1. The network-on-chip protocol
Packetized communication
PAYLOAD
HEADER
TAIL
Packet
FLIT …
What does a switch do
when a packet arrives?
swA
swB
The way packets (or packet flits) reserve buffer and link bandwidth
is called the “switching strategy”

2. Packet-based switching
Store-and-forward A B
After A finishes receiving the entire packet,
A asks B if B is ready for the entire packet.
(2) B  A, ack
(3) A sends the packet to B.
Buffering for the largest packet is needed at each switch
Extremely high latency
Full packet storage latency experienced at each hop
Packet not decomposed into flits
It is actually broken down into words (phits), but there is no special control for them
Used for Time-Division Multiplexing in real-time NoCs

3. Packet-based switching
Virtual-cut-through A B
While A receives a part of the packet,
A asks B if B is ready for the entire packet.
(2) B  A, ack
A starts to send the packet to B
even when A has not yet received the entire packet.
(4) If NACK from B, the entire packet is stored in A
Same buffering requirements as for store-and-forward
Cuts down on latency
Packet propagation is started as soon as possible
Packet not decomposed into flits
Deadlock freedom easier to enforce

4. Flit-buffer switching
Wormhole switching
Pipelining on a flit
basis A B
flit size < packet size
Smaller data space
is needed than
store-and-forward
After A receives a flit of the packet,
A asks B if B is ready to receive a flit
(2) B  A, ack
(3) A sends a flit to B.
Reduced buffering requirements
Flit-based allocation of buffers and channels
Reduced latency wrt store-and-forward
May lead to starvation of other packets
Packet strung out over the network if head flit blocks

5. X X The blocking problem
Wormhole switching suffers from packet blocking problems
Idle
Idle
Blocked X X
Blocked! A B C
C cannot send it
and has no enough
space for a new flit D
An idle channel cannot be used because it is owned by a blocked packet….
Although another packet could use it!!!

6. Virtual-Channels Performance improvement using virtual channels
Multiple virtual channels multiplexed on a single physical link
Node 1
Node 2
Node 3
Node 4
Node 5
Destination of B
Node 1
Node 2
Node 3
Node 4
Node 5
Destination of B A B
Block
Body flits use the VC acquired by the head flit
Tail flit releases VC

7. Flow control ACK-NACK Flow control ON-OFF Flow control
In buffered switching, it determines the way the downstream node communicates buffer availability to the upstream node
Block A B C
Backpressure
Don’t
send
Buffer
full
Don’t
send
Buffer
full
ACK-NACK Flow control
ON-OFF Flow control
Credit based Flow control

8. ACK-NACK Flow control Makes optimistic assumptions for transmission
ACK/NACK propagation
ACK and buffering
Memory deallocation
NACK
Retransmission
Go-back-N
Makes optimistic assumptions for transmission
In case of long-lasting blockage, waste of power!
Long buffer occupancy at the upstream switch
Natively supports link-level error control

9. Credit based flow control
No of credits 2
credit
Rx Buffer B
Receiver gives N credits to sender
Sender decrements count
Stops sending if zero
Receiver sends back
credit as it drains its buffer
Bundle credits to
reduce overhead 1 1 H
credit 1 B T H 1
credit B H
Round trip time between buffer empty and flit arrival
More efficient buffer usage
Large amount of upstream signaling (e.g., short packets)
Error control pushed at a higher layer

10. On-Off flow control Incoming flits Outgoing flits Fon threshold GO!
Foff threshold
STOP!
Switch B
Potentially reduces the amount of upstream signalling
Upstream state: 1 control bit
Permission to send (on) or not (off)
Upstream signalling only to change this state
Buffering requirements to prevent overflow or idling

11. Simple on/off implementation: stall/go
This could be an input or output buffer of the switch
Every stage of the pipelined data path should have two slots
Not to loose the flit in flight while stall is propagated upstream

12. Link Pipelining A link pipeline stage is not just
a retiming stage but also a flow control stage,
hence requires a 2-slot buffer (in general, the number of
additional slots is flow control dependent)