Models & Pricing
SAKURA-II M.2 2280 Module
- 60 TOPS (INT8) / 30 TFLOPS (BF16)
- 8W typical power
- Up to 32GB LPDDR4x, 68 GB/s
- 20MB on-chip SRAM
- M.2 2280 form factor
- -40 to 105°C option
SAKURA-II PCIe Card
- Up to 120 TOPS (single and dual)
- Low-profile single-slot
- PCIe host interface
- Sparse computing support
- Batch=1 low-latency inference
- Enterprise and industrial servers
SAKURA-II System
- Up to 240 TOPS when deployed in systems
- Multi-accelerator scaling
- MERA heterogeneous inference
- Ubuntu, Docker, Jupyter ready
- Edge server deployments
- Defense, robotics, smart city
Specifications
Accelerator
EdgeCortix SAKURA-II with Dynamic Neural Accelerator (DNA) architecture
AI Performance
60 TOPS (INT8) / 30 TFLOPS (BF16)
Power Consumption
8W typical
Compute Efficiency
Up to 90% accelerator utilisation
DRAM Support
Dual 64-bit LPDDR4x — 8GB / 16GB / 32GB
DRAM Bandwidth
68 GB/s
On-chip SRAM
20MB
Temperature Range
-40 to 105°C (industrial); 40 to 85°C commercial variant
Package
19 × 19 mm BGA
Module Form Factor
M.2 2280 (60 TOPS)
Card Form Factor
Low-profile single-slot PCIe, up to 120 TOPS single/dual
System Scale
Up to 240 TOPS across multiple accelerators
Generative AI
Llama 2, Stable Diffusion, DETR and ViT within 8W
Precision
Software-enabled mixed precision with near-FP32 accuracy
Software Stack
MERA compiler and framework — HuggingFace, TensorFlow Lite, ONNX, TVM, MLIR
Runtime Environment
Ubuntu, Docker, Jupyter
Certifications
FCC, CE, UKCA, ICES, VCCI, BSMI, REACH, RoHS
Recognition
2024 Edge AI Product of the Year winner
Ecosystem Partners
MegaChips, SoftBank, BittWare, Renesas RZ/V, Arm/Raspberry Pi 5
Overview
EdgeCortix positions SAKURA-II as the accelerator for the generative-AI era of edge computing, and the numbers support the claim: 60 TOPS of INT8 and 30 TFLOPS of BF16 inference inside an 8W typical power envelope, with support for multi-billion-parameter models. That is a power budget a fanless industrial enclosure can absorb, unlike the discrete GPUs normally required to run comparable workloads.
The architectural foundation is EdgeCortix’s Dynamic Neural Accelerator (DNA), a run-time reconfigurable and highly parallelised array built for Batch=1, low-latency inference rather than throughput-batched datacenter work. The company pairs it with up to 32GB of LPDDR4x at 68 GB/s and 20MB of on-chip SRAM, claiming up to four times the DRAM bandwidth of competing edge accelerators plus sparse-computing support that further reduces memory pressure. The result is up to 90% sustained utilisation, which is where the effective performance gap usually opens up.
Deployment spans three tiers from a single silicon design. The M.2 2280 module delivers 60 TOPS for space-constrained systems. Single and dual low-profile PCIe cards reach 120 TOPS for industrial and enterprise servers. Multi-card systems scale to 240 TOPS. Every tier runs the same MERA compiler and framework, which accepts models from Hugging Face, TensorFlow Lite, ONNX, TVM and MLIR and deploys them across heterogeneous hosts — including EdgeCortix’s own announced Raspberry Pi 5 support and partner integrations with Renesas RZ/V and BittWare.
The silicon is specified from –40 to 105°C and carries FCC, CE, UKCA, ICES, VCCI and BSMI certifications, with a full REACH and RoHS materials package — unusual rigour for an accelerator of this class. EdgeCortix has also validated a radiation-resilient configuration for orbital and lunar missions, which says something about the design’s robustness. For defense, robotics, drone and smart-manufacturing programmes that need generative and vision AI at the edge without dragging in a datacenter power and cooling budget, SAKURA-II is one of the few credible options at 8W.
Target Applications
Defense and security systems, robotics and drones, smart manufacturing inspection, smart city video analytics, automotive sensing, edge AI servers running multimodal models, and any deployment needing Llama 2 or Stable Diffusion inference inside an 8W thermal envelope.
Related Products
Request a Quote — EdgeCortix SAKURA-II
QS Compute — global B2B sourcing for EdgeCortix SAKURA-II M.2 modules, PCIe cards and multi-accelerator systems. Evaluation units and volume allocation on request.
Get Your Quote →