Quadric devkit features q16 hybrid AI, DSP, computer vision accelerator

The Quadric devkit is an M.2 Key M module equipped with the company’s q16 edge processor offering a hybrid data-flow + Von Neumann machine for not only neural networks, but also computer vision, digital signal processing, BLAS (Basic Linear Algebra Subprograms), and other workloads.

This architecture allows the Quadric q16 to be more flexible than traditional AI accelerators and can deliver more effective solutions for heterogonous systems that may have multiple accelerators or require a powerful processor. The M.2 form factor enables easy integration into boards with an M.2 Key M socket such as the Gumstix Raspberry Pi 4 development board shown below with the Quadric devkit.

Quadric devkit specifications:

• Accelerator – Quadric q16x4 Processor with 256 Vortex Cores, 8 MB on-chip memory, 4 kB per core memory, multi-precision multiply accumulate; 16 x 16mm package
• Memory – 4GB LPDDR4, 32-bit, dual-rank up to 3677 MHz
• Host Interface – 2-lane PCIe Gen4 up to 4GB/s
• Power consumption – 5.5W
• Dimensions – 80x22mm (M.2 Key M module)

 

The Quadric M.2 module also ships with an M.2 copper heat pipe heatsink for cooling. The company provides the quadric SDK with Source and Graph modes with the former available now, and the latter in closed beta. Source mode gives developers C++ control of the processor’s architectural features, such as control flow, random access, and optimized data-parallel execution such as GEMM, 1D and 2D N-point FFT, and image filtering. The Graph Mode lets developers work with traditional models in ONNX or TensorFlow formats. You can jump to the documentation for the full details.

At the time of writing only the q16 chip, designed for edge applications, is available, but the company is also working on the q8 for embedded sensors, the q32 for larger embedded systems, and the q64 for server applications. The Quadric devkit is supposed to be available now, but it appears you have to contact the company directly to be able to get a sample. More details may be found on the product page.

 

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress