Blaize Xplorer X1600 (Edge AI, 7K TOPS/W)
Product Overview
Blaize (formerly ThinCI) is a US edge AI chip company, founded 2011, headquartered in Elk Grove, California. Xplorer X1600 is its flagship edge AI inference chip, 2024 H1 released, TSMC 16nm, 160 TOPS INT8, 22W TDP, efficiency 7.27 TOPS/W (one of the highest-efficiency AI chips in the industry). Based on proprietary GSP (Graph Streaming Processor) dataflow architecture, fully programmable (no CUDA black box).
Strategic significance: Blaize is the efficiency leader in Edge AI, competing with Hailo and Sima.ai, targeting automotive ADAS, industrial vision, smart cities, retail analytics etc. 2024 customers: Motional (autonomous driving), VisionBank (smart cities), Argus (connected cars).
Core Specs
| Item | Parameter |
|---|
| Architecture | Blaize GSP (Graph Streaming Processor) |
| Process | TSMC 16nm |
| Compute Cores | 2 GSP tiles (each 8x10 streaming processing array) |
| GSP Streams | 16 parallel streams |
| LPDDR4 | 4GB LPDDR4 |
| Memory Bandwidth | 51.2 GB/s |
| INT8 | 160 TOPS |
| FP16 | 80 TFLOPS |
| TDP | 22 W (one of the lowest TDP high-TOPS AI chips) |
| Efficiency | 7.27 TOPS/W (H100 ~2.16, Hailo-8 ~1.4) |
| Form Factor | M.2 / PCIe / SoM (System on Module) |
| Interconnect | PCIe Gen3 x4 / USB 3.0 |
| Mass Production | 2024 H1 |
| Unit Price | ~$200-500 (SoM module) |
GSP (Graph Streaming Processor) Architecture
| Dimension | Traditional GPU | Blaize GSP |
|---|
| Execution Model | Imperative (threads) | Graph streaming (dataflow graph) |
| Parallelism | Thread-level (1000s) | Operator-level (streams) |
| On-chip Memory | Shared L2 + registers | Large SRAM (8MB per tile) |
| Dataflow | Cache lines | Graph streaming (zero cache miss overhead) |
| Power | 70-700W | 22W |
| Efficiency | 0.1-1 TOPS/W | 7.27 TOPS/W |
| Programmability | CUDA | Blaize Picasso (graphical) |
| Target | Data center | Edge AI |
GSP Tile Detail
Single GSP Tile:
- 8x10 streaming processing array (80 PEs)
- 8MB SRAM
- 16 parallel streams
- DMA engine
Xplorer X1600:
- 2 GSP Tiles
- Total 16MB SRAM
- 160 PEs
- 16 parallel streams
Key advantages:
- Graph execution: model compiled to graph, streamed directly on GSP
- Zero cache overhead: data flows inside SRAM, no HBM wait
- High efficiency: 22W achieving 160 TOPS (no HBM power overhead)
7K TOPS/W Efficiency Comparison
| Metric | Blaize Xplorer X1600 | Hailo-8 | NVIDIA L4 | NVIDIA Jetson Orin NX |
|---|
| INT8 | 160 TOPS | 26 TOPS | 485 TOPS | 100 TOPS |
| TDP | 22W | 2.5W | 72W | 25W |
| Efficiency | 7.27 TOPS/W | 10.4 TOPS/W | 6.7 TOPS/W | 4 TOPS/W |
| Price | ~$300 | ~$200 | ~$2,500 | ~$600 |
| Software | Blaize Picasso | HailoRT | CUDA | CUDA + JetPack |
| Maturity | early | mass production | mass production | mass production |
Blaize efficiency advantage: 160 TOPS @ 22W is 2.5x NVIDIA Jetson Orin NX 100 TOPS @ 25W compute + 12% power savings. Key product for industrial / automotive / smart city.
Software Stack Blaize Picasso
| Layer | Tool | Description |
|---|
| AI framework | Blaize Picasso | graphical model compilation (no-code) |
| PyTorch / ONNX | model import |
| TensorFlow | compatible |
| Compiler | Picasso Compiler | model -> GSP binary |
| Runtime | Blaize Runtime | Edge deployment |
| Visualization | Picasso Studio | graphical debugging + profiling |
| Model Zoo | ModelZoo | YOLOv8, ResNet, EfficientDet, MobileNet |
Warning: Ecosystem limitation: Blaize Picasso only 2-3 years old, operator coverage ~60% (vs HailoRT 80%). YOLOv5/v8 fully supported, but LLM inference weak (only 7B INT4 quantized).
| Item | Content |
|---|
| Company | Blaize (formerly ThinCI) |
| Founder | Dinakar Munagala (CEO) + Satyam Dronamraju |
| Founded | 2011 (originally ThinCI, renamed Blaize 2020) |
| Headquarters | Elk Grove, California, USA |
| Funding | $180M+ (Series C 2024-Q1 led by: Temasek + Franklin Templeton) |
| Valuation (2025) | ~$500M |
| 2024 Revenue | ~$30M |
| Employees | ~300 |
| Fab | TSMC 16nm |
| Customers | Motional (autonomous driving), VisionBank (smart city), Argus (connected cars), Blaize AI Edge |
| Status | preparing SPAC IPO (2026-2027 expected) |
Use Cases
- ✅ Automotive ADAS (multi-camera 8-12 streams simultaneous inference)
- ✅ Industrial vision (production line defect detection)
- ✅ Smart cities (multi-channel video analytics)
- ✅ Retail analytics (foot traffic + product recognition)
- ✅ Robotics (real-time perception)
- ✅ Low-power AI (battery-powered 22W)
- ❌ LLM inference (only 7B INT4 quantized)
- ❌ Data center (22W compute insufficient)
- ❌ AI training (inference only)
Blaize Product Line
| Product | Released | Compute INT8 | TDP | Target |
|---|
| Xplorer X1600 | 2024 H1 | 160 TOPS | 22W | Edge AI flagship |
| Pathfinder | 2023 H1 | 80 TOPS | 12W | entry Edge |
| Xplorer S1 | 2025 H2 (est.) | 320 TOPS | 40W | high-end Edge |
| Xplorer A1 | 2026 (est.) | 640 TOPS | 80W | server Edge |
Blaize vs Hailo-8 vs Jetson Orin
| Dimension | Blaize Xplorer X1600 | Hailo-8 | Jetson Orin NX |
|---|
| Compute | 160 TOPS | 26 TOPS | 100 TOPS |
| TDP | 22W | 2.5W | 25W |
| Efficiency | 7.27 TOPS/W | 10.4 TOPS/W | 4 TOPS/W |
| Multi-stream Video | 12 streams 1080p YOLOv8 | 4 streams | 6 streams |
| Software | Picasso (new) | HailoRT (mature) | JetPack (mature) |
| Price | $300 | $200 | $600 |
| Best For | Multi-camera Edge | Single camera | Entry-level robotics |
Blaize killer feature: single chip 12 streams 1080p YOLOv8 real-time inference (22W), the ultimate HW solution for 100+ camera smart city deployments.
Key Features
- GSP graph streaming architecture: no HBM wait, ultra-high efficiency
- 22W TDP: battery / solar powered deployment
- 160 TOPS: largest 22W Edge AI compute in industry
- 12-stream YOLOv8: smart city grade multi-camera
- Drawbacks: software ecosystem 2-3 years, weak LLM inference