SambaFlow (SambaNova RDU runtime)
npu
Fields
id
engines/sambaflow-sambanova-rdu-runtimeLanguages
python
License
proprietary
First release
2021
Design goal
compile-once dataflow graphs onto Reconfigurable Dataflow Units (RDU); whole-model residency across a nodevendor runtime for SN40L/SN30 accelerators, not a general-purpose engine
Scheduling
none
Backends
npu
Formats
sambaflow-compiled-pef
Notable features
static dataflow compilation; model is laid out across RDU tiles at compile timebatch size and sequence length pinned per compiled graph
Hardware caveats
runs ONLY on SambaNova DataScale nodes (SN40L RDU); not portablelong compile times; one graph per (model, batch, seqlen) triple
Best for
vendor-benchmark rows where SambaNova publishes aggregate tok/s through its own stack
Avoid for
any cross-engine comparison that expects per-request decode parity with vLLM/llama.cpp
notes
Added 2026-10-14 to give benchmark.engine_id a real record for the Argonne LLM-Inference-Bench SN40L rows (arXiv:2411.00136). SambaFlow is SambaNova's closed-source compiler+runtime stack; the Argonne measurements use it because SN40L has no PyTorch-eager path for LLMs.
Notes
Added 2026-10-14 to give benchmark.engine_id a real record for the Argonne LLM-Inference-Bench SN40L rows (arXiv:2411.00136). SambaFlow is SambaNova's closed-source compiler+runtime stack; the Argonne measurements use it because SN40L has no PyTorch-eager path for LLMs.
Cross-references
This record references no other records — it is a leaf. It is cited by 5.