TMS320: the DSP family that taught fixed point

by cera2_admin | Sep 11, 2026 | DSP | 0 comments

The TMS32010 shipped in 1983. It had 144 words of on-chip data RAM, a 16 by 16 multiplier
that produced a result in one instruction cycle, and a Harvard architecture that let it fetch
an instruction and a data word at the same time. Nothing about it is remarkable by modern
standards. Everything about how signal processing is taught still carries its fingerprints.

Why a separate architecture existed at all

A filter is a multiply-accumulate in a loop. On a general-purpose processor of the period
that inner loop cost a multiply routine, several memory accesses and a branch. The TMS320
design collapsed it into a single instruction with automatic pointer updates, so the sample
rate a design could sustain stopped being a software question and became an arithmetic one.
The MAC instruction, circular addressing and the separate program and data buses are still
the three things any DSP course explains first.

Fixed point, and why it stayed

The early parts had no floating-point unit, so everything was Q-format fixed point: the
programmer decided where the binary point sat and carried that decision through every
coefficient, every intermediate result and every overflow check. It is tedious and it is
error-prone, and it is also why the parts were cheap enough to put in a modem, a hard disk
controller or a hearing aid. The C6000 generation added floating point for the designs that
needed it; the fixed-point parts kept shipping in the millions because most designs did not.

What the family became

The name now covers several unrelated architectures: C2000 for motor and power control,
C5000 for low-power fixed point, C6000 for high-throughput VLIW work. Very little code moves
between them. What moved instead was the way engineers think about a signal chain u2014 budget
the cycles, place the binary point, check the headroom u2014 and that habit predates every part
still in production.

Sourcing an old part

Designs from the eighties and nineties are still in service, and the original silicon is
long out of production. The realistic options are a drop-in from a broker with documented
provenance, an emulation on a modern part where the interface is simple enough, or a
redesign. Brokered parts are the fastest route and the one that most often ends in a
counterfeit, so anything safety-relevant gets decapsulated and checked before it goes into a
build.