Prediction and speculation - title slide
von Neumann's model
linear scaling of speed-quadratic scaling of trans#
modern microprocessors
ilp -instruction level parallelism
von neumann's porjection postulate
enabling technologies for ilp
intel 386
pipeline
control xfer instructions
control xfer in assembly
comparison control xfer - hll-assembly
instruction dependencies
pipeline at work
limits of pipelining
superscalar basics
superscalar/2
superscalar at work
power 5
pentium 4
feature size, frequency, complexity
SPEC
occurence of branches
branches by type
branches by frequency
branches by taken rate
speculation
speculative execution
Slide 29
Slide 30
misprediction effects
efficiency versus bp accuracy
Slide 33
15 years of x86
branch prediction methods
static branch prediction
semi-static branch prediction
Dynamic branch prediction Slide 35
2bc 2 bit saturated counter
bimodal predictor
branch correlations
2 level / correlation predictor
taxonomy of 2 level predictors
gshare
path correlated predictor
tournament/meta predictor
data compression
markov predictor
ppm
neural methods
perceptron
bipolar perceptron
perceptron training
perceptron limitations
bp with perceptrons
the dataflow limit
exceeding the dataflow limit
value locality
value prediction/1
value prediction/2
value prediction/3
speculation taxonomy
research areas
reverse engineering
Slide 65
Slide 66
Slide 67
scheduling
Slide 69
Slide 70
Author: roberto innocente
E-mail: inno@sissa.it
Author: Roberto Innocente