EM64T performance

[George Macdonald]

There have been lots of comments, including mine, on the absence of Xeon
EM64T 64-bit performance results... anywhere. The only thing I've come
across is this: http://www.computeractive.co.uk/news/1156849 - only one
test and one which would seem to favor memory bandwidth and prefetch. I'm
not sure how it, Gnu Privacy Guard, stacks up as a measure of expected
general performance but it looks like Intel's Xeon/Nocona is performing in
the same ballpark here as AMD's Opteron, if you scale up the 1.8GHz Opteron
in the test to say a 2.2GHz version.

This kinda leaves me wondering why Intel is being so coy/silent on the
subject of 64-bit performance. In the past, even under NDAs, they've
generally managed to find a way to leak favorable comments on their latest
widget.

Rgds, George Macdonald

"Just because they're paranoid doesn't mean you're not psychotic" - Who, me??

 

[Mark Hahn]

across is this: http://www.computeractive.co.uk/news/1156849 - only one

test and one which would seem to favor memory bandwidth and prefetch. I'm

I'm not sure I understand why you say that. the text says that it's
merely encrypting a 200 MB file, and that the program takes 22-29
seconds to run. that's only around 8 MB/s, so memory bandwidth is
NOT even on the radar here (these machines sustain ~3 GB/s.)

encryption (of the nontrivial sort) is normally fairly compute-intensive,
and in that sense, cache-friendly. this is consistent with the results,
which basically show that running in AMD mode helps Intel's Xeon a lot.

I can't see why the AMD-mode boost for Intel would be explained by
bandwidth and prefetch - I'd look instead to the larger number of
architectural registers.

not sure how it, Gnu Privacy Guard, stacks up as a measure of expected
general performance but it looks like Intel's Xeon/Nocona is performing in
the same ballpark here as AMD's Opteron, if you scale up the 1.8GHz Opteron
in the test to say a 2.2GHz version.

yes, the opt/244 is pretty much the entry-level chip from AMD,
so really quite inappropriate to compare to Intel's fastest.

This kinda leaves me wondering why Intel is being so coy/silent on the
subject of 64-bit performance. In the past, even under NDAs, they've
generally managed to find a way to leak favorable comments on their latest
widget.

no doubt they are still troubled by the fact that it's a non-Intel
innovation. perhaps they can "embrace and extend" it enough to make
it thoroughly incompatible, and then warm up to it

regards, Mark Hahn.

 

[George Macdonald]

I'm not sure I understand why you say that. the text says that it's
merely encrypting a 200 MB file, and that the program takes 22-29
seconds to run. that's only around 8 MB/s, so memory bandwidth is
NOT even on the radar here (these machines sustain ~3 GB/s.)

I guess I believed what they say about "ability to move data between memory
and the CPU". I'd also like to see a test which takes longer, for a
better comparison.

encryption (of the nontrivial sort) is normally fairly compute-intensive,
and in that sense, cache-friendly. this is consistent with the results,
which basically show that running in AMD mode helps Intel's Xeon a lot.

I can't see why the AMD-mode boost for Intel would be explained by
bandwidth and prefetch - I'd look instead to the larger number of
architectural registers.

Yes I can see that and it's something which gets totally ignored by most of
the err, "analysts" with their focus on 64-bit addres space as the reason
for AMD64. I don't know that particular algorithm but I'd assume that
there was no advantage taken from the wider registers - IOW a simple
recompile.

yes, the opt/244 is pretty much the entry-level chip from AMD,
so really quite inappropriate to compare to Intel's fastest.



no doubt they are still troubled by the fact that it's a non-Intel
innovation. perhaps they can "embrace and extend" it enough to make
it thoroughly incompatible, and then warm up to it

With the 4GB DMA blunder it looks like it'll be a little time before they
have even full compatibility.

Rgds, George Macdonald

"Just because they're paranoid doesn't mean you're not psychotic" - Who, me??

 

[Grumble]

There have been lots of comments, including mine, on the absence of Xeon
EM64T 64-bit performance results... anywhere.

http://anandtech.com/linux/showdoc.aspx?i=2158

These are both 64-bit CPUs, and so, all benchmarks are run on
64-bit OSes with 64-bit binaries wherever possible.

 

[George Macdonald]

http://anandtech.com/linux/showdoc.aspx?i=2158

These are both 64-bit CPUs, and so, all benchmarks are run on
64-bit OSes with 64-bit binaries wherever possible.

Now things get interesting. There are some baffling aspects of the bar
graph reporting here though, when compared with the build files of John the
Ripper, e.g. Maybe I'm misreading them since I'm not intimately familiar
with the tools or benchmarks... and the article is awfully short on precise
info on which "mode" the CPU was running the benchmarks... possibly rushed
out.

The superior performance of EM64T in some of the important benchmarks only
thickens the plot on Intel's reticence with this CPU.

Rgds, George Macdonald

"Just because they're paranoid doesn't mean you're not psychotic" - Who, me??

 

[JK]

http://www.anandtech.com/linux/showdoc.aspx?i=2163&p=1

 

[Yousuf Khan]

http://www.anandtech.com/linux/showdoc.aspx?i=2163&p=1

This new test changed the order of things pretty completely.

Yousuf Khan