Electronics Weekly
That is not the only change to the IP, which is aimed at 5G comms and automotive applications.
“With an enhanced instruction set architecture [ISA] and clock-speed increase, the ConnX B20 DSP processes parts of the communication processing chain up to 30x faster, and parts of the radar/lidar processing chain up to 10x faster compared to the [existing] ConnX BBE32EP DSP.”
B20 has a 512bit vector width, up to 128 MACs, can load 1024 bits of data each cycle, and is claimed to achieve 1.4GHz or more if built on a 16nm process.
“Customers have a variety of algorithm acceleration options to reduce cycle counts, including higher precision with either 32bit fixed-point operations – including operations to optimise MAC, FFT and FIR – with native complex support, or single-precision vector floating-point operations, or half precision, with half precision at 2x the single-precision throughput,” said the firm.
An extended vector floating-point option supports complex floating-point operations and doubles the real floating-point operations “at the same vector width typically used in the front end of the radar processing chain”, said Cadence.
A communications option accelerates forward-error correction in lower bit-rate communication applications and software-defined radio.
10x lidar/radar performance claimed above is attributed to native 32-bit support option and higher clock speed, and the 30x comms performance also to higher clock speed, as well as the forward-error correction acceleration.
Alongside this, ConnX B10 has half the vector width of the B20 for applications requiring less parallelism.
Both B20 and B10 are supported by compiler, debug and simulation tools as well as an optimised math library and application examples with source code.
B10 and B20 are software-compatible with earlier ConnX DSP parts – ConnX BBE16/32/64EP).
General availability for both B10 and B20 is planned for the second quarter of 2019, although early access is available to some.
ConnX IP is branded Tensilica
