|
|
COMP8320 Tutorial 05 -- week 7, 2011
Graphics Processing Units
Please read the articles mentioned below before the tutorial.
Afterwards, discuss in small groups the following questions.
- Why is CPU + GPU a useful programming model?
- In OpenMP, it is possible to incrementally parallelize
an existing serial application. Is this the case with GPUs? Discuss.
- What are the main similarities and differences between the
concepts of threading on chip multiprocessors (e.g. the
Niagara-2) and GPU?
- When programming GPUs, one aim to use a number of threads that is
much greater than the number of cores. Give (at least) two reasons for
this.
- On the S2050 GPU, one can have up to 1,536 threads organized into
warps of 32. How long a latency (in cycles) can be hidden?
- The Fermi SM actually has two warp dispatchers, which dispatches
an instruction from each on 16 cores (as opposed to a single
dispatcher which dispatched all 32 threads). Why is this? (I would
like to know!)
- Consider the matrix multiply calculation C += A*B where
C is nxn, A> is
nxk and B> is kxn -- all
these are double precision. Assume k<<n. A CPU-GPU
system such as Xe has an 8 Gbytes/s transfer speed (via PCI-e
Gen2). Assume that the CPU can perform a (large) matrix multiply at 6
GFLOP/s (note the multiply has 2n^2k FLOPs). How large does
n need to be before it becomes worthwhile to offload this
computation onto the GPU (assuming that the GPU can calculate
infinitely fast)? Repeat the above for a CPU speed of 40 GFLOP/s.
- Table 3 of the first article gives some `speedups' for various
application on a GPU? What precisely do the authors mean by this term?
How meaningful are these figures?
- Critically assess the claim that the future of GPUs such as the
Tesla is secure because of its market in the gaming industry.
i.e. what kinds of games really require such an accelerator?
- How valid is the objection I have N codes but one IT
budget? In other words, what (kind of) applications can be
accelerated effectively and what cannot?
- In terms of heterogeneous processing, where do GPUs fit
in the spectrum?
Last modified: 6/09/2011, 16:49
|
