Computer Architecture 2008 – Advanced Topics 1 Computer Architecture Advanced Topics

Computer Architecture 2008 – Advanced Topics 1

Computer Architecture

Advanced Topics


Pentium® M Processor


Pentium® M Processor Intel’s 1st processor designed for mobility

– Achieve best performance at given power and thermal constraints

– Achieve longest battery life

Banias Dothan

transistors 77M 140M

process 130nm 90nm

Die size 84 mm2 85mm2

Peak power 24.5 watts 21 watts

Freq 1.7 GHz 2.1GHz

L1 cache 32KB I$ + 32KB D$ 32KB I$ + 32KB D$

L2 cache 1MB 2MB


Performance per Watt Mobile’s smaller form-factor decreases power budget

– Power generates heat, which must be dissipated to keep transistors within allowed temperature

– Limits the processor’s peak power consumption

Change the target– Old target: get max performance

– New target: get max performance at a given power envelope Performance per Watt

Performance via frequency increase– Power = CV2f, but increasing f also requires increasing V

– X% performance costs 3X% power Assume performance linear with frequency

A power efficient feature – better than 1:3 performance : power– Otherwise it is better to just increase frequency

– All Banias u-arch features (aimed at performance) are power efficient


Higher Performance vs.Longer Battery Life

Processor average power is <10% of the platform– The processor reduces power in periods of

low processor activity

– The processor enters lower power states in idle periods

Average power includes low-activity periods and idle-time– Typical: 1W – 3W

Max power limited by heat dissipation

– Typical: 20W – 100W

Decision

– Optimize for performance when Active

– Optimize for battery life when idle

Display(panel + inverter)

33%

CPU10%

Power Supply10%

Intel® MCH9%

Misc.8%

GFX8%

HDD8%

CLK5%

Intel® ICH3%

DVD2%

LAN2%

Fan2%


Static Power

The power consumed by a processor consists of

– Active power: used to switch transistors

– Static power: leakage of transistors under voltage

Static power is a function of

– Number of transistors and their type

– Operating voltage

– Die temperature

Leakage is growing dramatically in new process technologies

Pentium® M reduces static power consumption– The L2 cache is built with low-leakage transistors (2/3 of the die transistors)

Low-leakage transistors are slower, increasing cache access latency The significant power saved justifies the small performance loss

– Enhanced SpeedStep® technology Reduces voltage and temperature on low processor activity


Less is More Less instructions per task

– Advanced branch prediction reduces #wrong instructions executed

– SSE instructions reduce the number of instructions architecturally

Less uops per instruction

– Uops fusion

– Dedicated stack engine

Less transistor switches per micro-op

– efficient bus

– various lower-level optimizations

Less energy per transistor switch

– Enhanced SpeedStep® technology

Power-awareness top to bottomPower-awareness top to bottom


Improved Branch Predictor Pentium® M employs best-in-class branch prediction

– Bimodal predictor, Global predictor, Loop detector

– Indirect branch predictor

Reduces number of wrong instructions executed

– Saves energy spent executing wrong instructions

Loop predictor

– Analyzes branches for loop behavior Moving in one direction (taken or NT)

a fixed number of times Ended with a single movement

in the opposite direction

– Detect exact loop count – Loop predicted accurately

PredictionLimitCount

=

+1

0


Indirect Branch Predictor Indirect jumps are widely used in object-oriented code (C++, Java)

Targets are data dependent– Resolved at execution high misprediction penalty

Initially, allocate indirect branch only in target array (TA)– If TA mispredicts allocate in iTA according to global history

Multiple targets allocated for a given branch

– Indirects with a single target predicted by TA, saving iTA space

Use iTA if TA indicates indirect branch + iTA hits

Target Array

iTA

Branch IP

Predicted Target

hitindirect branch

hit

target

HIT

global history

target


Dedicated Stack Engine PUSH, POP, CALL, RET update ESP (add or sub an offset)

– Use a dedicated add uop

Track the ESP offset at the front-end– ID maintains offset in ESP_delta (+/- Osize)

– Eliminates need for uops updating ESP

– Patch displacements of stack operations

In some cases, ESP actual value is needed– For example: add eax, esp, 3

– A sync uop is inserted before the instruction if ESP_delta != 0 ESP = ESP + ESP_delta

– Reset ESP_delta

ESP_delta recovered on jump misprediction


ESP Tracking Example

ESP = SUB ESP, 8

STORE [ESP-8], EBX

STORE [ESP-4], EAX

PUSH eax

PUSH ebx

INC eax

INC esp

ESP = ESP - 4

STORE [ESP], EAX

ESP = ESP - 4

STORE [ESP], EBX

EAX = ADD EAX, 1

ESP = ADD ESP, 1

Δ = Δ - 4

Δ = 0

Δ = - 4

Δ = - 8Δ = Δ - 4

EAX = ADD EAX, 1

ESP = ADD ESP, 1

Δ = - 8

Δ = - 8

Δ = 0

Δ = 0

Sync ESP !


Uop Fusion The Instruction Decoder breaks an instruction into uops

– A conventional uop consists of a single operation operating on two sources

An instruction requires multiple uops when– the instruction operates on more than two sources, or

– the nature of the operation requires a sequence of operations

Uop fusion: in some cases the decoder fuses 2 uops into one uop– A short field added to the uop to support fusing of specific uop pairs

Uop fusion reduces the number of uops by 10%– Increases performance by effectively widening rename, and retire bandwidth

– More instructions can be decode by all decoders

The same task is accomplished by processing fewer uops– Decreases the energy required to complete a given task


A 2-uop Load-Op

Decoder

add eax,[ebp+4*esi+8]

Scheduler

LD

MEU ALU

OP

LD

OP

LD OP

tmp=load[ebp+4*esi+8]

eax = eax + tmp

Load-op with 3 reg. operands

Decoded into 2 uops LD: read data from mem OP: reg ← reg op data

The LD and OP are inherently serial

OP dispatched only when LD completes


A 1-uop Load-Op

Decoder

add eax,[ebp+4*esi+8]

Scheduler

Cache

LD + OP

LD + OP

LD

ALUOP

eax = eax + load[ebp+4*esi+8]

Decoded into 1 uopFused uops has a 3rd source – new field in uop holds index registerIncrease decode BW

Increase alloc BW andROB/RS effective size

Dispatched twiceOP dispatched after LD

fused uop retires after both LD&OP complete Increase retire BW


Enhanced SpeedStep™ Technology The “Basic” SpeedStep™ Technology had

– 2 operating points – Non-transparent switch

The “Enhanced” version provides– Multi voltage/frequency operating points. The Pentium M processor 1.6GHz

operation ranges: From 600MHz @ 0.956V To 1.6GHz @ 1.484V

– Transparent switch– Frequent switches

Benefits– Higher power efficiency

2.7X lower frequency 2X performance loss >2X energy gain

– Outstanding battery life– Excellent thermal mgmt.

Voltage, Frequency, Power

0.0

0.4

0.8

1.2

1.6

2.0

2.4

2.8

3.2

3.6

4.0

0.8 1.0 1.2 1.4 1.6Voltage (Volt)

Fre

qu

ency

( GH

z

)

0

2

4

6

8

10

12

14

16

18

20

Ty

pic

al P

ow

er

( Wa

tts

)

Freq (GHz)

Power (Watts)

2.7X2.7X

6.1X6.1X

Efficiency ratio = 2.3


Trace Cache(Pentium® 4 Processor)


Trace Cache

Decoding several IA-32 inst/clock at high frequency is difficult– Instructions have a variable length and have many different options

– Takes several pipe-stages Adds to the branch mis-prediction penalty

Trace-cache: cache uops of previously decoded instructions– Decoding is only needed for instructions that miss the TC

The TC is the primary (L1) instruction cache – Holds 12K uops

– 8-way set associative with LRU replacement

The TC has its own branch predictor (Trace BTB)– Predicts branches that hit in the TC

– Directs where instruction fetching needs to go next in the TC


Traces Instruction caches fetch bandwidth is limited to a basic blocks

– Cannot provide instructions across a taken branch in the same cycle

The TC builds traces: program-ordered sequences of uops– Allows the target of a branch to be included in the same TC line as the branch

itself

Traces have variable length– Broken into trace lines, six uops per trace line

– There can be many trace lines in a single trace

Jump into the line

Jump out of the line

jmpjmp

jmpjmp jmpjmpjmpjmpjmpjmp


Hyper Threading Technology(Pentium® 4 Processor )

Based on

Hyper-Threading Technology Architecture and Micro-architecture

Intel Technology Journal


Thread-Level Parallelism

Multiprocessor systems have been used for many years– There are known techniques to exploit multiprocessors

Software trends– Applications consist of multiple threads or processes that can be

executed in parallel on multiple processors

Thread-level parallelism (TLP) – threads can be from– the same application– different applications running simultaneously– operating system services

Increasing single thread performance becomes harder– and is less and less power efficient

Chip Multi-Processing (CMP)– Two (or more) processors are put on a single die


Multi-Threading Multi-threading: a single processor executes multiple threads Time-slice multithreading

– The processor switches between software threads after a fixed period

– Can effectively minimize the effects of long latencies to memory

Switch-on-event multithreading – Switch threads on long latency events such as cache misses

– Works well for server applications that have many cache misses

A deficiency of both time-slice MT and switch-on-event MT– They do not cover for branch mis-predictions and long dependencies

Simultaneous multi-threading (SMT)– Multiple threads execute on a single processor simultaneously w/o switching

– Makes the most effective use of processor resources Maximizes performance vs. transistor count and power


Hyper-threading (HT) Technology HT is SMT

– Makes a single processor appear as 2 logical processors = threads

Each thread keeps a its own architectural state– General-purpose registers

– Control and machine state registers

Each thread has its own interrupt controller – Interrupts sent to a specific logical processor are handled only by it

OS views logical processors (threads) as physical processors– Schedule threads to logical processors as in a multiprocessor system

From a micro-architecture perspective– Thread share a single set of physical resources

caches, execution units, branch predictors, control logic, and buses


Two Important Goals

When one thread is stalled the other thread can continue to make progress– Independent progress ensured by either

Partitioning buffering queues and limiting the number of entries each thread can use

Duplicating buffering queues

A single active thread running on a processor with HT runs at the same speed as without HT – Partitioned resources are recombined when only one thread is active


Front End Each thread manages its own next-instruction-pointer Threads arbitrate TC access every cycle (Ping-Pong)

– If both want to access the TC – access granted in alternating cycles

– If one thread is stalled, the other thread gets the full TC bandwidth

TC entries are tagged with thread-ID – Dynamically allocated as needed

– Allows one logical processor to have more entries than the other

TC Hit TC Miss


Front End (cont.)

Branch prediction structures are either duplicated or shared– The return stack buffer is duplicated

– Global history is tracked for each thread

– The large global history array is a shared Entries are tagged with a logical processor ID

Each thread has its own ITLB

Both threads share the same decoder logic– if only one needs the decode logic, it gets the full decode bandwidth

– The state needed by the decodes is duplicated

Uop queue is hard partitioned– Allows both logical processors to make independent forward progress

regardless of FE stalls (e.g., TC miss) or EXE stalls


Out-of-order Execution ROB and MOB are hard partitioned

– Enforce fairness and prevent deadlocks

Allocator ping-pongs between the thread– A thread is selected for allocation if

Its uop-queue is not empty its buffers (ROB, RS) are not full It is the thread’s turn, or the other thread cannot be selected


Out-of-order Execution (cont) Registers renamed to a shared physical register pool

– Store results until retirement

After allocation and renaming uops are placed in one of 2 Qs– Memory instruction queue and general instruction queue

The two queues are hard partitioned

– Uops are read from the Q’s and sent to the scheduler using ping-pong

The schedulers are oblivious to threads – Schedule uops based on dependencies and exe. resources availability

Regardless of their thread

– Uops from the two threads can be dispatched in the same cycle

– To avoid deadlock and ensure fairness Limit the number of active entries a thread can have in each

scheduler’s queue

Forwarding logic compares physical register numbers– Forward results to other uops without thread knowledge


Out-of-order Execution (cont)

Memory is largely oblivious– L1 Data Cache, L2 Cache, L3 Cache are thread oblivious

All use physical addresses

– DTLB is shared Each DTLB entry includes a thread ID as part of the tag

Retirement ping-pongs between threads– If one thread is not ready to retire uops all retirement bandwidth is

dedicated to the other thread


Single-task And Multi-task Modes MT-mode (Multi-task mode)

– Two active threads, with some resources partitioned as described earlier

ST-mode (Single-task mode)– There are two flavors of ST-mode

single-task thread 0 (ST0) – only thread 0 is active single-task thread 1 (ST1) – only thread 1 is active

– Resources that were partitioned in MT-mode are re-combined to give the single active logical processor use of all of the resources

Moving the processor from between modes

ST0 ST1

MTThread 1 executes HALT

LowPower Thread 1 executes HALT

Thread 0 executes HALT

Thread 0 executes HALT

Interrupt


Operating System And Applications

An HT processor appears to the OS and application SW as 2 processors– The OS manages logical processors as it does physical processors

The OS should implement two optimizations:

Use HALT if only one logical processor is active– Allows the processor to transition to either the ST0 or ST1 mode

– Otherwise the OS would execute on the idle logical processor a sequence of instructions that repeatedly checks for work to do

– This so-called “idle loop” can consume significant execution resources that could otherwise be used by the other active logical processor

On a multi-processor system, – Schedule threads to logical processors on different physical processors

before scheduling multiple threads to the same physical processor

– Allows SW threads to use different physical resources when possible

Documents

Computer Architecture 2008 – Advanced Topics 1 Computer Architecture Advanced Topics