The C7x DSP core is a high-performance
digital signal processor (DSP) developed by Texas Instruments as part of their fixed
and floating-point DSP platform. It represents the cutting-edge of DSP technology
and is designed to handle complex mathematical operations and large data processing
tasks with efficiency and speed. It features a Very-Long-Instruction-Word (VLIW)
architecture, which allows for significant mathematical processing capabilities
through wide vector instructions and multiple functional units. With its vector
(SIMD) instructions and fully pipelined computational instructions, the C7x DSP core
can perform a large number of computations per cycle, making it ideal for
applications that require real-time processing.
- 64-bit Architecture: The
C7x core is a true 64-bit machine with 64-bit memory addressing and the
capability to perform single-cycle 64-bit base arithmetic operations.
- Fixed-Point and Floating-Point
Operations: The core supports both fixed-point and floating-point vector
instructions. It boasts 13 fully pipelined functional units, allowing up to 13
instructions to commence execution on every clock cycle.
- Fully Pipelined Functional Units: Supports the initiation of independent
instructions on every clock cycle, significantly enhancing computational
throughput.
- Single-cycle L2 memory
access: The Streaming Engine of C7x allows single-cycle L2 memory access
for up to 6-dimensional data patterns, bypassing cache.
- Performance Gain: When compared to the previous generation C66x DSP core,
the C7x core achieves a 4 to 8 times increase, or even more, in DSP processing
capacity.
The AM62D device incorporates the
C7504 flavor of the C7x core, featuring a 256-bit wide vector core. This robust core
is designed to handle a variety of operations with high efficiency
- Operations per Cycle: Capable of performing up to 40 FLOPS
(floating-point operations) per cycle, which includes:
- 2x arithmetic/logical operations with 256-bit width (excluding
multiply instructions)
- 2x multiply operation with 256-bit width
- 1x correlation operation or a standard arithmetic operation with
256-bit width
- 1x vector predicate manipulation operation
- 1x non-aligned 256-bit load or store operation
- MAC per Cycle: The C7x-256bV DSP core achieves impressive MAC
(multiply-accumulate) rates per cycle:
- Int16: Up to 64 MAC/cycle.
- Int32: Up to 16 MAC/cycle.
- Float: Up to 16 MAC/cycle.
- Matrix Multiply Accelerator (MMA): The MMA coprocessor augments the C7x
architecture’s scalar and vector capabilities. On the AM62D device, the MMA is
tightly integrated with the C7x core. Through the streaming engines (SEs), it
provides a multitude of MAC operations, matrix-optimized storage, and data
movement. These features are particularly beneficial for dense linear algebra
operations prevalent in applications like vision (CNNs, SfM, filtering, etc.),
speech (xNNs), audio (convolutions), radar (FFTs), and controls (reinforcement
learning, state updates).
- Single-cycle L2 memory
size: 1.25 MB.