Computational resources
About 350 GB of raw data are produced each night, transferred from Chile to Seoul and reduced the same day.
Control computers in Chile
On-site computing consists of sixteen Telescope Control Computers, one per operational unit, and a single Main Control Computer that coordinates the array. Each TCC drives its own mount, camera, focuser and filter wheel and writes exposures to local storage as they complete. The MCC dispatches observation commands through RTCSpy, aggregates data from every TCC, and manages transfer to the processing facility at Seoul National University over KREONET.
| Telescope control computers | 16, one per operational unit |
|---|---|
| Main control computer | 1, array-level command hub |
| Dispatch | RTCSpy over real-time network |
| Link to Korea | KREONET |
| Transfer protocol | GridFTP, ≈ 80 MB s⁻¹ |
| Typical transfer time | < 12 hours (ToO: tens of minutes) |
Proton
All 7DT data are reduced on Proton, a dedicated server with dual AMD EPYC 7513 processors providing 128 cores at up to 2.6 GHz, 512 GB of memory, and two NVIDIA A100 GPUs sharing memory over NVLink. A nightly volume of roughly 3,000 raw images requires an effective per-image processing time of about 30 seconds to complete within the daily budget; the current pipeline sustains a median end-to-end throughput of 66 ± 24 GB per hour and clears a typical survey night in about five hours of wall-clock time after transfer completes. GPU acceleration is available for preprocessing, though in this deployment the throughput gain over the CPU path is minimal — the pipeline is bound by I/O rather than by computation.
| CPU | 2 × AMD EPYC 7513 |
|---|---|
| Cores | 128 @ up to 2.6 GHz |
| Memory | 512 GB (8 × 64 GB) |
| GPU | 2 × NVIDIA A100, 82 GB VRAM |
| GPU interconnect | NVLink |
| Target throughput | ≈ 30 s per image |
Storage servers
A typical night yields about 3,000 raw frames of roughly 117 MiB each, some 350 GB before compression. Raw data are compressed on site and transferred by GridFTP at a typical 80 MB/s, a procedure that usually completes in under twelve hours; target-of-opportunity data skip the compression and the wait for sunrise, cutting latency to tens of minutes. Storage is provided by a growing set of servers named for the hydrogen transition series. Two are in service — Lyman, with two 1.2 PB volumes, and Balmer, with one — for a current capacity of about 3.6 PB, and further servers are added as the archive grows. Each is attached to the compute server as an NFS mount over a 10 Gbps class network, so that I/O buffering does not burden processing.
| Naming | Hydrogen transition series — Lyman, Balmer, … |
|---|---|
| Lyman | 2 × 1.2 PB (RAID 60) |
| Balmer | 1 × 1.2 PB, extensible |
| Current capacity | ≈ 3.6 PB, extended as required |
| Attachment | NFS over 10 Gbps class network |
| Nightly inflow | ≈ 350 GB (≈ 3,000 × 117 MiB) |
The processing center
