Feasibility of Accelerator Generation to Alleviate Dark Silicon in a Novel Architecture

This thesis presents a novel approach to alleviating Dark Silicon problem by reducing
power density.
Decreasing the size of transistor has generated an increasing on power consumption.
To attempt to manage the power issue, processor design has shifted from one
single core to many cores. Switching on fewer cores while the others are off helps
the chip to cool down and spread power more evenly over the chip. This means that
some transistors are always idle while others are working. Therefore, scaling down
the size of the chip, and increasing the amount of power to be dissipated, increases
the number of inactive transistors. As a result it generates Dark Silicon, which doubles
every chip generation [63]
One of the most effective techniques to deal with Dark Silicon is to implement
accelerators that execute the most energy consumer software functions. In this way
the CPU is able to dissipate more energy and reduce the dark silicon issue.
This work explores a novel accelerator design model which could be interfaced to
a Stack CPU and so could optimise the transistor logic area and improve energy efficiency
to tackle the dark silicon problem based on heterogeneous multi-accelerators
(co-processor) in stack structure.
The contribution of this thesis is to develop a tool to generate coprocessors from
software stack machine code. But it also employs up-to-date code optimisation
strategies to enhance the code at the input stage. Analysis of the cores using key
metrics, based on 65nm synthesis experiments and industry standard tool-sets. It
further introduces a novel architecture to decrease the power density of the accelerator.
In order to test these expectations, a large-scale synthesis translation experiment
was conducted, covering widely recognised benchmarks, and generating a large
number of cores (in the thousands). These were evaluated for a range of key metrics:
silicon di-area, timing, power, instructions-per-clock, and power density, both with
and without code optimisation applied.
The results obtained demonstrate that one of two competing core models, ‘Wavecore’
(which is proposed in this thesis), delivers superior power density to the standard
approach (which it refers to as Composite core), and that this is achieved without
significant cost in terms of critical metrics of overall power consumption and
critical path delays.
Finally, to understand the benefit of these accelerators, these auto-generated cores
are analysed in comparison to a standard stack-machine CPU executing the same
code sequences. Both the cores generation work and the benchmark CPU assume a
65nm CMOS process node, and are evaluated with industry standard design tools.
It is demonstrated that the generated cores achieve better power efficiency improvements
over the relatively CPU core.