|
Analog Devices ADSP-21020
The Analog Devices ADSP-21020, introduced in 1993, is a 32-bit
floating-point DSP. The fastest versions run at 33.3 MIPS; all
versions use a 5.0-volt supply. The ADSP-21020 has found use in
imaging applications (for example, medical imaging, graphic arts,
video effects), laboratory test instruments, and military
applications.
While still manufactured by Analog Devices, the ADSP-21020 has been
largely superseded by the faster, more feature-rich, and less
expensive ADSP-2106x, the first versions
of which became available in 1994. The ADSP-2106x and ADSP-21020 share
virtually identical core architectures and the ADSP-2106x executes
ADSP-21020 object code. However, the ADSP-2106x adds substantial
amounts of on-chip RAM, peripherals, and I/O interfaces not found on
the ADSP-21020.
The ADSP-210x0 family formerly included the 12.5 MIPS ADSP-21010,
which was similar in architecture to the ADSP-21020. The ADSP-21010
has been discontinued by Analog Devices.
The ADSP-21020 is notable for its powerful and orthogonal
instruction set, unusual 48-bit instruction word width, shadowing
of all arithmetic and address registers, and lack of on-chip memory.
As in Analog Devices' ADSP-21xx fixed-point DSPs, the ADSP-21020
provides three distinct arithmetic units: a multiplier-accumulator, a
shifter, and an ALU. Unlike the ADSP-21xx, on the ADSP-21020 these
units all access a common register file containing 16 40-bit
registers. Inputs to all arithmetic operations come from the register
file, and results of all arithmetic operations are delivered to the
register file.
The ADSP-21020 supports three data types: 40-bit IEEE floating-point,
32-bit IEEE floating-point, and 32-bit fixed-point. The ADSP-21020
provides nearly complete hardware support for the IEEE
single-precision floating-point format and arithmetic; the processor
additionally supports the IEEE single-extended precision
floating-point format (a 40-bit format that extends the
single-precision mantissa by eight bits for added precision).
The ADSP-21020 provides a 32-bit address space and uses an unusually
large 48-bit instruction word. Like Analog Devices' fixed-point
ADSP-21xx processors, the ADSP-21020 uses a modified Harvard memory
architecture with two memory spaces: program memory (PM) and data
memory (DM), each with its own address and data buses. The ADSP-21020
has no on-chip memory. However, it does include a 32-word on-chip
instruction cache. The separate address and data buses for program
memory and data memory are both brought off-chip. If both sets of
buses are connected to physically separate memories, then the
ADSP-21020 can in general complete one instruction fetch and one data
memory access per instruction cycle.
When an instruction that attempts to perform two data memory accesses
is executed for the first time, the second (PM) data memory access
conflicts with fetching of the next instruction. When this occurs, the
next instruction is stored in the cache, allowing subsequent
executions of the dual-move instruction to execute in a single
cycle. Thus, if the following instruction is contained in the cache,
the processor can complete two data memory accesses in a single
instruction cycle.
As of March, 1997, the 33-MIPS ADSP-21020 sold for $159.00 in quantity
1,000. All versions use a 223-pin PGA package. According to Analog
Devices, a radiation-hardened version of the ADSP-21020 is under
development by TEMIC of France.
A complete analysis of the ADSP-21020's successor, the ADSP-2106x
family, including BDTI Benchmark™ results, is contained in BDTI's
report, Buyer's Guide to
DSP Processors, 2001 Edition.
|