Revolutionize Digital Signal Processing for Flexibility, Adaptability, and Performance
Achieve high performance while maintaining the flexibility to meet today's DSP needs and seamlessly adapt to future AI challenges using Enhanced DSP block with AI Tensor.
A variable precision DSP block is a power and area-efficient hardened IP block that balances compute resources with precision. Each DSP block can be independently configured or cascaded to scale for high-throughput processing.
For example, it can support a single-precision floating-point arithmetic or twice the number of half-precision operations on the same hardware. In total, it supports 6xINT9 or 2xINT16 or 1xCINT16 or 2xINT18x19 or1xINT24 or 1xINT27 or2xBfloat16 or2xFP16 or 2xTFP32 or 1xFP32 operation all with the same hardware.
A tensor is a mathematical object used to represent multi-dimensional arrays of data. Tensors generalize scalars (0-dimensional), vectors (1-dimensional), and matrices (2-dimensional) to higher-dimensional arrays. In Intel® FPGAs, three tensors are provided in every single DSP block and each supports up to 40 INT8 operations per clock.
The Intel® DSP with AI Tensor includes the tensors inside the DSP block. Co-locating the tensor arrays inside the DSP block beside the logic fabric reduces the need for data transfers, bottlenecks, and reduces the challenges of the place and route process.