• 32 Parallel Coder+Judge Agent Pairs: Swarm architecture generates and validates kernels concurrently

• MAP-Elites Evolutionary Optimizer: 36 behavior cells with 4 specialized islands (memory_bound, compute_bound, fused_ops, tensor_cores)

• Pattern RAG System: 1,711 CUTLASS patterns + 113 Triton patterns for context-aware generation

• Up to 5x Speedup: Llama-3.1-8B (5.2x), Qwen2.5-7B (4.2x), Mistral-7B (3.4x), SDXL UNet (2.9x) vs torch.compile

• Multiple Input Types: HuggingFace model IDs, KernelBench tasks (250+), or custom PyTorch files

• Dual Output Formats: Triton (Python GPU kernels) or native CUDA C++ - drop-in PyTorch replacement

• Three Optimization Modes: --turbo (fast ~2min), default (balanced), --quality (maximum optimization)

• Interactive CLI: forge command launches wizard, plus forge browse for KernelBench task browser

• Session Management: Track past optimizations with forge session list

• Credit System: 1 credit per KernelBench/custom kernel, 1-2 credits for HuggingFace models

• 97.6% Correctness Rate: Tiered evaluation pipeline (Dedup → Compile → Test → Benchmark)

• 250k Tokens/Second: Powered by fine-tuned NVIDIA Nemotron 3 Nano 30B for fast generation

• Automatic Tensor Core Optimization: WMMA, TMA for Hopper architecture

• Cross-platform Install: npm, npx, curl (macOS/Linux), PowerShell (Windows)

• PyTorch Kernel Profiling: Profile custom PyTorch kernels with full NCU integration

• PyTorch Benchmarking: Run statistical timing analysis on PyTorch operations

• PyTorch Emulation: Test PyTorch kernels across 86+ GPU architectures without hardware

• Automatic Kernel Detection: Detects PyTorch kernels from your code automatically

• Cross-DSL Comparison: Compare PyTorch kernel performance side-by-side with CUDA/Triton implementations

• Unified Workflow: Same profiling methods (NCU Full, Fast, Static, Line-by-Line) work across all supported languages

• Multi-DSL Platform: CUDA, Triton, TileLang & CUTE support with automatic detection and compilation

• Unified Profiling Experience: Profile Triton and TileLang kernels with NCU just like CUDA kernels

• CUTLASS Auto-Detection: Automatically finds your CUTLASS installation across common locations

• Enhanced Language Features: Full semantic highlighting, hover documentation, and go-to-definition for all DSLs

• Context-Aware Help: AI automatically retrieves relevant GPU documentation based on your code and questions

• GPU-Specific Filtering: Only shows documentation compatible with your GPU architecture

• 100+ Documentation Sources: Comprehensive coverage of CUDA, Triton, TileLang, and CUTE APIs

• DSL-Aware Suggestions: AI understands which DSL you're working with and provides relevant guidance

• GPU Context Display: See your active GPU info directly in the chat

• DSL-Specific Metrics: Extracts Triton warps/stages, TileLang block sizes, CUTE tile dimensions

• Smart Performance Warnings: Get alerts for high warp counts, excessive sync points, or register pressure

• Benchmark Accuracy: Triton benchmarks now show correct DSL columns (BLOCK, Warps, Stages)

• Emulation Reliability: Fixed bug where Triton used real GPU instead of emulator when emulated GPU selected

• GPU Detection: Better architecture identification and compatibility checking

• Chat Interface: Cleaner message rendering with improved code display

• Triton Emulation: Now correctly uses emulation mode when emulated GPU is selected

• Benchmark Columns: Triton and TileLang show proper DSL-specific columns instead of CUDA defaults

• Parameter Handling: DSL parameters now correctly flow through profiling pipeline

• Mac Platform Support: Full macOS compatibility with Metal GPU detection for Apple Silicon (M1/M2/M3) and Intel GPUs

• Multi-GPU Support: Select and profile multiple GPUs simultaneously with side-by-side performance comparison

• GPU Filter Dropdown: Filter profiling results by specific GPU

• Remote GPU Support: Full SSH connection to remote GPU servers with seamless CUDA environment integration

• Dynamic Profiling Methods: 5 profiling options (NCU Full, Fast, Static, Line-by-Line, Kernel Replay) with quick-switch buttons

• Simplified Chat Modes: Streamlined to Agent, Gather, and Forge only (Coming soon)

• Chart Rendering: Better performance visualization for profiling data

• Double Credit Consumption: Fixed bug charging credits twice

• Profiling Data Persistence: No more data loss on view refresh

• SSL Handshake Issues: Fixed connection stability

• Iterate Mode: Removed and consolidated into simplified chat mode system

• New GPU Emulator built from scratch with cycle-accurate scheduling and multi-warp latency simulation

• PTX and SASS translation for deep low-level analysis and debugging

• Remote connection support (SSH + WSL) to run and profile kernels anywhere

• Agentic AI "Iterate Mode" that writes, profiles, and optimizes kernels automatically until peak performance

• Kernel fusion detection for better performance across sequential and parallel operations (Beta Users)

• Enhanced benchmarking and profiling with full metric breakdowns and bottleneck detection

• Local LLMs now work perfectly

• 96-98% emulator accuracy vs real GPUs

• Simulation speed under 100ms for 1,000 instructions

• 30% more accuracy than previous builds

• Remote GPU connection via SSH integration

• Native support for GPU cloud providers (RunPod, Google Cloud, AWS, Azure, Paperspace, Vast.ai, Lambda Labs)

• Seamless profiling on remote GPUs as if they were local

• Automatic GPU detection on remote machines

• Smart Profiling Terminal with AI-powered insights

• Automatic bottleneck detection (memory-bound vs compute-bound)

• NCU-compatible metrics without requiring hardware

• AI-generated optimization suggestions (memory coalescing, bank conflicts, occupancy, branch divergence)

• Fixed NCU GUI integration for report generation

• Enhanced profiling UI with collapsible sections

• WebWorker-based analysis for non-blocking performance

• LRU caching for instant re-analysis

• Improved error handling and fallback mechanisms

• Full GPU emulator for profiling without physical hardware

• 86+ GPU architectures supported

• Static kernel analysis engine (under 100ms)

• Roofline model implementation with ±15% accuracy

• Architecture comparison across multiple GPUs instantly

• Benchmarking Terminal for benchmark sweeps and custom kernel configurations

• Visual comparison between kernels

• Static Profiling with instant CUDA kernel insights in CodeLens

• Real-time registers, shared memory, and occupancy analysis while typing

• Profile with Configs - complete cycle with persistent configs and history

• Tools Detector (nvidia-smi, nsight compute, nvcc)

• Execution time, memory bandwidth, occupancy, SM efficiency, and register usage metrics

• Data size presets, warmup runs, and execution controls

• Grid/block optimization with automatic suggestions

• Multi-GPU support with device-specific benchmarking

• Session management with persistence across restarts

• Sortable results with performance indicators

• CSV export for sharing benchmark results

• OpenAI, Anthropic, Deepseek integration

• Local Ollama and vLLM support

• BYOK (Bring Your Own Key) flexibility

• Fill-in-the-Middle autocomplete

• NVIDIA Nsight Compute integration

• Real-time GPU performance metrics

• Hardware detection and optimization

• CUDA syntax highlighting and IntelliSense