Reduction in synchronisation in bundled data systems

1
Reduction in synchronisation in bundled data
systems

• Charlie Brej
• APT Group
• Manchester University

2
Outline

• Asynchronous Circuits
• Early Output
• Counterflow Networks
• Anti-Tokens
• Conclusions

3
Asynchronous Latch
Ri
Ro
Latch
Ao
Ai
Req
Ack
4
Asynchronous Pipeline
5
Pipeline merging
6
Pipeline merging
C
7
Pipeline merging
C
0
0
1
8
Early Output
Go!
0
0
?
9
Guarding

• Problem
• Inputs
• Late
• Unnecessary
• Acknowledge before ready
• Solution
• Validity signal (Vo)

Ri
Ro
Vo
Latch
Ao
Ai
10
Early Output with Guarding
Go!
0
0
C
1
11
Counterflow Networks

• Nodes fire when a threshold of neighbours have
  fired
• No Inputs/Outputs, just Links
• Neural network like operation

12
Counterflow Networks
13
Counterflow Networks
14
Counterflow Networks
15
Counterflow Networks
16
Anti-Tokens

• Reproduce the counterflow network behavior in the
  single-rail system.
• Inputs allow early acknowledgements
• Output latches can acknowledge early
• Propagate early acknowledgements

17
Anti-Token Generation
Go!
0
0
C
A
18
Anti-Token Propagation
Go!
A
A
C
A
19
Token Pass
T
T
T
20
Anti-Token Pass
A
A
A
21
Token Anti-Token collision
T
?
A
22
Token Anti-Token collision 2
T
T
A
23
Dual-Purpose Signals

• Arbiter free
• Req
• Token Request
• Acknowledge Anti-Token
• Ack
• Request Anti-Token
• Acknowledge Token

Req
Ack
24
Standard
25
Standard
26
Early Output
27
Anti-Tokens
28
Anti-Tokens
29
Conclusions

• Early output easy on most blocks
• Multiplexers
• Bit-level pipelining
• Counterflow
• no arbitration
• Increased performance
• Mixed use of techneques
• Less synchronisations