Re: G4 vs. Intel/AMD Performance
- From: Michelle Ronn <micron@xxxxxxxxxxx>
- Date: Tue, 22 Nov 2005 07:16:58 -0800
On 2005-11-22 03:56:00 -0800, "James Boswell" <jamesboswell@xxxxxxxxxxxxxxx> said:
Michelle Ronn wrote:Netburst is not Intel's biggest problem. Speed and latency with the memory controller is.
I'd say they both are, because of the interconnect with the memory controller, netburst is difficult to keep saturated, so it stumbles on performance, meanwhile the long high clocked pipes drive the thermals skyward.
Pentium M manages very well on the same bus protocol, because it's not an arch built around bandwidth/throughput in the same way as Netburst.
Pentium M does not have netburst (which you may be alluding to here).
Intel's dual core architecture is horrible. It is switched. Only one core can communicate at a time. AMD has a crossbar switch. Non blocking and WAY faster...
Aye, thankfully Yonah and onwards do away with that nonsense.
I have not seen anything out of Intel that supports that statement.
AMD gets to the memory with 60ns latency (local memory). Wost case for AMD is 180ns, if it has to go through two processors to get there. Intel takes 200ns to get to the North Bridge, best case.
your latency numbers seem a tad high?
testing with sciencemark opterons turn in around 50ns and dually Xeons turn in around 130ns
Doing the math on paper gers me these results. There is a 200ns delay to the North Bridge, best case, for Intel. The Opteron is 60ns per proc to memory interface. Again, on paper. Execution order, buffering will mask this.
still a tremendous ass whupping though :p
Then factor in, AMD has 3.5GB of memory bandwidth per processor socket. Quad proc systems can go 12GB sec to RAM. Intel, 3.5GB/sec best case, 1 proc or 4.
3.5GB/s ? where'd you find that figure?
by the specs it's 6.4GB/s per socket for the Opteron and 6.4GB/s for the entire system (using http://www.intel.com/design/servers/boards/se7525rp2/index.htm as a reference) for the Xeons ?
Are you talking stream numbers ?
I got these results running memory tests on Sandra 2005 x64 (I think SP3 of Sandra, just got the update a few weeks ago). Not going on paper math for this one, this is a benchmark. Not a perfect one, but I needed a constistant suite to run against x64 architectures.
Real world. On one of my server tests, an installation of Windows 2003 32 bit takes 40 minutes. On an Opteron, it takes 17 minutes. That memory latency adds up in the weirdest places.
I would be inclined to blame something other than memory latency for a difference as pronounced as that.
Compared an HP DL585 to an HP DL580. The major differences between the two systems were the procs, and HT.
I am consistantly seeing Opteron based systems outperform current x64 Xeon counterparts on OS installs. Drives and array controllers are the same. Memory size is the same. Difference is the procs. The Intel procs even have higher clock rates than the Opterons.
The only thing I can think of is that the Opteron is faster opening up the compressed packages. Not concerned enough about it to dive any deeper.
-JB
.
- References:
- Re: G4 vs. Intel/AMD Performance
- From: James Boswell
- Re: G4 vs. Intel/AMD Performance
- Prev by Date: Re: Apple Hardware Failure!
- Next by Date: Re: MAC Microsoft Advancing Constantly
- Previous by thread: Re: G4 vs. Intel/AMD Performance
- Next by thread: Re: G4 vs. Intel/AMD Performance
- Index(es):
Relevant Pages
|
