open source · apache 2.0 · single-node
HAL Simulator · Lite
Free, open-source simulator for the Adaptive OS hardware abstraction layer. Unified API across CPU, CUDA, OpenVINO NPU and a 24-qubit QPU simulator. Capped at 2 shards on a single leaf — perfect for laptops, Jetson and Pi 5.
backends
4 (CPU·CUDA·OVN·QPU-sim)
max shards
2
fabric
single leaf · 25GbE
price
$0
Tune the workload
Accelerator
Shards: 1 / 2 (lite cap)
Batch: 16
throughput
42.2 TFLOPS
latency p50
1.42 ms
power
304 W
Fabric view
leaf-01 ── cuda #0
└─ 25 GbE · rdma=off · lite profilePer-kernel sharding breakdown
filter kind
WARNoversized per-shard slice (134.0 MB)
Largest kernel slice exceeds 90 MB working set; risk of HBM pressure on NVIDIA CUDA (RTX).
↳ suggested: Enable auto-shard (expert split for MLP) or increase shard count.
| Layer / kernel | Kind | Strategy | Slice (MB) | Placement across shards |
|---|---|---|---|---|
| embed.tok | embed | replicate | 49.0 | s0 |
| block.0.attn | attn | tensor | 67.0 | s0 |
| block.0.mlp | mlp | expert | 134.0 | s0 |
| block.1.attn | attn | tensor | 67.0 | s0 |
| block.1.mlp | mlp | expert | 134.0 | s0 |
| block.2.attn | attn | tensor | 67.0 | s0 |
| block.2.mlp | mlp | expert | 134.0 | s0 |
| block.3.attn | attn | tensor | 67.0 | s0 |
| block.3.mlp | mlp | expert | 134.0 | s0 |
| block.4.attn | attn | tensor | 67.0 | s0 |
| block.4.mlp | mlp | expert | 134.0 | s0 |
| norm.out | norm | pipeline | 8.0 | s0 |
| lm_head | head | replicate | 49.0 | s0 |
replicate · pipeline · tensor · expert — selected per kernel from the model graph. Slices show the per-shard memory after split.
Kernel scheduling timeline
filter kind
t=0.00 msNVIDIA CUDA (RTX) · single leaf · 25GbE · all kernelst=7.80 ms
- 0.00 msbootqhaldiscover 1 shard(s) · NVIDIA CUDA (RTX)
- 0.60 msfabricfabric-monprobe single leaf · 25GbE · rtt=2.5µs
- 1.10 mscompilemodel.compile()lower embed+attn+mlp+norm+head → 13 kernels
- 2.00 msplansharderauto-shard: tensor+expert+pipeline
- 2.60 msplaceschedulerbind embed+head (replicate) → all shards
- 3.00 msplaceschedulerbind attn (tensor) → split across shards
- 3.40 msplaceschedulerbind mlp (expert) → shards
- 3.80 msplaceschedulerbind norm (pipeline) → shard 0
- 4.20 msdispatchshard-00launch attn shard #0
- 4.50 msdispatchshard-00launch mlp shard #0
- 4.80 mscollectivenccl/rcclall-reduce attn 32MB · ring across 1 shards
- 5.40 mscollectivenccl/rcclall-reduce mlp 64MB · ring across 1 shards
- 6.00 msrebalancefabric-monleaf-02 congestion 78% → migrate mlp shard #2
- 6.50 msrebalanceschedulerevict cold KV cache · reclaim 2.1GB
- 6.90 mscollectivenccl/rcclall-gather logits · ECMP path
- 7.40 msemittelemetryp50=lat ✓ · power ✓ · OTLP → Parquet
- 7.80 msdoneruntimestep complete · feedback → policy net
Need TPU / ROCm / real QPU, multi-rack spine fabric, automatic sharding across ≥8 nodes, or NUMA-aware scheduling? Upgrade to the Enterprise simulator.