LowEnergy MultiUser VSELP Vocoder for a Domain Specific Reconfigurable DSP Architecture

1
Low-Energy Multi-User VSELP Vocoder for a Domain
Specific Reconfigurable DSP Architecture
Roy A. Sutton University of California, Berkeley
CS 294-1 Spring 1998
2
Typical Digital Cellular phone
Encode
ADC
Modulate
Encode
Duplex
Controller
DAC
Demodulate
Decode
Use CODEC to compress the data stream!
3
Typical Digital Base Station
PBX
CH
R/F
...
Can we replace multiple CODEC with a single and
save Space? Cost? Energy?
4
How to Save Energy?

• Eliminate idle time Weiser, etal., OSDI-94
• Reduce energy per Op Chandrakassan, etal.
  ISLPED-96
• Work just fast enough and use most efficient
  processing for most repetitious work!

5
CODEC Data Stream Observations

• Users dont talk continuously (Data is sporadic
  and bursty)
• During silence, very little processing is
  required
• No benefit from processing data faster than
  required
• Data arrives in packets which need be processed
  just fast enough for steady state

6
Which CODEC? answer VSELP

• Compression factor 8x (64k bps to 8k bps)
• Frame size 160 bits

• Output rate 20ms / frame (160 b / 8k bps)
• Divided into 4 sub-frames

7
What is the Repetitious Part?
8
VSELP Repetitious Part (cont)

• 5 of code runs 70 of time and 92 of hits
  (ignoring theta)
• 5 found in two functions 1) dot_product 2)
  iiRfilter
• Make sure these two computations are efficient!

9
HW Implementation Strategy

• Run (the 5) repetitious part on special hardware
  optimized for these two functions
• Run (the 95) remaining part on general purpose
  processor
• Use Pleiades architecture template as guide to
  select architecture instance Pleiades

10
HW for dot_product and iiRfilter

• Can implement using address generators, memory
  elements, and MAC units
• dot_product
• 2 Add Gen
• 2 Memory
• 1 MAC
• iiRfilter
• 3 Add Gen
• 3 Memory
• 1 MAC

11
Architecture Instance
Memory
Memory
CPU
Memory
Add Gen
Add Gen
Add Gen
Interconnect Network
MAC
MAC
Add Gen
Add Gen
Memory
Memory
1 CPU, 12 Satellites, Networked A Domain Specific
Reconfigurable DSP Architecture!
12
HW Simplifying Assumptions

• Satellites and network configuration time is
  ignored
• Network is configured once at startup
  (statically)
• Satellites may be run-time configured
  (dynamically)
• Hardware performance tracks 12 gate ring
  oscillator over voltage (1 - 3 volts)
• Satellites can be accessed by one thread at a
  time
• The network consumes no energy!

13
Processing 1 Stream

• Stream processed by a single thread
• Monitor the stream input buffer level
• Adjust the task priority and hardware throughput
  as required
• Thread uses Satellite processors for repetitive
  code

Q
scheduler
bl1
tp
p1
Q
14
Processing 4 Streams
Q
Q
hw
hw
hw
Threads compete for Satellites Priority now
important
hw
hw
hw
15
Thread Scheduling

• Exists
• Use preemptive multithreaded scheduling
• Dispatch highest priority thread for next time
  slice
• Extensions
• Dynamically adjust each thread priority base on
  its workload
• Dynamically adjust total hardware throughput
  based on aggregate workload

16
Priority and Throughput Adaptation

• 4 Performance Levels (TP)
• for given fsample required, pick voltage via LUT
• Sub-frames mark queue levels (0, 40, 80, 120,
  160)
• Adjust processor throughput and task priority by
  viewing queue levels

17
Satellite Access Simulation Trace
18
Results
19
Conclusions

• Sporadic data stream with fixed throughput can be
  computed with reduced energy by stretching
  computation in time
• Using specialized processors for redundant
  computation can reduce time and energy
• Multiple sporadic data streams can be viewed as a
  single with aggregate duty with slight overhead
• Always compute using maximum time allowable (and
  adjust processor throughput) to minimize energy

20
Future Work

• Account for interconnect network energy
  consumption
• Investigate critical instance adaptation behavior
  and adaptation transients
• Account for satellite and network configuration
  time
• Hide configuration time of satellites by setting
  up next operation during current
• Consider different satellite selection /
  configuration
• When do energy reduction returns diminish?
• What about sensitivity for thread time slice?