Technology Trends and Design Challenges Flashcards
Explain the idea of von Neumann architecture
Linear memory storing the program and data. Control flow processors that execute instructions sequentially
Draw a diagram to show the idea of von Neumann architecture
.
Why do we have a memory hierarchy?
Ideally we would have an infinitely large memory capacity such that each word was immediately available. The best feasible option is to construct a hierarchy of memories, each of which has greater capacity than the preceding but is less quickly accessible
Draw the simplified hierarchy of a modern machine
.
State Moore’s Law
Prediction that transistor density on silicon would double every technology generation (18-24 months). This can be applied in different situations (eg. cost) as well as transistor density
How has the price per transistor changed? Why?
Dropped dramatically. This is due to the rising price of silicon /area combined with what Moore’s law tells us about the halving area/transistor
Draw a graph to show the change in price/mm^2
.
Draw a graph to show the change in mm^2/transistor (Moore’s law)
.
Therefore draw a graph to show the change in price/transistor
.
What is yield?
Proportional of fully functioning chips
Number of defects is proportional to ?
Chip area
Yield is proportional to ?
1/ chip area
What is the current indication about price/transistor?
It is starting to rise for the latest technologies. Semiconductor industry is bifurcating into cheap vs high-performance chips
Can Moore’s law be applied to cost?
Yes. We use Moore’s law (for transistor density) to work out area/transistor which is then combined with price/area to give price/transistor, which halves every two years to abide with Moore’s law. Although there are indications that the cost/transistor may increase so it will no longer apply
Can Moore’s law be applied to performance? Why?
No, not any more. Processor performance no longer doubles every two years because it is constrained by power, instruction level parallelism and memory latency
Can Moore’s law be applied to wire scaling?
Yes. Moore’s law gives us equations for resistance, capacitance and charge time. These tell us that wires don’t get any faster for the same chip area. But halving L makes the wire go 4x faster. Adding buffering/repeaters helps
What does Dennard Scaling say?
In every technology generation the transistor density doubles (Moore’s law), the circuit becomes 40% faster, and power consumption (with twice the number of transistors) stays the same
Is Dennard Scaling still applicable?
No. It broke down because:
1. We cannot further decrease supply voltage since transistors leak too much ie. static power increases dramatically compared to dynamic power
2. Wires don’t scale for the same chip area
3. Semiconductor technology no longer delivers on Moore’s law for performance, so there is more emphasis on computer architecture improvements instead
Is Moore’s law still delivering on transistor density improvements? How?
Yes just about, but we have to constrain the number of transistors that are switching (ie. dynamic power)
How can we constrain the dynamic power in a chip?
- Move towards more accelerators which are optimised for their specific purpose and only turned on when needed
- Move toward a range of processor cores - small efficient cores when lightly loaded and larger cores for heavy workloads, particularly in phones/tablets
What is Race-to-dark?
Turning off a core saves static as well as dynamic power, so this is another goal
What is static power?
Power consumption that is independent of transistor activity
What is dynamic power?
Power consumption due to the switching of transistors
Give the equation for the resistance of a wire
R ∝ L / (W x H)
