COMPUTER SCIENCE TRIPOS Part II – 2021 – Paper 9
Comparative Architectures (rdm34)
(a) A superscalar processor may speculatively execute loads even when one or more
earlier stores have not yet computed their memory addresses. In practice, we
would need to restart execution from the speculative load if a memory-carried
dependency is subsequently detected.
(i) With the help of some additional hardware it is possible to record which
loads cause such ordering violations. Briefly outline how this could be done
and how such a record could be used to help improve performance.
[3 marks]
(ii) Describe why such a scheme may unnecessarily delay the issuing of a load
even when the mechanism correctly recalls that the load has led to an order
violation between a store and load in the past? [4 marks]
(b) Why might it also be advantageous for a superscalar processor to predict whether
a particular load will hit or miss in the processor’s L1 data cache? [3 marks]
(c) You are asked to design hardware to run artificial neural network applications in
a high-performance and energy-efficient manner. Such workloads can typically
make good use of many multiply-accumulate (MAC) units operating in parallel
and narrow datatypes. Your system is required to support a range of different
neural networks that vary considerably in the type of computations they
perform. You consider three approaches: (1) to use a multicore processor; (2)
to design a single domain-specific accelerator; (3) to compose your design from
two or more domain-specific accelerators where each is specialised for different
types of neural network.
(i) What are the advantages and disadvantages of each approach? [6 marks]
(ii) Describe one possible way of organising the multicore processor and a
possible choice for the architecture(s) of its individual cores. Briefly justify
your design decisions. [4 marks]