Source linked

DxPTA sucht Photonic Transformers 15x schneller als erschöpfend

Eine neue Design-Space-Explorationsmethode findet Architekturen für photonische Transformator-Beschleuniger, die engen Flächen- und Leistungsbeschränkungen entsprechen und 15,2x schneller als Brute Force suchen.

dxptaphotonic computingtransformer acceleratorsdesign space explorationoptical dataflowhw sw co design

Photonic transformer accelerator design just got a search algorithm that finds a working architecture in 15.2× less time than brute force, hitting area under 50mm², power under 5W, and latency under 10ms.

The Problem with Manual Photonic Accelerator Design

Transformer models like DeiT-T/S/B and BERT-B/L keep winning benchmarks, but their sheer size makes electronic acceleration energy-inefficient. Photonic transformer accelerators (PTAs) promise big speedups and efficiency gains over silicon, but building one that actually meets application constraints—area, power, energy, latency—has been a manual slog. No existing PTA design methodology systematically explores the trade-offs while respecting hard constraints. That means designers either overshoot the budget or waste weeks iterating.

DxPTA’s Optical Dataflow-Guided Search

The DxPTA approach flips that script. First, it identifies the key architecture knobs by analyzing the coherent optical dataflow—things like number of wavelength channels, microring resonator banks, and photonic tensor core dimensions. Then it ranks those parameters by impact on the constraints. Finally, it feeds that ranking into a constraint-aware search algorithm that prunes the space early, targeting only viable regions.

Results speak in concrete numbers. For a 50mm² area, 5W power, 50mJ energy, and 10ms latency budget, DxPTA delivers designs as tight as 26mm², 4.8W, 39mJ, and 6ms. That’s not just meeting specs—it’s leaving headroom. On the search-time front, DxPTA finishes in 1/15.2 the wall-clock of an exhaustive sweep, making iterative co-design with real transformer workloads practical.

What This Unlocks for AGI Hardware

Photonic computing has long been the “next big thing” that never shipped in production. DxPTA doesn’t solve the fabrication challenges or the packaging nightmare, but it does give architects a tool to find a viable PTA configuration before committing silicon. That’s the difference between a one-off lab demo and a repeatable design flow for diverse transformer models. The next step is clear: wire up DxPTA’s search with a real photonic foundry PDK and see if the 26mm² design actually tapes out.


Source: DxPTA: An Architecture Design Space Exploration with Optical Dataflow-guided Strategy for HW/SW Co-Design of Photonic Transformer Accelerators
Domain: arxiv.org

Read original source ->

External source stays available while the OJO article and comment thread stay local.

Comments load interactively on the live page.