (DAC) of the signal so that it can be transmitted via laser over fiber optic cable.
As a result, the receiver end of the link is performing an analog-to-digital conversion (ADC) so that the signal is back to a digital one.
This is a key factor to remember regarding fiber optic data transmission: fiber signals are converted from a digital signal and must be converted back again when received. Without getting into signal and sampling theory, there is no ‘perfect’ way to convert from digital to analog, or analog back to digital. There are technologies that allow for transmissions to take place within the limits which are acceptable for whatever the communications standard(s) dictate. However, the management of signals becomes more complicated as data rates and link distances increase. This requires more to be done by digital signal processors (DSPs); however, signal processing requires energy, and DSPs produce heat.
Not all transceivers require DSPs, but the need for additional signal processing arose as data rates and distances increased. A DSP can have digital clock recovery functionality, as well as dispersion compensation functionality. These functions help to reduce signal distortions – which can reduce bit error rate of the system to be within acceptable performance limits.
These functionalities are built into standard high-speed transceivers via the DSP. DSP circuitry consumes about 50% of the transceiver power. By moving DSP functionality to the host ASIC’s SerDes and removing from the optics, the goal is a reduction of the optics power of 50% and overall system power (related to signal processing) by 25%.
LPO/LRO Background
The linear-drive technology of the LPO was developed to remove the DSP/CDR chip in the transceiver and move those functions into the switch chip on the host.
What is left inside of the transceiver is the driver chip and the trans-impedance amplifier (TIA). The driver chip has Continuous Time Linear Equalization (CTLE) built in, and the TIA has equalizer functionality, which does provide some compensation of the signaling,
LPO Module
LPO (Linear Pluggable Optics) transceivers do not have DSP chipsets on-board, so as a result, LPO relies on the host to handle retiming and signal conditioning, unlike traditional fully retimed optical modules. By removing the DSP, LPO reduces power consumption without sacrificing high-speed data transmission, but requires careful engineering to ensure that effective links can exist.
Somewhat of a stop-gap or intermediate solution, an LRO-based transceiver focuses on the transmitter operation, keeping a single DSP on the transmit portion of the device. Because it removed the receiver DSP, these devices have become known as “half-retimed” as only the transmitter retains the functionality.
LRO Module
Removal of half of the DSP chipset provides some power reduction to the overall module, which is beneficial. It also retains some of the retiming functionality for the module. Just like with LPO, the responsibility for signal recovery shifts to the host system, which must be able to manage signal integrity to ensure optimal performance.
LPO modules and LRO modules require more integration and ‘tuning’ of the host with the transceiver, unlike a DSP-based solution. The LPO-based modules are more complicated than an LRO-based module, so LRO is still a viable option as LPO technology develops and becomes standardized.
There is an LPO MSA that has been created, and this will lead to standardization of communication, calibration, and interoperability. At this time, one of the biggest challenges in interoperability, so early deployments have been successful in systems with a single supplier for host and transceivers. Unlike traditional optics, the LPO/LRO optics must be calibrated once in the system, and it is an end-to-end calibration which involves the onboard ASICs.
When to Choose an LPO-based Transceiver?
An LPO-based transceiver is not ideal for all applications. Because there is no DSP, these transceivers are more sensitive to fiber issues (bends, splices, connectors) because there is nothing onboard to compensate for signal issues due to fiber plant.
It is for this reason that LPO transceivers are best suited for shorter reach applications (500m to 2km range). DSP-based transceivers are still the design of choice for longer reach applications (>10km) because they can compensate as needed.
By eliminating the in-optic DSP, an LPO-based transceiver is more power-efficient, as there isn’t a processor chip onboard. For an 800G transceiver, this could mean a reduction from ~12W of power to 4-6W per module. This also means that there is less heat being generated, which allows for designs of systems with higher density ports. In addition to the increased port density, this lower heat generation allows for a reduction in cooling requirements for these systems.
Power consumption, cooling solutions, and increased density are critical to designers of data centers. Coupled with ever-increasing bandwidth requirements, we start to see why LPO-based transceivers are vital for the next generation of data centers and AI deployments.
The good news is that most connection requirements for data centers are high data rate, and short distance. However, there is also the need to have lower latency solutions in many applications. This is another area where LPO-based transceivers outperform DSP-based designs. A DSP, by its nature, will have processing overhead that can impact certain low-latency applications (high frequency training, for example). No DSP means no overhead penalty.
This trade-off in transceiver design comes with some caveats.
An LPO-based transceiver has no digital signal processing or compensation on-board. But this must occur somewhere. As mentioned earlier, the host system must take on this task. The LPO-based transceiver solution much be tightly integrated with the host ASIC, as it will be handling signal processing and conditioning. These functions include modulation, compensation, equalization, and forward error correction (FEC).
When all the features and benefits of an LPO-based transceiver are combined, we find that they will be more cost-effective (no DSP component onboard) but will require host systems that can support the functionality removed from the transceiver.
They are not a universal solution, and DSP-based solutions will still have a place in network architecture.
The following is a high-level review of the two technologies and their key features, which provides guidance on which technology deployment should be considered:
● DSP-based transceivers:
○ Longer reach (>10km, typically metro and long-haul networks).
○ Plug-and-play deployment with minimal ‘tweaking’
○ Power consumption is not a primary concern
○ Latency is not a primary concern
● LPO/LRO-based transceivers:
○ Reducing power and cooling requirements is a primary concern
○ Increased port density is a primary concern
○ Low-latency connections are a requirement (applications like AI/ML workloads)
○ The host ASICs can be modified for signal conditioning and FEC.
Conclusion
LPO and LRO technology offers a compelling balance of power efficiency, cost, and speed. It's a key solution for modern data centers and AI clusters seeking to optimize performance while minimizing environmental impact. While the technology is relatively new, standardization efforts are underway via the LPO MSA to align the needs of the industry to the capabilities and plans of the developers and manufacturers of optical components and host systems.
References
OIF Energy Efficient Interfaces
OIF Common Electrical I/O (CEI)-112G