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
| Dimension | RPU/RDU | GPU (H100) |
|---|---|---|
| Programming model | Compiler auto-parallel | CUDA manual optimization |
| Core structure | Small tensor cores × hundreds | Tensor Core × 132 |
| Control flow | Static dataflow | Dynamic SIMT |
| Flexibility | High (each core independently programmable) | Medium (warp scheduling) |
| Performance/watt | High (compiler optimized) | Medium |
| Ecosystem | Small (proprietary SDK) | CUDA mature |
| Price | Low | High |
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)
- Tenstorrent Overview (Grayskull/Wormhole) - Full specifications
- Tenstorrent Blackhole - 2024 H1 6nm 120 Tensix cores 8GB SRAM 16 BF16 PF cluster Jim Keller architecture
SambaNova (Dataflow RDU)
- SambaNova SN40L - 2024-09 1,020 RDU cores 638 BF16 TF 128GB HBM3
- SambaNova SN50 - 2026 H2 estimated 3nm 256GB HBM3E 1.5 PF
Other RPU/RDU Architectures
- Blaize Xplorer - GSP graph stream 7K TOPS/W 22W 160 TOPS INT8 Edge AI
- Lightelligence (China) - Photonic computing 1.6 PF FP16 100W 16 TOPS/W CAS + Tsinghua
- LightOn Alfred (France) - Photonic computing 1.2 PF FP16 80W 15 TOPS/W Microsoft + OVHcloud
- Esperanto ET-SoC-1 - RISC-V heterogeneous 1332 cores 42 TOPS INT8 75W founded by Dave Patterson
Related Architectures
- GPU - General AI
- TPU - Data center
- WSE - Wafer-scale
- Complete Comparison Table