PHASER

Methods & sources

  1. 200–3,000× less energy per step. Modeled, not built (research/2026-09-14, Exps. 31 and 34; notes/energy-per-multiply.md). Compared with an equally capable dense recurrent network (echo-state network) on an 8-bit chip at 0.2 pJ per multiply-accumulate, with PHASER extrapolated to a million optical modes. Extrapolated from simulations up to 65,000 modes, assuming the measured equivalence of 16–64 optical modes per neuron continues, and with detector channels growing with the modes (about 6 × 10⁴ at 10⁶ modes). At today's simulated size (4,096 modes) PHASER is at parity with small digital networks, simple digital algorithms (NG-RC) remain cheaper on standard benchmarks, and sparse networks remove the advantage.
  2. 1 pJ vs 0.001 pJ per multiply. Today's AI chips spend roughly 0.7–2 pJ per multiply-accumulate at chip level (NVIDIA H100 INT8 ≈ 0.71 pJ, Google TPU v4 ≈ 1.24 pJ, H100 BF16 ≈ 2 pJ), and far more when weights stream from memory. PHASER: 0.06–1 fJ (0.00006–0.001 pJ) per equivalent multiply, modeled at 10⁶ optical modes. An equivalent multiply is PHASER's energy per input step divided by the multiply-accumulates per step (N² + N) of the dense recurrent network of N neurons it matches in task quality. Modeled, not measured; sources and derivation in the notes, Exp. 34.
  3. Up to 10 B steps per second. Modeled (Exps. 32 and 34): several light pulses in flight at once, each an independent input stream sharing one optical program. In simulation the per-stream quality held up to 80 pulses. 13 pulses in the tuned ring give 8.9 × 10⁹ input steps/s; the plate stack (3.9 × 10⁹ per stream) reaches 10¹⁰ with about 3. Without multiplexing: 0.75–4 billion steps/s per stream. Needs a gain medium that recovers within about 1 ns, pulse-to-pulse leakage in the cavity below 10⁻³ per trip, and about 13 GHz detection on every readout channel (256 channels). No pulsed hardware has been built.
  4. Light loses 10–30 % per round trip. Realistic builds, modeled in Exp. 33: mirrors, modulators and glass coatings; bulk absorption in the glass is about 0.3 % of the light per trip. The gain medium replaces the loss every trip, so the state keeps circulating; the pump is what PHASER pays for, along with detection and readout electronics (Exp. 34).
  5. Watch the light compute. A live simulation in the browser of the research's linear stack (arch.ts stackCavity, Exps. 30 and 33), with the research simulator's own config: a 5 % input/output coupler with the gain (global saturation, G₀ 1.35, Exp. 30's operating point) and its 10 mm host crystal; four fabricated fused-silica phase plates, 64 × 64 pixels at 20 µm, random programs (depth 0.1), 1 mm thick with 0.1 % coatings, 5 mm apart; and a concave end mirror (R 120 mm, 99.9 %). One round trip per input step. Scalar field on a 64 × 64 grid at 20 µm, angular-spectrum propagation, air. Round trip 0.255 ns including the glass; passive retention of the dominant mode 79 % per trip (Exp. 33: 79 %). A stack of today's transmissive LC panels keeps only about 0.5 % per round trip (Exp. 33), which is why the stack uses fabricated plates; a programmable reflective SLM (dielectric LCOS, about 95 %) can serve as the end mirror. The browser port matches the research simulator bit for bit over 4,000 round trips (apps/phaser/scripts/validate-sim.ts).
  6. The drawings. The hero and the machine section draw the same simulation at true proportions, inside a bench assembly modelled at true scale after real parts (16 mm cage rods, a 1/4-inch mirror in a kinematic mount, an OV3660 camera module, an ESP32-S3-DevKitC-1); the assembly is an illustrative build, not a tested design; each drawn wavefront's cross-section is the simulated |E|² at its plane on its round trip. The model is continuous-wave, so the many wavefronts in flight stand for time-multiplexed pulses; crests are stylised. The electrons-and-light figure is a physics-motivated illustration, not a device model: atoms drawn as points sampled from a 1s orbital density, electrons hopping between atoms by two-site tunnelling that mostly falls back, each completed hop shaking the lattice (heat); light as 650 nm wave packets whose fringes close up by n = 1.5 inside the glass. Depth of field is drawn: the front layer is in focus.
  7. 128 weeks. Wood Mackenzie (Aug 2025): power-transformer lead times of about 128 weeks in Q2 2025.
  8. 5 years+. Lawrence Berkeley National Laboratory, Queued Up (2026 edition): the median time from grid-interconnection request to operation is over 5 years.
  9. $400 B+. IEA, Energy and AI (Apr 2025) and its 16 Apr 2026 update: the five largest tech companies spent over $400 B in 2025.
  10. Footage. Real b-roll from Pexels (Pexels License), Mixkit (free licence) and the US Senate Committee on Energy and Natural Resources (public domain). "SuperMUC: First Commercial Hot-Water Cooled Supercomputer to Consume 40% Less Energy" by IBM Research, CC BY 3.0, via Wikimedia Commons (excerpt, re-encoded)
← PHASERλ 650 nm