Replies to post #12506 on Advanced Micro Devices Inc (AMD)

[kpf]

aixman

Regardless, very few companies will decide to maintain a separate AMD64 port if there is no need for larger address space.

If the programmers would decide what to do and what not in the future, that is. ;-)

In my vision there is a lot of "FUTUREPROOF! 64-bit-enabled" stickers on softwarepackages next year already. Maybe some of it can benefit from larger adress space, but the majority of it will not. No problem; as long as the application works, nobody cares. What you sell with the product then is the awareness of buyers/users to have made a decision for the future. Which is what AMD sells as well with AMD-64 (Where it is less deceptive, assuming there will be a need for 64-bit before the System sold this year will be obsolete) However, this awareness is valuable. Why noch charge for it..

Maybe the biggest thing with AMD-64 is that there is an opportunity to occupy the "64-bit" thing in the public recognition, which is valuable similar to the Pentium-brand.

K.

[DDB]

I don't want to argue about this but we're really getting into sematics. Pointers are sort of also variables, so even if the most integers are not getting bigger, pointers are. Depending on the code this may or may not be noticable.

Plus, I did say 0-15% slowdown, which is realistic.
a) 64bit int operations on Opteron take the same time like 32bit ops except mul (+1cycle) and div(~ 2x time of 32bit). That's also true for address operations. And when the CPU loads a 32bit value from cache or mem it already transfers 64bit at once.

b) The AMD64 ABI allows passing function/method parameters in integer and fp/sse registers. That gives another possibility for speedup since the push/pop pairs which are usually used are a source for slowdown because the CPU serializes them (IPC==1 then).

But to support ;) your conclusions with numbers I compared GCC 32bit code to 64bit code:

SPECint:
+11,6
+8,1
+2,9
-25,0
+48,2
-15,3
+49,2
-6,1
+1,5
+6,2
+5,0
-6,9
+4,8

SPECfp:
+3,9
-8,6
+18,2
+5,4
+30,2
+93,6
-0,5
+17,2
+22,1
+18,9

Regardless, very few companies will decide to maintain a separate AMD64 port if there is no need for larger address space.

.. or if customers don't want to buy a faster version.

Regards,
DDB

[DDB]

I don't want to argue about this but we're really getting into sematics. Pointers are sort of also variables, so even if the most integers are not getting bigger, pointers are. Depending on the code this may or may not be noticable.

Some real world example (a compiled c file with some SSE2 assembly for FFT calculations):

(size in bytes)
11902 test1
11592 test1_64

The latter one is compiled for 64bit mode.

Regards,
DDB