Anyone using a Phenom II X6 machine with Gnu Radio? If so, what’s your
impression?
–
Principal Investigator
Shirleys Bay Radio Astronomy Consortium
How many SSE instructions can AMD do per clock cycle nowadays? A
couple of years ago Intel could do 1 per clock cycle and AMD could
only do 1/2, which was a huge downer for AMD. If they can now do 1 SSE
per clock cycle, they might be back in the game, which would be a
really good thing in terms of competition.
On the other hand, I think CUDA and OpenCL are in the brink of
changing the whole game. I have a couple of NVidia GTX 295 cards which
are doing most of my crunching nowadays – they are easy to program
and they are much faster than any CPUs on the market.
BTW, wtf is this rant about newbies? People are just panicking with
their coursework – chill-out man.
juha
On 04/28/2010 08:46 PM, Juha V. wrote:
Last time I saw anything in the Gnu Radio context with CUDA, it was not
very
good. The overhead of setting up a transaction with the GPU was too
high
compared to the work you wanted it to do. This may have changed–I’d
be delighted
beyond belief if it were the case.
BTW, wtf is this rant about newbies? People are just panicking with
their coursework – chill-out man.
To be clear, I have nothing against newbs. We are all newbs at one
point or other in whatever
new endeavours we engage in. What I was ranting about was those folks
(newbs or other) who
seem to want a free ride, and get really shirty at you when you
can’t/won’t give them said free
ride.
Juha, I’ve been in engineering for over 30 years, and I’ve found that
there is a class of folks
who belong to the “I’m too busy and important to learn anything new,
why can’t you just
give me the answer”. That class is what I’m complaining about. The
class who feel “entitled”,
who feel that their own inability to make any progress on whatever it
is they’re working on is
not due to their own ineptness or inability/unmotivation to learn, but
rather because “those who
know stuff” are churlishly not helping them.
I worked for many years for one of the well-known, but now failing,
telecom engineering giants.
I used to, when I first got there, help run a technical support help
desk. Just about every single day
I did that job, there would be people calling in, software
developers, engineers, who just couldn’t
fathom that we weren’t there to do their job for them. That because
they were having trouble getting
their heads around how to write whatever 1000 lines of code they’d
been assigned that day, didn’t
make it the tech-support folks job to write it for them.
Anyway, I’m ranting again. I’ll stop. Someone pass me a beer.
Cheers
–
Marcus L.
Principal Investigator
Shirleys Bay Radio Astronomy Consortium
Juha passes beer to Marcus and orders one for himself,
I’m always curious of what hardware gives the most bang per buck – so
I think I’ll take a look at this new Phenom CPU. Maybe AMD is really a
serious contender again. From what I read, the AMD/ATI GPUs are faster
than NVidia offering currently, so I should be looking at the ATI
OpenCL toolkit, but I just haven’t got around to doing it yet. Maybe
the same also applies to CPUs as well.
On the GPU issue that you pointed out: yes, you have to minimize the
bandwidth between the GPU RAM and the CPU RAM. The CPU-GPU bus is
fast, but it is not fast enough yet. There is no sense in eg., sending
a vector for just a FFT and then copying it back. All my current
problems allow me the send the data to the GPU, where I do nearly all
the heavy lifting, before sending the results back to the CPU RAM.
Managing all this is much easier than SSE programming IMHO. The reason
why I am all hyped up about CUDA is that I got a 50x speed up with
very minimal knowledge of GPU programming. All I did was a trivial
parallelization of my functions. The CUDA environment is very close to
ANSI C, so it wasn’t very difficult to grasp for a non-programmer like
me. The learning and coding took only about one week – time well
invested Í think.
juha
On 04/28/2010 08:46 PM, Juha V. wrote:
How many SSE instructions can AMD do per clock cycle nowadays? A
couple of years ago Intel could do 1 per clock cycle and AMD could
only do 1/2, which was a huge downer for AMD. If they can now do 1 SSE
per clock cycle, they might be back in the game, which would be a
really good thing in terms of competition.
Looking at a few benchmarks, it looks like the Phenom II X6 1055T is
very very
comparable in FP performance to the Core 2 Quad QX9770, which is a very
nice
performer indeed.
–
Marcus L.
Principal Investigator
Shirleys Bay Radio Astronomy Consortium