Skip to content

10x-Engineers/QuantX-Hardware-Aware-Quantization

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 

Repository files navigation

QuantX-Hardware-Aware-Quantization

QuantX helps AI hardware teams evaluate MXFP, BFP, NVFP, and custom quantization formats against real LLM and VLM workloads. Make data-driven numeric format decisions, accelerate validation, and reduce pre-silicon risk.

QuantX

Hardware-Aware Quantization Platform for AI Inference Silicon

Inference IP Architects · Numeric Format Exploration · Hardware Validation

🚀 Request Evaluation · 🌐 Website · 📬 Contact


Overview

QuantX is a hardware-aware quantization platform built specifically for teams designing AI inference silicon.

Most quantization frameworks help machine learning engineers deploy models onto existing hardware. QuantX addresses a fundamentally different problem:

Which numeric format should be implemented in silicon before tape-out?

QuantX enables architecture teams to evaluate custom numeric formats against production-scale Large Language Models (LLMs) and Vision Language Models (VLMs), measure accuracy degradation before hardware implementation, and establish a software golden reference for RTL and silicon validation.

Built by 10xEngineers for AI accelerator companies, RISC-V processor vendors, inference IP teams, and custom AI hardware startups.


⚡ At a Glance

Target Users AI Inference IP Architects
Primary Use Case Numeric Format Design-Space Exploration
Model Range 1B – 14B Parameters
Supported Formats MXFP · BFP · NVFP · AMXFP · FP16+INT
Optimization Engine Automated Multi-Level Bit Allocation
Validation Flow Golden Reference for RTL & Silicon
Deployment Existing SDK Integration or Custom SDK Development
Roadmap Growth ~1 New Model Added Per Month

🔄 Workflow Comparison

Traditional Hardware Design Flow

Numeric Format Choice
        │
        ▼
RTL Development
        │
        ▼
Verification
        │
        ▼
Tape-Out
        │
        ▼
Run Real LLMs/VLMs
        │
        ▼
Discover Accuracy Loss

❌ Too Late To Change Numeric Format

QuantX Workflow

Select Numeric Format
        │
        ▼
QuantX Simulation
        │
        ▼
Run Real LLMs & VLMs
        │
        ▼
Evaluate Accuracy / Memory / Compression
        │
        ▼
Compare Candidate Formats
        │
        ▼
Make Silicon Decision With Data

✅ Before RTL Freeze

🏗️ QuantX Architecture

HuggingFace Models
        │
        ▼
Calibration Dataset
        │
        ▼
Meta Optimization Engine
        │
        ▼
Feature Selection
        │
        ▼
Quantization Engine  (RTN / GPTQ / GPTAQ)
        │
        ▼
Evaluation Layer  (LM-Eval / LMMS-Eval)
        │
        ▼
Benchmark & Validation Reports

🎯 Core Workflows

🔍 Design-Space Exploration

Evaluate numeric formats before hardware implementation.

  • MXFP exploration
  • BFP exploration
  • NVFP exploration
  • AMXFP exploration
  • Proprietary custom formats
  • Accuracy vs memory tradeoff analysis
  • Pareto frontier generation

🧪 Hardware Validation

Use QuantX as a software golden reference model.

  • RTL verification
  • Datapath validation
  • Silicon bring-up
  • Numerical accuracy closure
  • Regression testing

When RTL output diverges from QuantX output, engineers immediately know where to investigate.

🚀 SDK Deployment

Deploy QuantX-compressed models into production environments.

  • Existing inference SDK integration
  • Custom runtime development
  • Compiler stack integration
  • Hardware-specific deployment optimization

🧩 Numeric Format Support

Format Description
FP16 + INT FP scaling with integer elements
MXFP OCP Microscaling Format
BFP Block Floating Point
NVFP FP8 scale with FP4 elements
AMXFP Dual-scale Microscaling Format

Additional formats can be added through QuantX's modular architecture.


📐 Quantization Features

Rounding Methods

  • RTN (Round-to-Nearest)
  • GPTQ
  • GPTAQ

Outlier Reduction

  • Scaling
  • Clipping
  • Rotation
  • Reordering

Structural Optimization

  • Mixed Precision
  • Tensor Granularity Control
  • Per-Channel Quantization
  • Per-Group Quantization
  • CodeBook Quantization

🧠 Meta-Optimization Engine

The QuantX Meta-Optimization Engine automates what traditionally requires weeks of manual experimentation.

Level 1 — Transformer Block Allocation

Scores transformer blocks based on:

  • Cosine similarity
  • Token similarity
  • Span score
  • Z-score analysis
  • Domain expertise priors

High-importance blocks receive more precision.

Level 2 — Tensor-Level Refinement

Independent scoring of:

  • Query projections
  • Key projections
  • Value projections
  • Output projections
  • FFN Up
  • FFN Gate
  • FFN Down

Level 3 — Format-Aware Feature Selection

Automatically activates compatible algorithms depending on:

  • Numeric format
  • Model architecture
  • Memory constraints
  • Hardware limitations

🤖 Supported Models

Language Models

Llama Family Qwen Family
Llama 2 Qwen 2
Llama 3.1 Qwen 2.5
Llama 3.2 Qwen 3

Vision Language Models

  • Qwen 3 VL
  • Llava-Next 1.6 7B
  • SmolVLM
  • SmolVLM2

Generative & Legacy Models

  • CLIP
  • OPT
  • Stable Diffusion 1.5
  • Stable Diffusion 3.5
  • Stable Diffusion XL

📊 Benchmark Highlights

Model Format Result
Llama 3.2 1B INT + FP5 8% accuracy-gap recovery vs RTN
Llama 3.1 8B BFP Surpasses baseline GSM8K score
Qwen 3 VL 2B INT + FP5 4% accuracy-gap recovery
Qwen 2.5 VL 7B BFP Near-baseline ChartQA accuracy

Across all evaluated configurations, QuantX consistently outperforms vanilla RTN quantization.


⚙️ QuantX + Baltoro

QuantX integrates directly with Baltoro, 10xEngineers' AI compiler stack.

Pretrained Model
       │
       ▼
    QuantX
       │
       ▼
Compressed Model
       │
       ▼
    Baltoro
       │
       ▼
Optimized Machine Code
       │
       ▼
Custom AI Hardware

Together they provide a complete path from model compression to deployment on custom silicon.


💡 Why QuantX?

Traditional Challenge QuantX Solution
Numeric format selected using intuition Data-driven format exploration
Accuracy validated after hardware exists Accuracy validated before tape-out
Weeks of manual quantization sweeps Automated optimization engine
No software golden reference Hardware validation workflow
Limited support for custom formats Native MXFP, BFP, NVFP, AMXFP support

🚀 Getting Started

  1. Schedule a technical discussion with the QuantX team
  2. Share your target hardware architecture
  3. Define candidate numeric formats
  4. Run QuantX exploration on representative workloads
  5. Review accuracy, memory, and implementation tradeoffs
  6. Select the optimal format before silicon commitment

📬 Contact & Access

Website https://10xengineers.ai
Request Evaluation https://10xengineers.ai/contact-us
Email contact@10xengineers.ai
Location San Diego, CA, USA
LinkedIn https://linkedin.com/company/10x-engineers
YouTube https://youtube.com/@10xengineers

🔗 Related Products

Product Description
Baltoro MLIR-based AI compiler stack
Compiler & Toolchain Services Custom compiler development
AI Infrastructure Software End-to-end inference software
RISC-V AI Solutions Hardware-software co-design services

Copyright © 2025 10xEngineers · All Rights Reserved

About

QuantX helps AI hardware teams evaluate MXFP, BFP, NVFP, and custom quantization formats against real LLM and VLM workloads. Make data-driven numeric format decisions, accelerate validation, and reduce pre-silicon risk.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors