Guys...
Dropped in to give you this...thought you might be interested.
Tom
TechWeb August 31, 1998
DSP Architectures Target Next-Gen Cell Phones
By: Stephan Ohr and contribution by Ron Wilson
New digital signal-processing architectures are on the way from two major
industry powers. DSP Group is coming out with its long-expected Palm core, a
multiple-multiply-accumulator core targeted at cellular telephone applications.
Later this month, at DSP World in Toronto, Philips Semiconductors is scheduled
to roll out a pair of architectures under the umbrella name R.E.A.L. DSP.
DSP Group's multi-MAC PalmDSPCore can execute at up to 450 million
instructions per second (Mips), the vendor said. VLSI Technology, in San Jose,
Calif., has built the first hardware implementation of the soft core, which is
specifically aimed at such new-generation cellular services as wideband
code-division multiple access (CDMA), the Universal Mobile Telecommunications
System (UMTS), and IMT-2000 international mobile communications.
The PalmDSPCore is the latest product of a five-year collaboration with VLSI
through which the DSP Group has delivered architectures and VLSI has implemented
them. VLSI has done 80- and 130-Mips OakDSP cores, sometimes combining them with
ARM7TDMI cores to meet the combined control and signal- processing needs of
cellular telephony and other applications.
The latest implementation of the Oak-plus-ARM combo, for example, uses a 0.
25-micron, 2.5-volt process and delivers 130 Mips with a 130-MHz clock. Thus,
VLSI has been able to offer dual-core embedded processors not only for Global
System for Mobile Communications (GSM) and CDMA cell phones, but also for such
computing and entertainment applications as DVD, cable modems, and digital
television, said Ray Slusarczyk, director of marketing for the company's
Embedded Processor Division, in Tempe, Ariz.
Slusarczyk said he believes embedded DSP will exhibit 40 percent compound
annual growth and expects the segment to represent more than a $ 6 billion total
available market in 2002.
The PalmDSPCore implementation anticipates such third-generation cellular
services as Internet browsing and multimedia messaging. But there are
indications that the DSP Group, whose engineering design resources are located
in Tel Aviv, Israel, may go head-to-head with Lucent and Texas Instruments in
the cellular base station market, as well.
Even in central switching facilities, DSPs for telephone line cards and
asymmetric digital subscriber line ( ADSL) modems must meet stringent size and
low-power constraints to minimize heat and power dissipation in tight spaces,
said Bat-Sheva Ovadia, DSP Group department manager for DSP architecture and
algorithm development. That application would put the PalmDSPCore up against
power-trimmed versions of TI's C6x, which uses a very long-word instruction
(VLIW) architecture, and with Lucent's dual-MAC architecture.
Indeed, Ovadia acknowledged, VLIW was among the architectures considered by
the DSP Group's engineering team, as was a superscalar architecture. The
multi-MAC architecture, with parallel multiply-accumulate units, was ultimately
chosen because in the team's view it represented the best choice for portable
cellular handsets.
"This architecture is known in the DSP world for its silicon efficiency,"
Ovadia said. "It also provides the best programming model."
In addition to high code density, the PalmDSPCore offers users a choice of
fixed-point math-processing capability: Users can harness 16-, 20-, or 24-bit
precision. That not only minimizes die size, but also lets the user avoid buying
more capability than is required, Ovadia said.
She said her group did a careful statistical analysis of cell phone
operations and algorithms and discovered that most could make use of fast
Fourier transforms (FFTs) to execute operations more quickly and efficiently
than is possible via the traditional approach of implementing finite impulse
response (FIR) filters as convolutions in the time domain.
For the FFT-based method, the decision to use multiple MACs rather than VLIW
or superscalar made the most sense, Ovadia maintained. An additional feature of
the multipliers is they can be cross-coupled with the arithmetic logic unit and
arithmetic summing unit (ALU and ASU). That parallelization of resources
supports telephony operations that need parallel operations, such as modem pools
and ADSL (that technology effectively puts 255 parallel modems on one line).
Since the PalmDSPCore will execute three OakDSPCore instructions on each
cycle, its sustained execution rate is 450 Mips with a 150-MHz clock. Its power
peaks at about 2,700 Mips (18 operations times 150 MHz), said Ovadia.
The PalmDSPCore is intended to be code-compatible with the OakDSPCore,
offering an easy upgrade path to OakDSPCore users, though additional
instructions for cell phone support (such as Viterbi detection) have been built
into the new machine. The multi-MAC architecture also makes it possible for
certain operations to be more streamlined than a simple parallelization of
OakDSPCores would allow.
FFTs with rounding take 16 cycles with the OakDSPCore, for example, and only
two cycles with the PalmDSPCore. Complex multiplication takes five cycles with
the Oak and two cycles with the Palm. A complex FIR filter operation gets about
a 2X speedup with the Palm. But Viterbi decoding takes only two cycles on the
Palm, compared with 13 cycles with the Oak.
Since the PalmDSPCore is a "soft core," available for license, its
performance and power consumption will depend on the layout and process
technology used by the implementer. In the VLSI implementation, available in
mid-1999, the PalmDSP Core will use a 0.20-micron, 1.8-V process to produce 450
Mips with a 150-MHz clock. Power consumption is projected to be on the order
of 0.95 milliwatts per MHz.
Philips, meanwhile, is reportedly ready to roll a pair of internally
developed DSP architectures. The company declined to release technical details,
but sources indicated there will be both a conventional, single-MAC family and a
dual-MAC family. The primary feature will be Philips' ability to customize the
cores and their peripherals to precise customer needs.
|
