The world's first predictive latency management layer for gaming PCs.
You have a 7800 XT. You're hitting 144+ FPS. But your game still stutters every 30 seconds.
NeuroPace fixes that.
Your GPU isn't slow. Windows is in the way.
Every second, the OS processes thousands of Deferred Procedure Calls (DPC) and hardware interrupts. When these pile up — even for 2 milliseconds — your perfectly rendered frame gets delayed. Your mouse input arrives late. Your GPU clock drops 400MHz because the thermal controller overreacted to a tiny temperature spike.
Current solutions react after the damage is done. NeuroPace predicts the problem before it reaches your screen.
| Metric | Before | After | Change |
|---|---|---|---|
| 1% Low FPS (CS2, 1080p) | 138 fps | 182 fps | +32% |
| Frametime StdDev (Warzone) | 4.2ms | 2.8ms | -33% |
| Input Latency (Valorant) | 12.4ms | 8.1ms | -35% |
| Thermal Throttle Events/hr | ~14 drops | ~2 drops | -86% |
| Clock Speed Variance | +/- 400MHz | +/- 50MHz | 8x more stable |
| CPU Overhead | — | < 0.5% | Virtually zero |
Benchmarked on AMD Radeon RX 7800 XT | Ryzen 7 7800X3D | 32GB DDR5 | Windows 11 23H2 Measured with PresentMon and CapFrameX. Full methodology in
docs/benchmarks.md.
Every optimizer on the market today is reactive — it sees a stutter, then tries to fix it. By then, the damage is already on your screen.
NeuroPace is predictive. It continuously analyzes rolling windows of 1,000+ telemetry samples — DPC latency patterns, interrupt density, GPU queue depth, frametime variance, CPU core state, VRAM pressure — and calculates the probability of the next frame stuttering.
When risk is detected:
- Thread migration to isolated cores
- GPU scheduling priority elevation
- Interrupt queue rebalancing
- Frame pacing curve smoothing
The stutter never reaches your monitor.
Current Frame → Analyze 1000-sample window → Predict next frametime
↓
Risk > threshold?
↓
Preemptive mitigation (< 5ms)
Why this matters: No consumer tool predicts future frametimes. This is the shift from reactive optimization to predictive stability.
Windows treats all system interrupts equally. Your GPU interrupt, mouse polling, network driver, and background Windows Update service all compete for the same attention.
NeuroPace builds a real-time interrupt priority map:
| Source | Priority | Strategy |
|---|---|---|
| GPU Render Queue | Critical | Dedicated core, zero preemption |
| Mouse / Keyboard | Ultra-Critical | Lowest possible latency path |
| Network Stack | Adaptive | Scaled by game state |
| Telemetry / Logging | Deferrable | Batched during idle frames |
| Background Services | Low | Migrated to efficiency cores |
The result: Software-Defined Latency Management — a concept that doesn't exist in consumer software today.
Standard GPU tools see three things: FPS, utilization percentage, and temperature. But your RDNA GPU produces dozens of hardware signals that never reach the OS:
- Command processor stall events
- Wavefront occupancy levels
- Cache hit/miss ratios
- Render queue saturation
- Shader pipeline bubbles
NeuroPace fuses hardware-level GPU telemetry (via AMD ADLX) with OS kernel-level data (via ETW) into a single closed-loop feedback system:
RDNA Hardware Sensors Windows Kernel Events
(ADLX SDK) (ETW API)
↓ ↓
┌────────────────────────────────────────┐
│ Telemetry Fusion Engine │
│ 100Hz polling · 28-feature vector │
└────────────────────┬───────────────────┘
↓
Prediction Engine (ONNX)
Sub-millisecond inference
↓
Hardware Actuator (Ring-3)
Thread · Power · Priority
Your PC doesn't just run your game. It learns how to run it better.
This is not a simulation. These are real sensor readings from a physical AMD Radeon RX 7800 XT:
[ADLX] AMD Radeon RX 7800 XT initialized for telemetry.
[AGG] Aggregator started - telemetry: 10ms, dashboard: 33ms
[# 200] GPU: 30MHz 45C 23W | VRAM: 5366/16368MB | Clients: T:1 D:1
[# 400] GPU: 25MHz 45C 24W | VRAM: 5358/16368MB | Clients: T:1 D:1
[# 600] GPU: 5MHz 45C 18W | VRAM: 5355/16368MB | Clients: T:1 D:1
- AMD Radeon RX 7900 XTX / 7900 XT / 7900 GRE
- AMD Radeon RX 7800 XT / 7700 XT
- AMD Radeon RX 7600 XT / 7600
- Any GPU supporting AMD ADLX SDK
Four decoupled microservices. Zero game memory access. Named Pipe IPC with < 1ms transport latency.
| Module | Language | Role |
|---|---|---|
| Telemetry Collector | C++ | ETW kernel events + ADLX hardware polling at 100Hz |
| Prediction Engine | Python/ONNX | 28-feature inference, sub-ms latency, rolling window analysis |
| Hardware Actuator | C++ | Thread affinity, TGP limits, priority optimization via Win32 API |
| Control Center | Web | Real-time dashboard, module orchestration, audit logging |
NeuroPace will never trigger anti-cheat systems:
- Zero game memory access — never reads or writes to any game process
- Zero kernel drivers — purely Ring-3 (User-Mode), no
.sysfiles - Zero code injection — no
WriteProcessMemory, noCreateRemoteThread, no DLL injection - Zero restricted handles — static validation against
VM_READ/GET_CONTEXT - Full audit trail — every decision logged with timestamps
Compatible with EasyAntiCheat, Vanguard, BattlEye, FACEIT AC, and all major platforms.
See: docs/anti-cheat-compliance.md
Download the latest release from the Releases page.
NeuroPace-RDNA-v0.1.0-Win64.zip
├── bin/
│ ├── neuropace-telemetry.exe
│ └── neuropace-actuator.exe
├── scripts/ # Prediction engine
├── dashboard/ # Web Dashboard (Vue/Node.js)
├── models/ # Pre-trained ONNX model
├── NeuroPace.exe # Smart System Tray Launcher
└── README.md
Run NeuroPace.exe. It will seamlessly start the Node.js backend and open the web dashboard in your browser.
git clone https://github.com/l1ve709/NeuroPace-RDNA.git
cd NeuroPace-RDNA
# Bootstrap vcpkg
git clone https://github.com/microsoft/vcpkg.git external/vcpkg
.\external\vcpkg\bootstrap-vcpkg.bat
# Build Telemetry
cmake -S telemetry -B telemetry/build ^
-DNEUROPACE_USE_ADLX=ON ^
-DADLX_SDK_DIR="C:/path/to/ADLX" ^
-DCMAKE_TOOLCHAIN_FILE="external/vcpkg/scripts/buildsystems/vcpkg.cmake"
cmake --build telemetry/build --config Release
# Build Actuator
cmake -S actuator -B actuator/build ^
-DCMAKE_TOOLCHAIN_FILE="external/vcpkg/scripts/buildsystems/vcpkg.cmake"
cmake --build actuator/build --config Release
# Python dependencies
pip install -r ai-engine/requirements.txt# Terminal 1 — System Tray Web Launcher
.\releases\NeuroPace-RDNA-v0.1.0\NeuroPace.exe
# Terminal 2 — Prediction Engine
python ai-engine/src/main.py
# Terminal 3 — Actuator (target your game's PID)
.\actuator\build\Release\neuropace-actuator.exe --pid <game_pid>| Spec | Value |
|---|---|
| Telemetry Rate | 100 frames/sec (10ms resolution) |
| Feature Vector | 28 dimensions per inference |
| Prediction Latency | < 1ms per cycle |
| Action Response | < 5ms detection-to-mitigation |
| CPU Overhead | < 0.5% single core |
| Memory Footprint | ~45MB resident |
| IPC Transport | Named Pipes, JSON serialization |
NeuroPace-RDNA/
├── telemetry/ # C++ Sensor & ETW Collector
│ ├── include/ # telemetry_data, etw_collector, adlx_sensor
│ └── src/ # Implementation
├── ai-engine/ # Python Prediction Engine
│ ├── src/ # predictor, feature_engineer, ipc
│ └── models/ # Pre-trained ONNX models
├── actuator/ # C++ Hardware Actuator
├── dashboard/ # Web Control Center
├── shared/ # Cross-module protocol definitions
├── scripts/ # Utility scripts
└── docs/ # Documentation, benchmarks, compliance
Will this get me banned? No. NeuroPace never touches game memory. It only reads OS-level metrics and adjusts system parameters. See Anti-Cheat Compliance.
Does it work with NVIDIA? Not yet. NeuroPace is built for AMD RDNA using the ADLX SDK. NVIDIA support would require NVAPI/NVML integration.
How much CPU does it use? Less than 0.5% of a single core. Lock-free data structures and hardware-accelerated ONNX inference.
Do I need Admin rights? Recommended. Telemetry needs Admin for ETW kernel access (DPC/ISR). GPU metrics via ADLX work without elevation.
NeuroPace RDNA
Predictive latency management for the competitive edge.
Copyright (c) 2026 Ediz Sonmez. All rights reserved.
