This page outlines extensions planned for releases after v0.10.2. These features are under active development or under consideration for future releases; none are part of the v0.10.2 shipped compiler unless explicitly marked otherwise.
Everything here is roadmap for the open compiler unless a chip says otherwise. Some of it (distributed primitives, model serving, GPU backends) already ships in the commercial mind-runtime.
Phase 13: BCI & Neuroscience
Optimizations for brain-computer interface and real-time neural processing:
Ultra-low latency paths
Streaming tensors
Pre-allocated memory pools
Signal processing primitives
@realtime annotation
Distributed Training
Multi-node training support for large models (see Distributed Execution Guide):
Data parallelism
Model parallelism
Pipeline parallelism
Collective communication
Elastic training
Commercial The distributed primitives (TCP transport, NCCL/Gloo backends, RingAllReduce, pipeline parallelism, fault tolerance) belong to the commercial mind-runtime; deterministic GPU and multi-node float execution is roadmap.
Production Deployment
Full-stack deployment infrastructure (see Deployment Guide):
One-command deployment
Containerized serving
A/B testing and canary deployments
Model versioning and rollback
Built-in monitoring
Commercial Model serving infrastructure ships with the commercial mind-runtime; cloud deploy, canary rollouts and model versioning are still roadmap.
Sparse Tensors
First-class support for sparse data:
Sparse tensor types
Sparse-aware autodiff
Optimized sparse-dense operations
Graph neural network primitives
Quantization
Built-in quantization for efficient inference:
INT8/INT4 quantization
Mixed-precision training
Quantization-aware training
Post-training quantization tools
Hardware Targets
- Shipped
- Commercial
- Planned
| Target | Status | Notes |
|---|---|---|
| x86-64 CPU | Stable | AVX2/AVX-512 vectorization |
| ARM64 CPU | Stable | NEON vectorization |
| NVIDIA GPU (CUDA) | Enterprise | Production CUDA 12.8+ backend via Enterprise license; cuBLAS/cuBLASLt/cuDNN, 8-stream pool, caching arena allocator. |
| AMD GPU (ROCm) | Commercial | ROCm backend targeting rocBLAS / hipStream; ships in the commercial mind-runtime. |
| Apple Silicon (Metal) | Commercial | Metal Performance Shaders backend; ships in the commercial mind-runtime. |
| WebGPU | Commercial | Browser and native execution via WGSL shader codegen; ships in the commercial mind-runtime. |
| WebNN (CPU/GPU/NPU) | Commercial | W3C WebNN graph builder with CPU/GPU/NPU device selection; ships in the commercial mind-runtime. |
| Google TPU | Planned | TPU execution backend; roadmap item. |
| On-device NPU | Planned | Dedicated neural accelerator backends (ANE, Hexagon, OpenVINO); roadmap item. |
| Groq LPU (TSP) | Planned | LPU execution backend; roadmap item. |
| DPU (BlueField / Pensando) | Planned | Data processing unit offload backend; roadmap item. |
| FPGA (Versal / Agilex) | Planned | FPGA execution backend via HLS lowering; roadmap item. |
| ASIC | Planned | Custom ASIC execution dialect; roadmap item. |
| Cerebras (WSE-2 / WSE-3) | Planned | Wafer-scale engine backend; roadmap item. |
| Taalas (Hardware Models) | Planned | Hardware model provenance integration; roadmap item. |
| Tenstorrent (Wormhole / Blackhole) | Planned | TT-Metalium / Tensix mesh execution backend; roadmap item. |
| SambaNova (RDU) | Planned | Reconfigurable dataflow unit backend; roadmap item. |
| Graphcore IPU (Bow / Mk2) | Planned | IPU BSP execution backend; roadmap item. |
| Intel Gaudi (2 / 3) | Planned | Gaudi accelerator execution backend; roadmap item. |
Developer Tooling
Language Server Protocol (LSP)
Formatter
mindc fmt (Mindcraft) provides deterministic, idempotent formatting today; see the Roadmap.Debugger
Profiler UI
Learn More
See the full future extensions specification at mind-spec/future-extensions.md and the Roadmap for timeline information.