OVRO-LWA Time Machine System

• The OVRO-LWA Time Machine system is a two-stage voltage buffering and transient processing architecture that continuously captures raw voltage data from 352 dual-polarization antennas.
• It integrates large-scale DRAM and FPGA buffering with GPU-accelerated pipelines to achieve millisecond to sub-microsecond transient detection over a nearly full-hemisphere field of view.
• The system supports multiple applications including gravitational wave follow-up, cosmic-ray detection, RFI localization, and array diagnostics with precise timing synchronization.

The OVRO-LWA Time Machine system is a two-stage voltage buffering and transient processing architecture developed at the Owens Valley Radio Observatory Long Wavelength Array (OVRO-LWA). It facilitates continuous raw voltage capture and retrospective analysis at the antenna level, enabling high-sensitivity searches for fast radio transients, such as prompt coherent emission associated with compact object mergers, as well as supporting array diagnostics, RFI localization, and cosmic-ray studies. The system integrates large-scale DRAM and FPGA buffering, high-throughput data pipelines, and GPU-accelerated processing to provide millisecond and sub-microsecond transient capture capabilities across a nearly full-hemisphere field of view (Kosogorov et al., 21 Dec 2025, Nelles et al., 2019).

1. System Architecture and Buffering Strategy

OVRO-LWA consists of 352 dual-polarization, wideband dipoles, arranged with 241 antennas in a central 200 m core and 111 antennas in an extended array reaching 2.4 km, observing 13–86.5 MHz. The Time Machine system deploys a two-stage voltage data buffering scheme (Kosogorov et al., 21 Dec 2025):

• Stage 1 (Ring Buffer): High-rate (197.1 MS/s per polarization per antenna, 8-bit ADC) raw samples from all antennas are streamed into a circular DRAM buffer with 30-minute capacity array-wide. The buffer retains the full 55–86 MHz band, realized via analog frontend filtering, enabling pre-alert transient recovery. The total buffer plus correlator archive spans approximately 5 PB.
• Stage 2 (Offline Voltage Dump): Upon receipt of an external trigger (such as a gravitational wave alert with FAR ≤ 1/10 yr), the ring buffer is frozen and a selected 30-minute segment is offloaded to NVMe/spinning disk storage. The transient search typically analyzes the 69–86 MHz subset to optimize dedispersion (Kosogorov et al., 21 Dec 2025).

A parallel implementation based on FPGA block RAM (for sub-microsecond transients) is used in specialized workflows, where each antenna features 16k-sample circular on-FPGA buffers (12-bit, 196.608 MSPS), supporting rapid internal and external triggered scrapes of up to ≃81 µs of pre- and post-event data (Nelles et al., 2019).

2. Data Flow and Major Hardware Components

The data flow is structured as follows (Kosogorov et al., 21 Dec 2025, Nelles et al., 2019):

1. Front-end ADCs: Each antenna’s dual polarization output is sampled at 197.1 MS/s (8-bit) or 196.608 MSPS (12-bit, for sub-microsecond path), routed through analog filters.
2. Buffered Acquisition: Circular DRAM or on-FPGA RAM forms the primary buffer, with the ring buffer on 32 servers providing 30 min of retention, and FPGA buffer supporting ≃81 µs look-back.
3. Triggered Dump: Voltage data is selectively written to persistent storage upon detection of external or internal triggers, using high-throughput interfaces (PCIe DMA, 10 GbE), achieving dump rates up to 0.8–1 GB/s (Nelles et al., 2019).
4. Back-end Processing: A GPU farm comprising 12 CPU nodes and 16× NVIDIA RTX A4000 GPUs (16 GB each) runs the Bifrost streaming framework with custom DSP blocks, supporting upchannelization, calibration, beamforming, dedispersion, and transient search. An offline disk cluster archives the data products.

Timing synchronization is maintained by GPS-disciplined rubidium clocks with <1 ns drift, distributing a 10 MHz reference plus 1 PPS to all digitizers, and encoding GPS timestamps in each buffer packet to ±5 ns (Nelles et al., 2019).

3. Transient Search Processing Pipeline

3.1 Retrospective Coherent Beamforming

After gain calibration (by bandpass fitting with 10 s Cassiopeia A transits in CASA), buffered voltages are upchannelized to Δν=0.7 kHz resolution. Geometric delays are applied to synthesize “tied” beams. The voltage time series for beam $b$ is

$$y_b(t) = \sum_{i \in \text{core}} w_{i,b} \cdot x_i(t + \tau_{i,b}),$$

where $w_{i,b}$ are complex calibration gains, $x_i(t)$ the antenna voltages, and $\tau_{i,b}$ are geometric delays (Kosogorov et al., 21 Dec 2025). This configuration yields time resolution Δt=1.3 ms and supports 51 beams tiling a 50% probability gravitational wave localization region.

3.2 Dedispersion and Candidate Search

For each coherent beam:

• Dedispersion is performed incoherently across trial DMs 100–400 pc cm$^{-3}$ with DM spacings such that residual intra-trial smearing

$$\Delta t_\text{smear} \approx 4.15\ \text{ms} \cdot \Delta\text{DM} \cdot (f_\text{low}^{-2} - f_\text{high}^{-2}) \leq \Delta t_\text{samp} = 1.3\, \text{ms}$$

is below the sample time (Kosogorov et al., 21 Dec 2025).

• Boxcar filtering is applied with window widths 1–200 ms to match expected transient scattering.
• RFI excision uses Spectral Kurtosis and Savitzky–Golay filtering.
• Candidate selection involves SNR computation over all (t, DM, w) tuples, thresholding at SNR ≥ 5 for triggers, followed by DBSCAN-based clustering, and a final SNR = 7 detection threshold, yielding ≲1 false alarm per ≳10$^{12}$ trials.

4. Performance Metrics and Limits

4.1 Data Rates and Storage

• Array-wide raw stream: 197.1 MS/s per pol × 8 bits × 352 antennas ≈ 0.56 TB/s.
• Buffer dump: ≈130 GB/min; ~4 TB per 30 min set.
• Dedispersion load: 10$^6$ time samples × 2.5×10$^4$ channels × 3×10$^4$ DM trials ≈ 7.5×10$^{13}$ operations per beam.

4.2 Latency

• Alert-to-recording start: ~3 min 30 s, set by external GW alert latency.
• DRAM buffer write→storage latency: ~3 hr.
• Processing (51 beams over 30 min): presently ≳1 month due to compute limits; search is typically restricted to the 50% localization region.

4.3 Sensitivity

• Fluence threshold (7σ): ~150 Jy ms for $\Delta t = 1.3$ ms.
• Luminosity upper limit at 95% c.l.: $L_{95} = 4.0 \times 10^{41}$ erg s$^{-1}$ over $\Delta \nu = 17$ MHz, via $L = 4 \pi D^2 F \Delta \nu$ (Kosogorov et al., 21 Dec 2025).

5. Scientific and Engineering Applications

The Time Machine system underpins several major science and engineering use cases (Kosogorov et al., 21 Dec 2025, Nelles et al., 2019):

• Gravitational wave-triggered transient searches: Enables recovery of pre-alert voltage data and retrospective, wide-field fast transient search in response to GW event alerts. The S250206dm campaign exemplified this usage, recovering ~2 min of pre-alert data, forming 51 beams across the 50% localization region, and establishing a population-level constraint on merger-associated radio emission (no detection above 150 Jy ms) (Kosogorov et al., 21 Dec 2025).
• Cosmic ray air-shower science: FPGA-level buffering with sub-microsecond latency captures impulsive air-shower signatures, with an observed event recovery efficiency of ~98% (2% loss for buffer overrun when triggers occur <80 μs apart) (Nelles et al., 2019).
• RFI localization: Buffer capture and retrospectively beamformed localization enables spatial reconstruction of impulsive RFI sources with meter-scale precision.
• Array diagnostics: Playback of pulser signals enables end-to-end calibration and diagnosis of gain mismatches, polarization swaps, and timing errors (e.g., 5 ns clock glitches).

6. Implementation Lessons and Prospective Improvements

Hardware and system-level feedback indicate the following constraints and lessons (Nelles et al., 2019):

• Memory limitations: On-FPGA RAM only supports ≃81 μs per antenna; extension to >1 ms requires integration of external DDR3, which increases system complexity.
• Data uplink bottlenecks: The 10 GbE links saturate at ≳5 Hz trigger rates; upgrades to 40 GbE are under commissioning.
• Automated diagnostics: Nightly self-tests and real-time “watchdog” RFI monitors now mitigate early faults.
• Timing precision: GPS+Rubidium backbone achieves sub-nanosecond drift, with observed timing glitches being resolved via RFI phase tracking.
• Pipeline computational load: The current dedispersion and candidate search pipelines are computationally intensive; GPU kernel optimizations and higher throughput storage are planned to reduce turnaround to sub-day timescales and enable full-sky analyses (Kosogorov et al., 21 Dec 2025).

A plausible implication is that, as data rates and sky coverage increase, scaling of processing and rapid data movement will become the main technical limitations.

7. Context and External Comparison

The antenna-level voltage buffering approach aligns the OVRO-LWA Time Machine system with similar strategies under development in SKA-LOW, LOFAR, ACTA, and Parkes, where pre-correlation buffering, nanosecond-level event capture, and low-latency triggering enable both astrophysical and engineering goals (Nelles et al., 2019). The OVRO-LWA implementation couples these capabilities to large-array architectures, leverages both central and distributed buffering, and demonstrates unique data recovery for GW follow-up.

The system establishes a template for population-level constraints on coherent radio emission from compact binary mergers, with anticipated future campaigns expected to advance upper limits on such emission by an order of magnitude (Kosogorov et al., 21 Dec 2025).