Majestic Labs Software Development Planning

HW & Emulations Delivery Firmware System & Core AI Kernels Compiler & Toolchain Model Zoo
6
HW & Emulations Delivery · 0 done
1
Firmware · 0 done
30
System & Core · 0 done
21
AI Kernels · 0 done
40
Compiler & Toolchain · 0 done
1
Model Zoo · 0 done
99
Total items
Vertical / MonthJun 26v0.5
Jul 26v0.6
Aug 26v0.7
Sep 26v0.8
Oct 26v0.9
Nov 26v0.95
Dec 26v1.0
Jan 27v1.1
Feb 27v1.2
Mar 27v1.3
Apr 27v1.4
May 27v1.5
Jun 27v1.6
Jul 27v1.7
Aug 27v1.8
Sep 27v1.9
Oct 27v2.0
Nov 27v2.1
Dec 27v2.3
HW & Emulations Delivery
+ add
Palladium▾
+ add
+ add		+ add+ add+ add
Tape Out▾
+ add		+ add+ add+ add+ add+ add
Evaluation Boards▾
+ add		+ add
Full Majestic System in our lab▾
+ add		+ add+ add+ add
Customer Early Access▾
+ add
Firmware
SCP boot flow closed▾
+ add		+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add
System & Core
E2: Core Graph Execution▾
+ add
E2: XDMA Engine▾
+ add
+ add
+ add
E2: ARM QEMU Env▾
MDD: API ready▾
+ add
E2: Kernel driver for device access▾
MDD: Initial features (?)▾
+ add
MDD: User Space mgmt Tool▾
+ add
+ add
+ add
MDD: MMU (device side)▾
+ add
E2: Graph Optimizations▾
+ add
E2: Fault tolerance▾
+ add
+ add
E2: Data digestion support▾
+ add
E2: Advanced Scheduling▾
+ add
E2: Multi-Chip / Multi-E2▾
+ add		+ add+ add+ add
AI Kernels
+ add
Func: GEMM BF16; Softmax▾
Perf: GEMM FP32; RMSNorm▾
Precision: exp; reciprocal▾
+ add
Func: FlashAttention; GEMM FP8, FP4 with Blocked Scaling▾
Perf: GEMM BF16, FP8; Softmax▾
SystemC: Performance Regression Suite▾
+ add
+ add
+ add
+ add
+ add
P0.0 perf: rng; pointwise; Reduce; RoPE▴
Conv3D; Conv1D (dilated); ConvTranspose1D; interpolate▴
+ add
+ add
+ add
+ add
+ add
+ add
+ add
+ add
+ add
+ add
+ add
+ add
Compiler & Toolchain
Triton: Raise grid / scratchpad caps▴
+ add
+ add
+ add
+ add
+ add
+ add
+ add
+ add
Triton: Sliding-Window Attention (window=512)▾
Triton: FP8 KV-cache load/store kernels▾
+ add
Triton: MoE top-8 routing + expert dispatch▾
+ add
+ add
Triton: Flash-decode 131k context▾
+ add		+ add+ add+ add+ add+ add+ add+ add
Model Zoo
Laguna - Functional▾
+ add		+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add+ add
SystemC MLAMP→ Aug 26▾
CL: Simple host to ML1 communication layer, replace current TCP implementation→ Sep 26▾
MOS: Buffer Allocation & Kernel Metadata→ Sep 26▾
MDD: Planning→ Sep 26▾
E2: Buffer Allocation & Kernel Metadata→ Aug 26▾
E2: SOA, Boot & Core Graph Execution→ Sep 26▾
CL: Finalize host to ML1 communication layer→ Dec 26▾
E2: Advanced Scheduling (1)→ Dec 26▾
E2: Advanced Scheduling (2 - NUMA)→ Feb 27▾
CL: Multiple ML1 connection, trivial orchestration (round robin)→ Mar 27▾
MOS: MMU (device side)→ Mar 27▾
MDD: Graph Execution Feature Complete→ Feb 27▾
E2: MMU (device side)→ Mar 27▾
CL: Multiple ML1 orchestration (?)→ Jun 27▾
MOS: Fault tolerance (?)→ Jun 27▾
MDD: Fault Tolerance→ Jun 27▾
E2: Debugging→ Jul 27▾
MDD: Data Digestion Support→ Sep 27▾
Interface Definition for Graph Compiler→ Jul 26▾
GEMM Fused-Epilogue (bias + act); Activation; Pointwise (Add/Op/Scale/Set); Reduce; LayerNorm→ Oct 26▴
Fused rms_norm; layer_norm + residual; RoPE (rotary); fused_dense / MLP→ Oct 26▴
mha_varlen_fwd; mha_fwd_kvcache (paged-KV decode); FA2 feature axes; SDPA_attributes→ Oct 26▴
rng: Philox core, state mgmt, uniform draw, normal draw→ Oct 26▴
P0.0 perf: attention (mha_fwd / varlen / kvcache); GEMM fused-epilogue→ Jan 27▾
Conv2D forward (3×3, 1×1); GroupNorm→ Jan 27▾
Kernel Infra Performance→ Mar 27▾
Pooling; BatchNorm; FFT (1D complex, 1D real, STFT)→ Mar 27▾
P0.0 Performance Cleanup→ Mar 27▾
P1 perf: Conv2D / 3D / 1D; interpolate; FFT→ May 27▾
P1 perf: Pooling, BatchNorm, GroupNorm, ConvTranspose1D; final P0.0 + P1 performance pass→ May 27▾
Robustness / Integration / Stress Testing→ Dec 27▾
FE: Slicing Algorithm→ Aug 26▾
FE: Cluster Algorithm→ Sep 26▾
FE: DMA Operator Support→ Jul 26▾
FE: MCMA Support→ Oct 26▾
FE: factory-op rewrite productionised→ Jul 26▴
FE: MCMA wrapper as Inductor codegen target→ Aug 26▾
Triton: Adopt TritonGPU dialect→ Jul 26▾
Triton: Port majestic codegen onto TritonGPU IR→ Aug 26▴
Triton: Link riscv64 libTritonMajesticRuntime→ Jul 26▴
Triton: Atomic-op codegen to XM via GW→ Jul 26▴
FE: Cluster Algorithm→ Sep 26▾
BE: FX → MLIR Translation→ Jul 26▾
BE: Cluster→ Aug 26▾
BE: MajesticRunner→ Jul 26▾
BE: MCMA Lowering & Generation→ Sep 26▾
FE: Inductor auto-generated code for Majestic→ Nov 26▾
FE: ATen native dispatch + device register→ Sep 26▾
BE: XDMA Function Support (in / out kernel)→ Sep 26▾
BE: Memory Materialisation & Bufferisation→ Oct 26▾
FE: ATen factory + elementwise→ Oct 26▾
Triton: Tile / warp / num_stages → CMA shape mapping→ Sep 26▾
Triton: Validate 88-op set vs TritonCPU baseline→ Sep 26▾
Triton: Tile-size sweeps for top-8 hot ops→ Oct 26▾
Triton: FP8 / BF16 / INT4 dtype lowering→ Oct 26▾
BE: Kernel Lowering — AOT→ Nov 26▾
Triton: Decommission TritonCPU path→ Nov 26▾
BE: Kernel Lowering — JIT (EW · Matmul · Reduction)→ Feb 27▾
FE: ATen Linear/SDPA hand-tuned→ Feb 27▾
BE: LLVM Backend Optimisation (RISC-V codegen)→ Mar 27▾
FE: PTMage cost-model on FXIR→ Feb 27▾
Triton: Block-pointer + swizzling in GEMM→ Feb 27▾
Triton: Triton → Cortex MLIR with CMA hints→ Mar 27▾
Triton: Persistent-kernel rewrite (Linear/SDPA)→ Feb 27▾
BE: Full CMA Catalog + PTMage Cost Model→ Feb 27▾
Triton: SW-pipeline GM→XM xDMA with compute→ Apr 27▾

Appendix A — AI Kernels: Full Operator Classification by Library

Source: kernel-roadmap §2 · Priority 0 = LLM must-have → 5 = classical sci-compute · — = not supported

Summary — Operator counts by library and priority

LibraryP0P1P2P3P4P5Not supportedTotal
blas — cuBLAS / rocBLAS70000264073
dnn — cuDNN / MIOpen690060425
flash-attention — FlashAttention v2/v3 + cuDNN fMHA1700030121
thrust-cub — Thrust + CUB0034000135
sparse — cuSPARSE / rocSPARSE00014001024
graph-analytics — cuGraph + cugraph-ops / pylibcugraphops0006300366
linalg-solvers — cuSOLVER / hipSOLVER00000273966
rng — cuRAND / hipRAND1500004928
fft — cuFFT / rocFFT0120003520
dataframes — cuDF0006700168

Priority legend: P0 LLM inference must-have   P1 Diffusion / video / audio   P2 Thrust/CUB parallel primitives   P3 GNN / tabular NN   P4 Training-only   P5 Classical sci-compute   — No plan to support