[NTLUG:Discuss] Re: athlon 64 3200+ -vs- P4 3.4Ghz -- NUMA/HyperTransport v. AGTL+, /lib64 v. /lib

[Bryan J. Smith]

On Sat, 2004-11-06 at 07:33, Jack Snodgrass wrote:
> Fry's

We could really use those east of the Mississippi.  :-(

> has a athlon 64 3200+ motherboard CPU for around $240

I assume it is a Socket-754.  It's really worth the extra $50 to go
Socket-939 IMHO.

I'm personally waiting on PCI-Express mainboards to come out for
Socket-939 late _this_ month.  The AMD NUMA/HyperTransport platform has
so much raw throughput, yet a cheap mainboard with only a single
0.125GBps PCI bus for _all_ PCI/LPC functionality wastes so much of that
power.

Although this is an example of a "cheap" dual-Opteron mainboard, take
away the second processor, and it is _very_typical_ of today's A64
mainboards:  
  http://www.samag.com/documents/s=9408/sam0411b/0411b_f6.htm  

You now have ATA, NIC and Audio fighting over the same I/O interconnect.
That's not good, especially when today's ATA and NIC can burst 50MBps.
The AMD8151+AMD8111 combination leaves so much I/O untapped.  The only
alternative today is to pay $400+ bucks to get the full AMD8151+AMD8131+
AMD8111 with (2) PCI-X busses, but that's overkill for desktops.

PCI-Express will solve the issue nicely by adding dedicated 0.125GBps
PCI-Express x1 channels.  A lot of people focus on PCI-Express for
graphics, but the _power_ it brings is in segmented I/O for ATA and
NIC.  Here's an example of the new nVidia nForce4 (labeled CK4"Pro")
that comes out late this month for Socket-939 (just remove the 2nd
Opteron and AMD8131 on the left):  
  http://www.samag.com/documents/s=9408/sam0411b/0411b_f7.htm  

Now instead of a measly 0.125GBps, you're getting almost 1GBps of I/O
_in_addition_ to the video, through an 8.0GBps (16+16 at 2000MT)
HyperTransport interconnect that is also _separate_ from the 6.4GBps of
a _true_ 128-bit dual-DDR memory.  The result is the aggregate
front-side throughput (AFST) is 14.4GBps for the Socket-939 A64:  
  http://www.samag.com/documents/s=9408/sam0411b/0411b_t1.htm  

> and a P4 3.4Ghz with ( cheap ) motherboard for about the same price. 

Is it the original P4 or the new Prescott core?  The latter is slower
MHz for MHz.

> It's a good deal on the 3.4Ghz CPU.... I've been tempted to get one of 
> those AMD 64bit chips though.... 

AMD NUMA/HyperTransport is powerful on its own, especially for servers
or workstations.  If you game, AMD roasts P4 now at everything.  If you
do engineering simulations, AMD gives you SSE speed with FPU accuracy as
its uses its FPU for compuations -- Intel uses "lossy" math in its SSE
pipes, results are not guaranteed to be precise (reproducible) if you
use the SSE pipes.

The 64-bit capabilities are just in addition.  On a server/workstation
with more than 4GB of RAM, _only_ AMD can guarantee memory-mapped I/O. 
Even the Intel IA-32/EM64T implementation of AMD x86-64 is still a
"32-bit" platform outside the chip.  AMD x86-64 is a true 48-bit (the
most compatible with the i486).

Again, what does this mean on the desktop?  Not quite as much.  But
"cheap" mainboards really hurt when you have ATA, NIC and Audio
contending over the same, shared PCI bus.  If you wait less than a
month, and spend less than $100 more, you can go Socket-939 and see a
significant boost in performance -- as well as reduced I/O contention. 
ATA and NICs are fast enough today to cause such.

> does anyone have any real-world experiences with the 64bit cpus and 
> linux?

The kernel runs in LONG mode, it can run either 32-bit or 64-bit apps.

64-bit apps use LONG mode, and that means they _must_ use 64-bit
libraries.  These are in /lib64.

32-bit apps then use 32-bit libraries.

64-bit apps can_not_ use 32-bit libraries and vice-versa**.  If an
application may use closed source 32-bit libraries/plug-ins (e.g.,
Mozilla), then it ships with a 32-bit version.

The 64-bit version of UT2004 show significant performance increases over
the 32-bit version on Linux.  AnandTech is where I read the review.

[ **NOTE:  Actually, this is what XP 64-bit does, via WoW, Win32 on
Win64.  The result is that not only does UT2004 32-bit beat UT2004
64-bit on their respective XP releases, but UT2004 32-bit also beats
UT2004 64-bit on 64-bit XP according to AnandTech, because of those
"translation" calls through WoW.  For those of us around the release of
"pure" Win32 NT 3.1, this is the same performance issue.  The reason? 
Same as in the NT 3.1 days, Microsoft's codebase is _not_ very portable,
and porting Win32 to Win64 is _impossible_. ]

> I don't mind re-compiling my kernel and favorite apps...

There are x86-64 distros.  They typically handle the "mess" of making
sure apps are linked against /lib64 and /lib appropriately.  It's not
easy to handle all that.

> is it worth it....

Worth what?

The design of AMD NUMA/HyperTransport?  Hell yes.
64-bit?  Depends on the OS.

What I've been telling people is if they want to run Linux/x86-64, then
go Socket-939 A64.  It's worth the extra money for the added memory
channel directly connected to the CPU, especially for the PCI-Express
mainboards.

But if they run Windows, then go Socket-754 Sempron (lacks 64-bit),
because it will be a _long_time_ before 64-bit Windows offers anything
but performance _degredation_ because of the library portage issues (and
use of WoW in their place).

> or should I just get the Intel CPU and stick with  that? I'm currently
> running a 3.0ghz cpu right now. 

You're not going to see _any_ improvement then.  Especially if that
3.4GHz is a new Prescott core, which is _slower_ per MHz.

If you've got a 3GHz P4, then I'd wait for the Socket-939 PCI-Express
mainboards to come out.


-- 
Bryan J. Smith                                  b.j.smith at ieee.org 
------------------------------------------------------------------ 
"Communities don't have rights. Only individuals in the community
 have rights. ... That idea of community rights is firmly rooted
 in the 'Communist Manifesto.'" -- Michael Badnarik