Infrastructure
Flagship hardware reference
Reference classes per suite — compute, memory, GPU, I/O, and the monitoring chain each department signs off on. Classes, not SKUs: procurement picks the current generation.
Suites
- Compute
- Apple-silicon Ultra-class or Threadripper PRO-class workstation
- Memory
- 128 GB unified / ECC
- GPU
- Dual-decoder media engine or 24 GB+ discrete GPU
- I/O
- NVMe scratch (7 GB/s class) + 25GbE to shared storage
- Monitoring
- Dual 27″ 5K editorial displays + client monitor via 12G-SDI I/O card
Cuts 8K raw multicam without proxy; proxies still generated for remote review.
- Compute
- Threadripper PRO-class, dual-GPU Resolve build
- Memory
- 256 GB ECC
- GPU
- 2× 24 GB+ discrete GPUs (one dedicated to viewer)
- I/O
- 12G-SDI DeckLink-class output, calibrated signal chain
- Monitoring
- Grade-1 HDR mastering display (1000-nit class), bias lighting, panel control surface
Reference chain is calibrated end-to-end; grades are never approved on the GUI viewer.
- Compute
- Mac Studio-class DAW host + dedicated DSP
- Memory
- 64 GB
- GPU
- Integrated (video playback only)
- I/O
- Thunderbolt audio interface, 32+ channel, sub-2ms round trip
- Monitoring
- Treated room, 5.1/7.1 calibrated monitors + Atmos-ready option, LUFS metering
Mix law: dialogue anchor −16 LUFS short-form, −23 LUFS broadcast deliverables.
- Compute
- High-core-count CPU + top-tier CUDA GPU
- Memory
- 192 GB
- GPU
- 48 GB-class GPU for sim + render
- I/O
- Local NVMe cache array, 25GbE
- Monitoring
- Color-managed 4K reference-adjacent display
Heavy sims farm out; the box is for lookdev and iteration speed.
- Compute
- GPU inference node (local) + provider API lanes (cloud)
- Memory
- 128 GB
- GPU
- 24 GB+ VRAM for local upscale/interp; gen-video runs on provider cloud
- I/O
- 10GbE minimum to MAM
- Monitoring
- Standard editorial display
Voice/avatar/gen-video are provider-cloud first; local GPU handles enhancement passes.
Shared infrastructure
All-NVMe shared array, 25/100GbE spine, project-per-volume
Active cuts, dailies, conform media
High-capacity HDD array with NVMe cache tier
Completed projects awaiting archive decision
LTO-9-class tape pairs + geo-separated object storage copy
Masters, camera originals, 3-2-1 rule enforced
GPU encode nodes with hardware AV1/HEVC encoders, queue-managed
Transcode, stitch, deliverable matrix renders
25GbE to suites, 100GbE spine, 10GbE floor minimum
Shared-storage editing without local media copies
12G-SDI + HDMI 2.1 client feeds, frame-accurate remote review streams
Director/client review sessions, in-room and remote
Full spec with tier fallbacks in docs/HARDWARE.md.