[geeks] sun ultra 24

Mark md.benson at gmail.com
Wed Oct 17 13:26:46 CDT 2007


On 17 Oct 2007, at 17:05, nate at portents.com wrote:

>> My Mac has fully buffered ECC memory.
>
> Yes it does, because it is multi-socket Intel.  And FB-DIMMs currently
> consume a lot more power than non-FB ECC DDR2.

Yeh, may be one reason the Mac Pro has a 1.25kW PSU :)

> 8GB of non-FB DDR2 memory uses 36.94 watts under full load vs.  
> 101.2 watts
> for 8GB of FB-DIMMs.  When idle, non-FB DDR2 memory uses 14.32  
> watts vs.
> 83.34 watts for FB-DIMMs.

Yeh, the RAM is by FAR the least thermally efficient part of the Mac  
Pro.A quick look at iStat Pro in Dashboard reveals the FB-DIMMs are  
running at 49 (working) and 38 (idle) Celsius per module. I've only  
just woken it from sleep too - they usually sit at around 52 Celcius.  
The only other component that is approaching that temperature is the  
Northbridge at 52 Celcius - but that's more than understandable given  
it's the central bus controller.

> This adds up quickly in data centers where you have lots of memory
> installed in lots of servers.

I would estimate the 2GB (4x512MB) I have in mine accounts for a good  
quarter of it's consumption.

> FB-DIMMs do have some advantages (latency not being one of them,  
> and their
> performance issues are largely masked by the massive on-chip L2 caches
> that Intel uses as well as a lot of extra work they put into their  
> memory
> controllers).  FB-DIMMs can be set up to have failover modules so in a
> situation where a module goes bad in use, the motherboard can stop  
> using
> it and switch to it's mirrored module (similar to RAID) without  
> missing a
> beat.

That's very impressive. RAM is one area that really can cause serious  
problems if a module fails, and failover like that makes a lot of sense.

Something I have notice, which may be related, is that the FB-DIMM  
controller in the Intel 5000X chipset is 4-channels in 2 banks. Bank  
A has 2 and Bank B has 2. This is clearly illustrated in the Mac Pro  
as it has 2 Risers, one for each bank. By default Apple only fill  
slots A1 and A2. RAM must be fitted in pairs, so this is the minimum  
they can get away with. Effectively you only get half the RAM  
bandwidth on an Apple stock system as a result (but it costs Apple  
less - of course). Thus the first thing most folk do is fill B1 and  
B2. However it is recommended that you match B1 and B2 with the same  
size of FB-DIMMs. I was going to sling 2x 1GB in B1 and B2 but in the  
end I put 2x 512MB in instead and saved the cash. Apparently this  
improve performance a shade and ensures the machine runs smoothly. I  
have to say the difference is ASTOUNDING in daily use. Not only does  
the extra 1GB of RAM give me extra headroom for apps, the extra pair  
of DIMMs seems to have made the whole machine a shade more perky.  
It's probably in the 1-10s of percent, but it does make a difference.

While the Ultra24 looks like a nice 'light' workstation (pay  
attention Mr. Jobs - you need one of these in your Professional  
product lineup!!), in order to put up a comparable x86-64 workstation  
that looks favourable alongside Dell's Precision 490 and 690, and  
Apple's Mac Pro, I think they must provide a good workstation based  
on the multi-socket 5000X board. Maybe they are waiting for 45nm CPUs  
though?

> And there are companies working on new FB-DIMM modules which will
> use less power and generate less heat, and should be coming to  
> market in
> the not too distant future.

An Apple will likely be one of the first to use them. Including 200  
grams of high grade aluminium on a RAM DIMM makes them RATHER expensive!

> However contrary to what some people have written on this list, I  
> don't
> consider 36w vs. 101w a negligible power difference.

It is a bit of a letdown, but the performance returns on Intel multi- 
CPU Xeon hardware is pretty well worth it at least in a workstation.  
In a server farm you'd possibly get a few extra AMD Opteron systems  
in that spare thermal bracket which may well out weight the losses in  
performance.

-- 
Mark Benson

My Blog:
<http://mdblog.68kmac.org>
68kMac.org:
<http://www.68kmac.org>
Visit my Homepage: <http://homepage.mac.com/markbenson>

"Never send a human to do a machine's job..."



More information about the geeks mailing list