Skip to main content

RPU / RDU (Reconfigurable Dataflow) Architecture

What is RPU / RDU

RPU (Reconfigurable Processing Unit) is Tenstorrent's reconfigurable processing unit, and RDU (Reconfigurable Dataflow Unit) is SambaNova's analogous architecture.

Core concept: Use compiler-driven automatic parallelization to map neural networks onto many small tensor cores, eliminating traditional SIMT scheduling overhead, achieving better performance/watt than traditional GPUs.

Representative products:

  • Tenstorrent Grayskull (2021): 150W entry
  • Tenstorrent Wormhole (2023): 16 cores, 200W
  • Tenstorrent Blackhole (2024): 120 cores, 300W
  • SambaNova SN40L (2023): 1,020 cores, 1,000 TFLOPS

RPU/RDU vs GPU

DimensionRPU/RDUGPU (H100)
Programming modelCompiler auto-parallelCUDA manual optimization
Core structureSmall tensor cores × hundredsTensor Core × 132
Control flowStatic dataflowDynamic SIMT
FlexibilityHigh (each core independently programmable)Medium (warp scheduling)
Performance/wattHigh (compiler optimized)Medium
EcosystemSmall (proprietary SDK)CUDA mature
PriceLowHigh

Tenstorrent Series

Grayskull (e75 / e150)

  • Released 2021-07
  • 120 Tensix cores (each with 5 RISC-V + 1 tensor engine)
  • 150W TDP
  • Primarily for developers

Wormhole (n150 / n300)

  • Released 2023
  • 8 / 16 Tensix cores
  • 200W TDP
  • Edge servers

Blackhole (p100a / p150a)

  • Released 2024
  • 120 Tensix cores
  • 300W TDP
  • Data center class

SambaNova SN40L

  • Released 2023
  • 40 GB HBM / 1,020 RDU cores
  • 600 TFLOPS BF16
  • 150W TDP (unbelievable)
  • Customer: San Diego Supercomputer Center (SDSC)
  • Ideal for single-card 70B+ LLM inference (40GB HBM + RDU efficient execution)

Software Ecosystem

Tenstorrent

  • TT-Buda (compiler, originally acquired Buda)
  • TT-Metalium (low-level SDK)
  • PyTorch / TensorFlow integration
  • Open source ecosystem

SambaNova

  • SambaFlow (compiler)
  • DataScale SN40L system
  • Enterprise LLM inference service

Use Cases

  • Large model inference (SN40L single-card 70B+)
  • ✅ LLM training (Tenstorrent Blackhole multi-card)
  • ✅ Data center energy efficiency optimization (compiler auto-parallel)
  • ✅ Research / HPC (open source RISC-V architecture)
  • ❌ General GPU computing (small ecosystem)
  • ❌ Mainstream deep learning (CUDA performance more mature)

Detailed Product Pages

Tenstorrent (RISC-V Tensor Core RDU)

SambaNova (Dataflow RDU)

Other RPU/RDU Architectures