Communicating

with Peripherals
(Interfaces Part II)
Advanced Digital VLSI Design I
Bar-Ilan University, Course 83-614
Semester B, 2021
13 May 2021

Heavily based on the wonderful lecture

“Interfaces: External/Internal, or why
CPUs suck” by Tzachi Noy, 2019
What do a car and a router have in common?

© Adam May
Teman,
13, 2021
Both the car and the router have interfaces

So let’s go ahead and try to build a router…

© Adam May
Teman,
13, 2021
 Reminder: Memory-Mapped I/O
• Registers and I/O Devices are given an address in the system’s memory map:
• Everything is treated the same as memory.
• To communicate with an I/O, we write to and read from these addresses.
• These are achieved with simple load and store assembler commands.
int peek (char *location) {
// Read from a memory-mapped address
• In C, we can define two functions, return *location;

peek and poke, to accomplish this }

void poke (char *location, char newval) {
easily: // Write to a memory-mapped address
(*location) = newval;
• Now to access a register, just }

define its address, and use #define DEV1 0x1000

these functions ...
dev_status = peek(DEV1)
...
5 poke(DEV1,8); © Adam May
Teman,
13, 2021
 General Purpose I/O (GPIO)
• Most microcontrollers have a set of general purpose input/output (GPIO) pins.
• Can be configured as input pins or output pins.
• Can be programmed by software for various purposes.
GPIO D
Config Q
Register

Output External
Memory- Enable GPIO Pin
Mapped D
Output Q
Register

Memory-
Mapped D
Input Q
Register
6 © Adam May
Teman,
13, 2021
 Example: Blinking a LED
• First, configure the GPIO to be an output.
• Next, create an infinite loop that:
• Toggles the state of the output. #define GPIO_CONFIG_REG 0x10000000
#define GPIO_OUTPUT_REG 0x10000001
• Waits for a given period. #define GPIO_BLINK_PIN 0b00000001
#define BLINK_PERIOD 1000000
int main () {
0 // Set LED connected GPIO PIN to output
toggle_config |= peek(GPIO_CONFIG_REG);
poke(GPIO_CONFIG_REG,toggle_config);
0→1→0→1
while (true) {
// Toggle the state of the GPIO output register
output_status = peek(GPIO_OUTPUT_REG);
poke(GPIO_OUTPUT_REG,output_status~GPIO_BLINK_PIN);
// Wait for a predefined delay
wait(BLINK_PERIOD);
}
}
7 © Adam May
Teman,
13, 2021
 Communicating Off-Chip
• What if we want to communicate with something more sophisticated
than a LED or a button?
• We need a communication protocol.
• Introducing UART
• The Universal Asynchronous
Receiver/Transmitter

Baud Rate = 1
• Baud Rate bit time
• Number of bits per unit time
UART1 UART2
data bits
BW =  Baud Rate
• Bandwidth frame bits
• Data per unit time
8 © Adam May
Teman,
13, 2021
Can we use UART for our router?
BAUD RATE: 115,200 bits/sec

SAMPLE RATE: 230,400 samples/sec

CODE: 40 instructions/sample

OVERHEAD: 9,216,000 instructions/sec

9.2% of CPU time * Assuming a 100MHz

clock frequency
© Adam May
Teman,
13, 2021
 Offload the CPU with a controller
• UART is a slow serial protocol
• One bit is transferred at a time at a low baud rate (e.g., 1200-115200 bits/sec).
• Integrate a specific UART controller that offloads the CPU
• Communicate with the UART through a wider register (e.g., byte, 32-bit).
• Use a Shift Register to serialize/deserialize parallel data
1-byte → 16X speedup
• Use a FIFO to buffer several CPU transactions 64 byte FIFO→512X speedup
Stop 7 6 5 4 3 2 1 0 Start
Shift U0Tx Stop 7 6 5 4 3 2 1 0 Start
1 Data 0
clock Shift 1 Data 0 U0Rx
Transmit shift register clock OE BE PE FE Receive shift register
16-element
FIFO TXEF Fifo empty flag 12-bit, 16-element
RXFE Fifo empty flag
FIFO
TXFF Fifo full flag RXFF Fifo full flag
Write data UART0_DR_R Read data UART0_DR_R
Transmit data register Receive data register Source: Bard, EE319K
11 Transmit Parallel-In Serial-Out (PISO) © Adam May
Receive Serial-In Parallel-Out (SIPO) Teman,
13, 2021
 But how do we know when it’s done?
• How can the CPU know when a new byte of data is received?
• Simple way: “Polling” do {
// Play games
• Check on the status of the UART every ...
so often to see if data has been received // Poll to see if we're there yet.
status = areWeThereYet();
(or if it is ready to receive new data) } while (status == NO);

• Polling can be carried out with a “busy-wait” loop:

Busy-wait on input from the UART Busy-wait on writing to the UART
while (TRUE) { current_char = mystring;
// Wait until a new character has been read // Continue until the end of string
while (peek(UART_IN_STATUS)==0); while (*current_char != ‘\0’) {
// Read the new character // Wait until the UART is ready
achar=(char)peek(UART_DATA); while (peek(UART_OUT_STATUS)!=0);
} // Send character to UART
poke(UART_DATA_OUT, *current_char);
// update character pointer
current_char++;
12 } © Adam May
Teman,
13, 2021
 Interrupts
• An interrupt is an asynchronous signal from a peripheral to the processor.
• Can be generated from peripherals external or internal
to the processor, as well as by software.
• Frees up the CPU, while the
peripheral is doing its job.

• Upon receiving an interrupt:

• The CPU decides when to handle the interrupt
• When ready, the CPU acknowledges the interrupt Source: Computers as
Components
• The CPU calls an interrupt service routine (ISR)
• Upon finishing, the ISR returns and the CPU continues operation

14 © Adam May
Teman,
13, 2021
RATE: 115200 bit/sec
0.1152 Mbps

RANGE: 15m
 Ethernet
• Widely used for realization of Local Area Networks (LANs)
• Bus with single signal path
• Nodes are not synchronized → Collisions
• Arbitration: “Carrier Sense Multiple Access
with Collision Detection (CSMA/CD)” Source: Computers as
• If collision → wait for random time → retransmit. Components

• Ethernet packet:
• Addresses
• Variable-length data payload: 46 – 1518 bytes
• Throughput:
• 10M = 2.5 x 4bit 100M = 25 x 4bit 1G = 125 x 8bit

21 © Adam May
Teman,
13, 2021
 Side note: The OSI Model Ethernet
• The Open Systems Interconnection (OSI) model defines seven network layers.
1. Physical: electrical and physical components
2. Data Link: Peer2Peer communication across
a singe physical layer.
3. Network: basic routing over the link.
4. Transport: ensure data is delivered in the proper
order and without errors across multiple links.
5. Session: interaction of end-user services
across a network
6. Presentation: defines data exchange formats
including encryption and compression.
7. Application: interface between the network
Source: Computers as
and end-user Components
22 © Adam May
Teman,
13, 2021
Ethernet

© Adam May
Teman,
13, 2021
Let’s try a simple interface: APB
• 32-bit bus
• Two phase access:
• Address phase
• Read/Write phase

Source: ARM

© Adam May
Teman,
13, 2021
Is APB Sufficient?
ETH RATE: 109 bits/sec
APB TRANSFER WIDTH: 32 bits
APB RATE: 2 cycles/transfer
CLOCK: 108 cycles/sec
APB THROUGHPUT: 1.6 x 109 bits/sec

© Adam May
Teman,
13, 2021
So let’s make it faster: AHB
• Wider bus (>32 bits)
• Pipelined address and R/W phases (X2 throughput)
• Supports Bursts

Source: ARM
© Adam May
Teman,
13, 2021
Is AHB fast enough?
ETHERNET RATE: 2 x 109 bits/sec

AHB TRANSFER WIDTH: 64 bits

AHB RATE: 1 cycle/transfer
CLOCK: 108 cycles/sec
AHB THROUGHPUT: 6.4 x 109 bits/sec

© Adam May
Teman,
13, 2021
Can the CPU support this?
ETHERNET RATE: 2 x 109 bits/sec

CPU WORD: 32 bits

INSTRUCTIONS PER SW/LW: 3 inst/load
CLOCK: 108 cycles/sec
CPU THROUGHPUT: 1.1 x 109 bits/sec

© Adam May
Teman,
13, 2021
 DMA
• Direct memory access (DMA) is a bus operation that allows reads and writes
not controlled by the CPU.
• A DMA transfer is controlled by a DMA controller (DMAC)
that requests control of the bus from the CPU.
• After gaining control, the DMA controller performs read and write operations
directly between devices and memory.

• DMA adds two new signals:

• Bus request
• Bus grant

Source: Computers as
Components
31 © Adam May
Teman,
13, 2021
 DMA Registers
• The CPU controls the DMA operation through registers in the DMA controller.
• Starting address register
• Length register
• Status register – to start and stall the DMA
• After the DMA operation is complete,
the DMA controller interrupts the CPU to tell it that the transfer is done.
• DMA controllers usually use short bursts (e.g., 4-16 words) to only occupy the
bus for a few cycles at a time

Source: Computers as
32 Components © Adam May
Teman,
13, 2021
Data Rate or Packet Rate?
BIT RATE: 2 x 109 bits/sec
each packet includes 20 bytes of overhead

DATA RATE: BIT-RATE x P/(P+20)

P=64 → 2 x 0.76 x 109 bits/sec
Larger packets mean
interconnect is busier P=1518 → 2 x 0.98 x 109 bits/sec

PACKET RATE: BIT-RATE / ((P+20) x 8)

P=64 → 2 x 1.48 x 106 packets/sec
Smaller packets mean
CPU has to do more P=1518 → 2 x 81.2 x 103 packets/sec
© Adam May
Teman,
13, 2021
 How does this affect the CPU?
• The CPU (in a router) needs to handle the packet
• i.e., figure out where to send the packet to.
• So, all it cares about is packet rate
• How much work can the CPU do on each packet?
• For packets with 1518 bytes of data, the packet rate is about 160K packets/sec
• At 100 MHz → 615 instructions per packet → Not that much
• For 64 byte packets → 34 instructions per packet → Infeasible!

• What can we do???

• Trivial solution: Raise the frequency
• Still not enough: Add additional CPUs (ASIPs)
• Better solution: Integrate dedicated hardware (Accelerators)
35 © Adam May
Teman,
13, 2021
A typical router SoC

HOST
ASIP ASIP ASIP
TRAFFIC 0 1 2
PARSER CLASSIFIER POLICER
MANAGEMENT

I/F I/F I/F I/F I/F I/F

0 1 2 3 4 5

© Adam May
Teman,
13, 2021
 But what about memory?
• We now have a processor that can communicate with peripherals,
with an off-chip network, etc.
But what about memory?
• Our router needs a lot of memory:
• To buffer packets
• To store routing tables
• To host the operating system
• …
• The on-chip memory (~MB) is nowhere near enough.
We need to use DRAM
37 © Adam May
Teman,
13, 2021
 DRAM Organization

Rank

Source: Onur Mutlu

Chip
Bank

38 Source: Bruce Jacob © Adam May

Teman,
13, 2021
 8K
DRAM Organization
• So, for an 1GB DIMM (i.e., 8 chips),
we need chips with 1Gb of memory
• e.g., 128k x 8k

128K 1 Gb
• But that is a lot of rows…

© Adam May
Teman,
13, 2021
 8K
DRAM Banks
• So, we break it into 8 banks of 16k rows 16K 128 Mb
• And readout an entire row to a buffer
• Once the row is buffered, we can directly
access any byte in the buffer.

8K

16K 128 Mb
Source: Computers as Components
© Adam May
Teman,
13, 2021
 8K
DRAM Banks B0/R0 B0/R0
B1/R1 B1/R1
16K 128 Mb B0/R0 B0/R0
B1/R1 B1/R1
B0/R0 B0/R0
B1/R1 B1/R1
8K B0/R3 B0/R0
B0/R0 B1/R1
16K 128 Mb B1/R2 B0/R3
B1/R1 B1/R2
© Adam May
Teman,
13, 2021
 AXI
WRITE ADDR CHANNEL

WRITE DATA CHANNEL

MASTER SLAVE
WRITE RESPONSE CHANNEL

READ ADDR CHANNEL

WRITE DATA CHANNEL

© Adam May
Teman,
13, 2021
CACHE
 But what happens on Startup?
• Start by reading from a BootROM
• A small piece of memory, which contains
Source: Intel
the very first code that is executed upon reset.
• Either hard-wired (mask-ROM) or rewriteable (embedded Flash)
• Can use bootstraps/fuses to change configuration.
• Then move on to the Bootloader
• Usually stored on rewriteable flash (i.e., SD card)
• Configures the chip and some of the peripherals
• Loads the end application (e.g., OS) from storage (flash, SSD, HDD)
• Passes control to the end application
• The BootROM and bootloader can be combined
• The BootROM of an x86 system is called the BIOS
45 © Adam May
Teman,
13, 2021
Summary
• Processors are great at processing.
• They are not so great at data movement.
• They are not so great at doing simple tasks.

• For a well-defined task, dedicated hardware

will never lose to a processor.

• With Processors we gain flexibility.

• SW development is faster than HW.
• Bugs in SW are much cheaper to fix.
© Adam May
Teman,
13, 2021
 Main References
• Tzachi Noy, “Interfaces: External/Internal, or why CPUs suck”, 2019
• Wolf, “Computer as Components - Principles of Embedded Computing System
Design,” Elsevier 2012

48 © Adam May
Teman,
13, 2021