Skip to main content

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

ItemParameter
ArchitectureBlaize GSP (Graph Streaming Processor)
ProcessTSMC 16nm
Compute Cores2 GSP tiles (each 8x10 streaming processing array)
GSP Streams16 parallel streams
LPDDR44GB LPDDR4
Memory Bandwidth51.2 GB/s
INT8160 TOPS
FP1680 TFLOPS
TDP22 W (one of the lowest TDP high-TOPS AI chips)
Efficiency7.27 TOPS/W (H100 ~2.16, Hailo-8 ~1.4)
Form FactorM.2 / PCIe / SoM (System on Module)
InterconnectPCIe Gen3 x4 / USB 3.0
Mass Production2024 H1
Unit Price~$200-500 (SoM module)

GSP (Graph Streaming Processor) Architecture

DimensionTraditional GPUBlaize GSP
Execution ModelImperative (threads)Graph streaming (dataflow graph)
ParallelismThread-level (1000s)Operator-level (streams)
On-chip MemoryShared L2 + registersLarge SRAM (8MB per tile)
DataflowCache linesGraph streaming (zero cache miss overhead)
Power70-700W22W
Efficiency0.1-1 TOPS/W7.27 TOPS/W
ProgrammabilityCUDABlaize Picasso (graphical)
TargetData centerEdge 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

MetricBlaize Xplorer X1600Hailo-8NVIDIA L4NVIDIA Jetson Orin NX
INT8160 TOPS26 TOPS485 TOPS100 TOPS
TDP22W2.5W72W25W
Efficiency7.27 TOPS/W10.4 TOPS/W6.7 TOPS/W4 TOPS/W
Price~$300~$200~$2,500~$600
SoftwareBlaize PicassoHailoRTCUDACUDA + JetPack
Maturityearlymass productionmass productionmass 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

LayerToolDescription
AI frameworkBlaize Picassographical model compilation (no-code)
PyTorch / ONNXmodel import
TensorFlowcompatible
CompilerPicasso Compilermodel -> GSP binary
RuntimeBlaize RuntimeEdge deployment
VisualizationPicasso Studiographical debugging + profiling
Model ZooModelZooYOLOv8, 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).

Vendor Information

ItemContent
CompanyBlaize (formerly ThinCI)
FounderDinakar Munagala (CEO) + Satyam Dronamraju
Founded2011 (originally ThinCI, renamed Blaize 2020)
HeadquartersElk Grove, California, USA
Funding$180M+ (Series C 2024-Q1 led by: Temasek + Franklin Templeton)
Valuation (2025)~$500M
2024 Revenue~$30M
Employees~300
FabTSMC 16nm
CustomersMotional (autonomous driving), VisionBank (smart city), Argus (connected cars), Blaize AI Edge
Statuspreparing 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

ProductReleasedCompute INT8TDPTarget
Xplorer X16002024 H1160 TOPS22WEdge AI flagship
Pathfinder2023 H180 TOPS12Wentry Edge
Xplorer S12025 H2 (est.)320 TOPS40Whigh-end Edge
Xplorer A12026 (est.)640 TOPS80Wserver Edge

Blaize vs Hailo-8 vs Jetson Orin

DimensionBlaize Xplorer X1600Hailo-8Jetson Orin NX
Compute160 TOPS26 TOPS100 TOPS
TDP22W2.5W25W
Efficiency7.27 TOPS/W10.4 TOPS/W4 TOPS/W
Multi-stream Video12 streams 1080p YOLOv84 streams6 streams
SoftwarePicasso (new)HailoRT (mature)JetPack (mature)
Price$300$200$600
Best ForMulti-camera EdgeSingle cameraEntry-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