Singapore Deep-Tech · SEEDER Group Partnership Open

Algorithmic Efficiency
for AI Inference

ANNIE — Algorithmic Neural Network Inference Engine — is a software layer that reduces AI inference compute energy by 60–80% without hardware changes. We are opening a joint research programme with the NUS ECE SEEDER Group targeting ACML 2026 and Singapore's SS 715:2025 green data centre standard.

60–80%
Compute Reduction
4
Core Techniques
0
Hardware Changes
Jun 26
ACML Deadline

No peer-reviewed methodology exists
for algorithmic inference efficiency

Singapore's SS 715:2025 Data Centre Standard mandates energy efficiency but defines no measurement framework for the IT layer — the algorithms running inside the hardware. IMDA and EDB cannot currently compare or validate competing efficiency claims from data centre operators. A joint ANNIE Systems–NUS ECE SEEDER Group paper fills that gap.

CONTEXT 01

SS 715:2025 Mandate

Singapore's green data centre roadmap requires measurable energy efficiency improvements. The PUE metric captures infrastructure efficiency but ignores algorithmic efficiency entirely.

CONTEXT 02

No IT-Layer Standard

No peer-reviewed benchmark methodology currently exists for evaluating inference efficiency at the algorithm level. Vendors self-report. Regulators cannot verify.

CONTEXT 03

The Research Opportunity

A joint NUS ECE SEEDER Group–ANNIE Systems paper proposing and validating a benchmark methodology would position Singapore as the origin of the global inference efficiency standard.

Four coordinated techniques.
One efficiency layer.

ANNIE is a software layer that wraps existing transformer models and replaces computationally expensive operations with mathematically equivalent alternatives. The techniques are independently validated and compound in their efficiency effect.

TECHNIQUE BSW

Bit-Shift Weight Quantization

Replaces floating-point multiply-accumulate operations with integer bit-shift arithmetic. Eliminates the primary source of GPU energy consumption in weight-matrix computation.

TECHNIQUE CRI

Centered Recursive Interpolation

Replaces the exponential function in softmax attention with an integer approximation. Removes transcendental function overhead from the attention mechanism.

TECHNIQUE PwL

Piecewise-Linear Activation

Replaces the GELU activation function — which requires erf() evaluation — with a piecewise-linear approximation using only integer slopes and thresholds.

TECHNIQUE CSC

Column-Sparse Attention Routing

Routes attention computation only through non-negligible attention scores. Reduces attention complexity from O(n²) to O(nnz) for sparse attention patterns.

Software-Only. Hardware-Agnostic.

ANNIE wraps any HuggingFace-compatible transformer model in a single function call. No hardware procurement, no infrastructure changes, no retraining. The efficiency gains are immediately measurable on existing H100, A100, or any CUDA-capable GPU — making H100 benchmark data on AWS Singapore immediately actionable for SS 715:2025 compliance reporting.

Joint programme with SEEDER Group (NUS ECE). AWS Credits. ACML 2026.

We are proposing a two-phase research collaboration with Asst. Prof. Kelvin Fong's SEEDER Group (Semiconductor & Emerging Electronic Devices Exploratory Research) at NUS ECE. Phase 1 targets ACML 2026. Phase 2 extends into hardware-algorithm co-design for the Digacus-Nano silicon programme.

P1

Phase 1 — ACML 2026 Paper (Deadline: 26 Jun)

Joint ANNIE Systems–SEEDER Group benchmark paper establishing a peer-reviewed methodology for measuring algorithmic inference efficiency on H100 infrastructure in AWS Singapore (ap-southeast-1). SEEDER researchers named as co-authors.

$$

AWS Research Credits — Zero Cost to SEEDER

We are applying for AWS Research Credits with Asst. Prof. Kelvin Fong (SEEDER Group) as co-PI. All H100 compute costs on AWS Singapore (ap-southeast-1) are covered by the credits — zero cost to NUS ECE or the SEEDER group.

P2

Phase 2 — Hardware-Algorithm Co-Design (2027)

Extending into Digacus-Nano ASIC co-design on GlobalFoundries 22FDX (Fab 3E, Woodlands, Singapore) — a custom integer compute array purpose-built for ANNIE inference. SEEDER's device/circuit/algorithm co-design expertise maps directly to this phase. Joint NRF/EDB IAF grant application.

IP

IP Framework

ANNIE Systems retains background IP. Jointly developed research results are co-owned with NUS under a standard university-industry IP agreement. SEEDER Group researcher co-authorship on all publications.

What We Bring

Working implementation of all four ANNIE techniques on GPT-2 and LLaMA-3 architectures. Benchmark framework measuring tokens/sec, active watts, and per-technique compute reduction. Full technical documentation available under NDA.

What We Need

A named SEEDER Group / NUS ECE co-PI for the AWS Research Credits application. Hardware measurement methodology expertise. SEEDER graduate researcher involvement in the benchmark design and paper authorship.

The Team

ANNIE Systems is a Singapore-incorporated deep-tech company founded at the intersection of algorithmic research and sovereign AI infrastructure.

A
Arnold Alagar
Co-Founder & CTO · Primary Inventor
Architect of the ANNIE efficiency methodology and primary inventor of the CIP patent claims. Background in sovereign-grade AI infrastructure and high-performance systems engineering across ASEAN and the US.
S
Steve Shattil, Ph.D.
Chief Scientist
Inventor of US Patent 11,640,522 B2. Deep specialisation in computational mathematics, signal processing, and energy-efficient neural network arithmetic. Track record of battle-tested, hyper-scale patent portfolios.

Connect with ANNIE Systems

Directed to the NUS ECE SEEDER Group (Asst. Prof. Kelvin Fong). We also welcome enquiries from NTU, SUTD, and A*STAR research groups working on energy-efficient computing and hardware-algorithm co-design. Technical documentation available under NDA.

arnold@annie.systems