FALCON-XR : flexible and hybrid low-precision floating-point compute engine for energy-efficient XR perception acceleration

Sharma, Vijay Pratap (avtor), Venkatpurwar, Shekhar (avtor), Kumar, Shubham (avtor), Pilipović, Ratko (avtor), Lokhande, Mukul (avtor), Vishvakarma, Santosh Kumar (avtor)

The convergence of extended reality (XR) and artificial intelligence is driving increasingly demanding on-device perception workloads, yet the accelerators that serve them must meet strict constraints: low latency for responsive interaction and energy-efficient design suitable for wearable devices. These workloads, particularly within the XR perception pipeline, exhibit heterogeneous numerical sensitivity, motivating hybrid-precision compute to balance accuracy and energy. This paper presents FALCON-XR, a runtime-adaptive low-precision floating-point accelerator for energy-efficient and low-latency XR perception. The architecture is built around the Hybrid Low-Precision Floating-Point Compute Engine (HL-FPCE), which supports FP4 and FP8 execution through a shared reconfigurable datapath that operates either as a single FP8 MAC unit or as two parallel FP4 MAC units within a five-stage pipeline. Implemented in 28 nm CMOS, the design achieves a 1.78 GHz operating frequency within a 5908 μm$^2$ area and a 29.42 mW power envelope, yielding a compute efficiency of 4.12 pJ/op, an energy-delay product of 0.93 ⋅ $10^{−5}$ fJ⋅s, an energy efficiency of 79.68 GFLOPS/W, and a compute density of 122.7 GFLOPS/mm$^2$. Evaluation across representative XR perception workloads demonstrates that FP8 execution preserves near-baseline accuracy while FP4 delivers substantial reductions in latency and energy with acceptable accuracy degradation, confirming the suitability of FALCON-XR for real-time, power-constrained XR edge deployment.

• Jezik: Angleški jezik
• Ključne besede: XR perception acceleration, low-precision floating-point arithmetic, FP4/FP8 MAC engine, energy-efficient edge AI
• Vrsta gradiva: Članek v reviji
• Tipologija: 1.01 - Izvirni znanstveni članek
• Organizacija: FRI - Fakulteta za računalništvo in informatiko
• Status publikacije: Objavljeno
• Različica publikacije: Objavljena publikacija
• Leto izida: 2027
• Št. strani: Str. 1-11
• Številčenje: Vol. 112, art. 102902
• PID: 20.500.12556/RUL-190098
• UDK: 004.8:004.3
• ISSN pri članku: 0167-9260
• DOI: 10.1016/j.vlsi.2026.102902
• COBISS.SI-ID: 301494787
• Datum objave v RUL: 22.10.2026

Gradivo je del revije

• Naslov: Integration
• Skrajšan naslov: Integration
• Založnik: Elsevier
• ISSN: 0167-9260
• COBISS.SI-ID: 25627136

Licence

• Licenca: CC BY 4.0, Creative Commons Priznanje avtorstva 4.0 Mednarodna
• Povezava: http://creativecommons.org/licenses/by/4.0/deed.sl
• Opis: To je standardna licenca Creative Commons, ki daje uporabnikom največ možnosti za nadaljnjo uporabo dela, pri čemer morajo navesti avtorja.

Sekundarni jezik

• Jezik: Slovenski jezik
• Ključne besede: pospeševanje zaznavanja za razširjeno resničnost XR, aritmetika s plavajočo vejico nizke natančnosti, MAC-pogon FP4/FP8, energijsko učinkovita robna umetna inteligenca

Projekti

• Financer: ARIS - Javna agencija za znanstvenoraziskovalno in inovacijsko dejavnost Republike Slovenije
• Številka projekta: P2-0359
• Naslov: Vseprisotno računalništvo

• Financer: MeitY - Ministry of Electronics and Information Technology (MeitY), Government of India
• Številka projekta: Grant EE-9/2/21-R&D-E
• Naslov: Special Manpower Devel-opment Program for Chip to Start-Up (SMDP-C2S)