Product Spotlight: SiT5977 Super-TCXO Enables New Architecture for Greater AI Workload Efficiency in Datacenters

|by |

Artificial intelligence (AI) is accelerating advancements unlike any technology before. AI can analyze massive datasets and deliver answers in fractions of a second. It's driving progress, everything from immensely capable virtual assistants to new medical breakthroughs and an even better understanding of the universe's origins. Along with these advancements, datacenters must contend with the challenges of increased energy consumption and rapidly rising costs. To help address these challenges, AI workload efficiency is taking center stage.

The new SiTime SiT5977 Super-TCXO^® helps system engineers better optimize AI workload efficiency by enabling new architectures in smart network interface cards (Smart NICs), acceleration cards, switches and compute node applications. The new SiT5977 Super-TCXO is unique in the timing industry-enabling AI workload efficiency with 3X better synchronization and 800G network connectivity in a 4X smaller footprint.

How Precision Timing Can Enable Greater AI Workload Efficiency

By improving AI workload efficiency, datacenters can lower total cost of ownership (TCO) and improve energy profiles and carbon footprints. "Improving AI workload efficiency to reduce energy consumption and carbon emissions is an industry-wide challenge. Precision timing is one of the approaches to help solve this problem," said Dave Altavilla, president and principal analyst at HotTech Vision & Analysis. "SiTime is the only semiconductor company fully dedicated to developing innovative timing solutions required for the complex scaling of today's AI datacenters."

Large AI clusters contain tens of thousands of accelerators, and AI workloads are divided among all the accelerators. The maximum utilization of the cluster depends on both end-to-end network synchronization and high data bandwidth.

Synchronizing the data transmission with precision timing enables proper orchestration of AI data workloads among the accelerators to ensure maximum efficiency. Precision timing's highly tuned synchronization also enables precise telemetry, allowing insights for operators to eliminate underperforming accelerator nodes. In addition, tighter synchronization allows architects to reduce the guardband between data packets, called the window of uncertainty. A smaller window of uncertainty reduces the chance of overlaps, therefore reducing the frequency of data re-transmits.

Equally important to synchronization is fully utilized network bandwidth. To ensure that AI workloads are done as quickly and efficiently as possible, it's critical to remove bandwidth bottlenecks and ensure that the high bandwidth can be fully leveraged by the compute node. Together, precision synchronization and higher bandwidth utilization increase the efficiency of the entire datacenter.

SiT5977 Super-TCXO Enables Streamlined AI Architecture

The SiT5977 Super-TCXO streamlines AI system architecture. This new chip integrates a low-jitter PLL to support 800G bandwidth with SiTime's ultra-stable DualMEMS™ stability engine. The SiT5977 further integrates digital frequency control to enable system-level synchronization. It's the only single-chip timing solution that delivers the most resilient performance for AI compute nodes with high bandwidth and network synchronization.

The SiT5977 Super-TCXO performs much better than quartz timing devices which have poor resistance to environmental vibration and airflow. Thermal shock and airflow in the datacenter can disrupt quartz oscillator precision and lead to dropped packets and poor synchronization. The SiT5977 operates at ±1 ppb/°C stability over temperature slope (dF/dT) to ensure a precise time reference under thermal shock and airflow. This means no performance throttling, fewer dropped packets and smaller windows of uncertainty.

In addition to the leading dF/dT of the Si5977, its ultra-low jitter, just 80 fs from 12kHz to 20MHz, is crucial to enabling 800G bandwidths for the most capable AI datacenters. Bottlenecks in bandwidth cause the accelerators to idle while they wait for data, sometimes up to 50% of their run time. High bandwidth in the datacenter infrastructure, including the NIC, is critical to reducing the accelerator idling time. SiT5977's class-leading jitter performance removes data bottlenecks, ensuring that accelerators can run as efficiently as possible.

By integrating the high stability over temperature for nanosecond synchronization, ultra-low phase jitter for 800G bandwidth and high-resolution digital frequency control into a 5 mm x 3.5 mm package, the jitter cleaner used in the traditional quartz architecture can be completely removed. The 4X smaller footprint, compared to the competing architecture, enables larger processors in compact systems. This allows system architects to fully leverage high-speed 800G bandwidth network connectivity and maximize utilization of the AI cluster.

SiT5977 Features:

• Environmentally robust with ±1 ppb/°C frequency slope (dF/dT) for optimum performance under airflow, thermal shock.
• Capable of driving 800G and higher links via 80 fs phase jitter and LVDS outputs.
• Enables embedded control loops with precise digital tuning of output frequency (DCTCXO), ±400 ppm pull range and 0.05 ppt (5e-14) resolution via I2C/SPI.
• Eliminates link flaps from quartz timing activity dips or micro jumps.
• Resistant to shock, vibration and board bending.
• Eliminates external LDOs via on-chip voltage regulators.
• ±100 ppb frequency stability over temperature with DualMEMS ™ architecture.
• 156.25 MHz output frequency enabling high-speed SerDes and 800G links.

Distributed applications, such as AI training and inference, need precision timing to synchronize workloads. The SiT5977 enables a new, more efficient architecture to lower energy costs and optimize bandwidth utilization, helping to unlock the full promise of AI.

If you would like to learn more, sign up for the New Architecture for Greater AI Workload Efficiency in Datacenters webinar.

The SiT5977 Super-TCXO is in production and samples are available now.