INTEGIZER
Integer Zero-Loss Encoding & Refactoring
Neural Network Weight Refactoring Between Floating-Point and Integer Domains
📋 Read Technical Abstract
The Method That Unlocks Integer AI
»» Jump to Economics and Business Value Proposition »»The Core Insight: The world's most powerful AI hardware already runs integer operations faster than floating-point. NVIDIA H100 Tensor Cores are specified at roughly double the peak dense throughput for INT8 against FP16 — 3,958 TOPS against 1,979 TFLOPS. That is a silicon specification; measured production deployments cluster lower, at 1.4–1.8×. Apple's Neural Engine, Qualcomm's Hexagon DSP, Google's TPUs, and Intel's AMX extensions are all integer-first. Billions of ARM Cortex-M microcontrollers deployed in IoT devices are pure integer processors with no floating-point unit at all. The silicon is ready — what has been missing is a structured method to bring trained neural network intelligence into the integer domain. The INTEGIZER is that method.
Key Innovation: Unlike conventional quantization which permanently degrades model weights, the INTEGIZER's encode-and-scale operation with metadata-tracked exponents enables potentially lossless conversion. The full number sequence is preserved where the destination hardware range accommodates it. Refactored weights can be deployed on alternative architectures, importing intelligent behavior — the patterns, associations, and decision boundaries learned through training — from one computational domain to another.
📥 Input Floating-Point Weights (e.g. 12.345678)
↓
⚙️ Encode & Scale: 12.345678 × 10⁶ = 12345678 (INT64)
↓ ↘
🔢 Integer Computation Engine 💾 Store Exponent(s) as Metadata
┌─────┬─────┬─────┬─────┐
│ ALU │ Bit-Shift │ LUTs │ FP Fallback │
└─────┴─────┴─────┴─────┘
↓ ↓
🏭 High-End Scale-Up 📱 Low-End Minimal-Scale
Patent Figure 64: Integer computation module and deployment paths
🔄 Potentially Lossless Conversion
Where the destination integer range accommodates the full number sequence, the encode-and-scale operation preserves every significant digit, and the stored scale exponents enable exact restoration of the original floating-point values from integer results and vice versa. Where the destination range is narrower, fewer digits survive and the round trip is no longer exact — in proportion to how much of the sequence the container cannot hold. Both cases worked through below →
⚡ Higher Throughput Per Watt
Current GPU silicon is specified at roughly double the peak dense throughput for integer against floating-point, and published production measurements of 8-bit inference deliver 1.4–1.8× against an FP16 baseline. The INTEGIZER produces the refactored weight database that exploits it, and re-targets that same database to other architectures without going back to the floating-point master. See the full measured distribution →
🧮 Integer Computation Engine
Performs inference using an integer arithmetic logic unit with pre-calculated lookup tables for non-linear functions (softmax, sigmoid, tanh), eliminating runtime floating-point calculation overhead entirely.
🛡️ Floating-Point Fallback
Maintains a floating-point fallback capability for values outside lookup table range, ensuring mathematical completeness without sacrificing integer-first performance for the vast majority of operations.
📱 IoT & Edge Deployment
Enables deployment of trained SENN intelligence onto billions of ARM Cortex-M microcontrollers and other integer-only processors that have no floating-point unit — bringing AI to devices that could never run it before.
🏭 Data Center Economics
Accelerates AI data center expansion by enabling inference on cheaper integer chipsets — compressing build timelines, lifting scale within a fixed power envelope, and cutting cost across power, cooling and hardware.
The Problem: Floating-Point Bottleneck
Current limitations in the art:
- Expensive Hardware: Floating-point processing units consume significantly more silicon area, power, and cost than integer ALUs
- Wasted Precision: Float64 carries ~18 decimal digits of precision — much of which is noise, not signal, in trained neural network weights
- Destructive Quantization: Existing quantization methods permanently degrade model weights with no path to restoration
- IoT Exclusion: Billions of deployed integer-only processors cannot run any neural network inference today
- Data Center Costs: AI inference at scale consumes enormous power and requires expensive GPU/TPU hardware
- Missing Pipeline: No structured method exists to convert, deploy, and optionally restore neural network weights across computational domains
| Capability | Conventional Quantization | INTEGIZER |
|---|---|---|
| Conversion Direction | 🔶 One-way (FP → INT only) | ✅ Bidirectional (FP ↔ INT) |
| Weight Preservation | ❌ Permanently Degraded | ✅ Potentially Lossless |
| Scale Exponent Tracking | ❌ Not Preserved | ✅ Stored as Metadata |
| Original Value Restoration | ❌ Not Possible | ✅ Full Restoration via Stored Exponents |
| Non-Linear Functions | 🔶 Runtime FP Calculation | ✅ Pre-Calculated Lookup Tables |
| FP Fallback Capability | ❌ No Structured Fallback | ✅ Built-in for Out-of-Range Values |
| IoT / No-FPU Deployment | ❌ Limited / Ad Hoc | ✅ Structured Pipeline |
| Cross-Architecture Portability | 🔶 Architecture-Specific | ✅ Refactored Weight Database |
| Intelligence Transfer | 🔶 Same Model, Degraded | ✅ Imports Intelligent Behavior to New Architecture |
What "Potentially Lossless" Means — Two Worked Cases
Case 1 — the destination range accommodates the full sequence
Encode × 10⁵ 1,234,567,890 → held comfortably in INT32 or INT64
Exponent stored 5 → retained as metadata, travels with the weights
Decode ÷ 10⁵ 12345.67890 → exact, in both directions
All ten significant digits fit inside the destination container. Nothing is lost on the way in, and the reverse-scale operation returns the original value exactly. This is the lossless case.
Case 2 — the destination range is too narrow
Encode to fit 12,345,678 → only eight digits survive the container
Exponent stored 3 → adjusted to match the digits retained
Decode ÷ 10³ 12345.678 → the final two digits are gone
Whether the surviving digits are simply truncated, or rounded up or down at the .5 mark, is an implementation convention chosen for the deployment. Either way the round trip is not exact. The narrower the destination container, the fewer significant digits survive, and the larger the departure from the original value.
The conversion is lossless where the destination integer range accommodates the full number sequence, and lossy in proportion to how much of that sequence the destination cannot hold. Whether a given deployment is one or the other is decided by the target chipset's available range, not by the method.
A separate question: does the model still give the same answers?
The two cases above concern numeric fidelity — whether a number survives the round trip. That is arithmetic, and it is decidable. It is not the same question as whether a quantised network produces the same outputs, which is a statistical property of the model as a whole.
For deployment-grade 8-bit inference the answer is no, not exactly. INT8 is a narrow container — 256 representable values — so any INT8 deployment is Case 2 by construction, and the measured consequence at model level is roughly 1 to 3 percent accuracy degradation for well-tuned INT8 weight-and-activation quantisation, with some model families more sensitive than others and per-model validation required. (arXiv 2411.02355, over 500,000 evaluations; LLaMA3-70B exception per arXiv 2408.15301.)
What the INTEGIZER changes is the nature of that trade. The retained scale exponent makes it recorded and reversible rather than silent and permanent, and the same refactored weight database can be re-targeted to another destination architecture without returning to the floating-point master. The trade-off is not eliminated — it is put under control. Case 1 is where the wide-container, high-end scale-up path in Figure 64 lives; Case 2 is where the low-end minimal-scale path lives. Both are settings of one method.
INTEGIZER Technical Architecture (Figure 64)
- Receive Trained Weights: Accept neuron weights encoding intelligent behavior — patterns, associations, and decision boundaries learned through training
- Encode & Scale: Multiply weights by deterministic scaling factor (e.g., 12.345678 × 10⁶ = 12345678 as INT64) to shift significant digits into target integer range
- Store Exponents: Preserve scale exponents as metadata, enabling exact reverse-scale restoration
- Integer Computation: Execute inference via integer ALU (addition, subtraction, multiplication — no FPU required), bit-shifting for power-of-2 multiplications, and pre-calculated lookup tables for non-linear activation functions
- FP Fallback: Engage floating-point fallback processor where available for values outside lookup table range
- Deploy: Assemble refactored weight database and deploy on destination architecture, importing intelligent behavior from source system
Quantization Spectrum (Figure 65)
Float64 / Float32
Full Precision → Moderate Precision. Industry-standard training formats. Float64 carries ~18 decimal digits (much is noise); Float32 carries ~7 digits. The INTEGIZER can convert from either into integer representations while preserving significant information.
INT16 / INT8 / INT4
Reduced to Low Noise. 65,536 / 256 / 16 possible values respectively. Fewer values force resolution — coarser but more decisive weight assignments. INT8 is already the native fast path on current GPU Tensor Cores, TPUs, and mobile neural engines.
Ternary {-1, 0, +1}
Maximum Compression. Only three states per weight. Computation may be bypassed entirely — replaced by simple sign-conditional addition. Minimal noise. Enables AI on the most resource-constrained devices imaginable.
The Hardware Is Already Waiting
The INTEGIZER method enables AI deployment across hardware that already exists — from data centers to wristbands. The silicon can already do massively parallel integer computation.
🖥️ GPU Tensor Cores
NVIDIA H100, A100, and successors are specified at ~2× peak dense INT8 throughput against FP16 on identical silicon (measured production gains are lower — see the business case)
📱 Mobile Neural Engines
Apple Neural Engine (INT8), Qualcomm Hexagon DSP — shipping in billions of phones and tablets
🔧 Intel AMX / TPU
Intel AMX INT8 matrix ops, Google TPUs in INT8 inference mode — server-class integer acceleration
🌐 IoT / ARM Cortex-M
Billions of pure-integer microcontrollers deployed worldwide — no FPU, but the INTEGIZER brings AI to them
🏭 Data Centers
Cheaper integer chipsets could accelerate AI infrastructure expansion with lower power, cooling, and hardware costs
🔋 Edge & Embedded
PLC controllers, industrial sensors, automotive ECUs, medical devices — anywhere compute budget is limited
Market Opportunity
The summary cards below give the shape of the opportunity. For the detailed economic argument — the power constraint, a fully sourced 100 MW worked model, where the measured throughput figures come from, what 8-bit inference costs in accuracy, and honest answers to the FP8/FP4 objection — see the INTEGIZER business case →
💰 Data Center Cost Reduction
Current GPU silicon is specified at 2× peak dense INT8 throughput against FP16, but peak silicon is not deployment performance. Published production measurements of 8-bit inference cluster at 1.4–1.8× against an FP16 baseline. At a conservative 1.5×, a power-capped facility yields roughly 50% more serving capacity from the same grid connection — which matters most where megawatts, not capital, are the binding constraint. Full worked model →
🌍 IoT AI Enablement
Billions of integer-only ARM Cortex-M devices currently cannot run any neural network. The INTEGIZER opens an entirely new market for on-device AI in industrial, medical, agricultural, and consumer IoT.
⚡ Power & Sustainability
Integer computation consumes significantly less power than floating-point. At data center scale, the energy savings from INTEGIZER-enabled integer inference contribute meaningfully to sustainability targets.
📈 An Advantage That Compounds
The industry trajectory is toward more integer computation, not less. Every generation of AI hardware adds faster integer support, and the population of integer-capable devices grows every year. The INTEGIZER method becomes more valuable over time.
The Future of AI Compute
The hardware is ready. The silicon already runs integer faster than floating-point. The INTEGIZER is the method that bridges trained AI intelligence to the integer domain — where it runs faster, cheaper, and on hardware that couldn't participate in AI before.
From hyperscale data centers to the smallest IoT sensor, the INTEGIZER enables AI deployment on the hardware the world already has.
Partner with INTEGIZER Development
Read the business case before you write →