Linus Torvalds ：Core 2 is good, but the K8 core isn\'t bad either

Edison

http://realworldtech.com/forums/ ... =77238&roomid=2

Name: Linus Torvalds (torvalds@osdl.org) 3/10/07

ShaidarHaran (meepmeep@beep.com) on 3/10/07 wrote:
>
>My question is:
>How?
>
>What might AMD have done to improve Barcelona's
>architecture that much over even the vaunted Core
>microarchitecture?
Well, the K8 was pretty good, but the K8 has a few obvious
stumbling blocks:

- the 64-bit paths (and thus all the splitting of
   instructions) for MMX/XMM. Just fixing that will help
   improve vector loads a ton. Core 2 just wiped
   the floor with everybody thanks to having a fundamentally
   stronger full 128-bit vector unit.

- the memory pipeline was fairly unaggressive, and would
   not re-order much at all. Core 2 just did a lot better.

The first one AMD talked about long ago, and according to
ars technica, Barcelona does much more aggressive memory
pipeline reordering now too. That could account for a lot
of it.

Ars technica also says AMD now does the same kind of stack
optimizations that Intel started doing with the Pentium M.
That can be quite a big deal, because breaking the stack
pointer dependencies makes for potentially a lot better
re-ordering capabilities for OoO.

There have been studies that say that IPC on x86 is heavily
limited by stack pointer dependencies under certain loads,
but since those dependencies are often trivial and can be
short-circuited for a lot of cases, having a separate stack
engine that can resolve them can help quite a bit.

Still, being 40% better than Core 2 probably means that
they picked a very specific benchmark. Quite possibly one
where they get the added advantage of the integrated
memory controller, or of the better integration of the
four cores (if there is a lot of data sharing, the AMD
quad core approach should be much better).

Core 2 is good, but the K8 core isn't bad either, and I
suspect that they should be able to make up for most of
the ground they lost with the combination of the full
128-bit ops, the memory re-ordering and the stack engine.
Add a few tweaks here and there, and I bet we'll see some
interesting competition again.

Core 2 has been so dominant lately that getting a
bit of competition must be a good thing ;)

                  Linus