You are not logged in.

#1 2014-12-18 07:36:03

jt
Member
From: Bermuda Triangle, NC USA
Registered: 2014-05-21
Posts: 1,407

Xserve Proc -> MDD/G4/1.33GHz and doubled L3 Cache . . .

.  .  .  ran across this notion hanging out with the band.

http://macos9lives.com/smforum/index.ph … icseen#new

I can see why Xserve might have ways to put twice as much L3 cache to good use under X than Apple provided for the MDD series processor cards. But for general computing under either OS, could improvement be all that substantial? I wonder if it might do much of anything but look really cool in ASP under OS9?

My notion is that it might not be all that useful under X for anything other than Xserver side of things?

Just curious what you architectural gurus' take on this might be?

Offline

#2 2014-12-18 18:00:33

ClassicHasClass
Member
From: Electron Alley
Registered: 2014-05-26
Posts: 1,099
Website

Re: Xserve Proc -> MDD/G4/1.33GHz and doubled L3 Cache . . .

Comparing servers to servers, my POWER6 has two cores with 64K L1 each, 4MB L2 each, and 32MB of L3 each off-chip. For modern systems where running VMs is typical (my system runs a single AIX LPAR), you need more cache if you have a higher virtual CPU to CPU ratio or you'll get lots of thrashing when it switches execution contexts.

For the Xserve, this was less of an issue in its heyday; its main use for all that L3 was simply processing, especially data work. The closer the data is to the CPU, the faster it chugs through. Even my old dual 1.25GHz 7450 MDD card got close to my current dual 1.8GHz 7447A in some tasks because the 7450 had L3 and the 7447A doesn't. (On tasks that are not as data-bound, the 7447A crushes it, however.)

The G4 is especially sensitive to L3 because of its fairly slow FSB, even for its day. If the bus is faster, you can simply drive up L2 to achieve the same effect, and some systems like the HP PA-RISC (admittedly due to somewhat impaired cache latency) had incredible amounts of L2 cache. The C8000 workstation on my shelf has dual PA-8900 CPUs at 1.1GHz, with a breathtaking 1.5MB per core of L1 (!) and 64MB each (!!) of L2. Even the G5 does this to some extent; no G5, not even the Quad, has L3, though I have long argued the 1MB of L2 each core has in the Quad isn't enough either.

Offline

#3 2014-12-18 22:08:13

jt
Member
From: Bermuda Triangle, NC USA
Registered: 2014-05-21
Posts: 1,407

Re: Xserve Proc -> MDD/G4/1.33GHz and doubled L3 Cache . . .

ClassicHasClass wrote:

For the Xserve, this was less of an issue in its heyday; its main use for all that L3 was simply processing, especially data work. The closer the data is to the CPU, the faster it chugs through. Even my old dual 1.25GHz 7450 MDD card got close to my current dual 1.8GHz 7447A in some tasks because the 7450 had L3 and the 7447A doesn't. (On tasks that are not as data-bound, the 7447A crushes it, however.)

So if it's addressable under OS9, a second MB of L3 would be great, but likely not as big a bump had Sonnet or someone 50% additional L2 to the processor card upgrade?

Large VidCap, Sound, Graphics, Spreadsheet and Database file sizes processes would benifit from the doubled L3?

THX, keep it coming! wink

Offline

#4 2014-12-19 01:19:24

ClassicHasClass
Member
From: Electron Alley
Registered: 2014-05-26
Posts: 1,099
Website

Re: Xserve Proc -> MDD/G4/1.33GHz and doubled L3 Cache . . .

More L2 is always better, but more L2 may not fit on a die (remember at the time the G4 was made, we were still at 180-200nm process sizes), so bring on the off-chip L3 when yield or cost becomes prohibitive, which is less integrated but better than hitting the SDRAM.

Nowadays even L3 is on-die in many designs, but it's still much less integrated with cores than the L2. Frequently it refers to a block of cache shared by all cores, such as in Haswell-E.

Offline

#5 2015-01-20 08:20:51

dogknight
Member
Registered: 2015-01-20
Posts: 1

Re: Xserve Proc -> MDD/G4/1.33GHz and doubled L3 Cache . . .

Back in OS 8 days, L2 cache made a huge difference. I added a 256k L2 cache card to my old Performa 5260/120. It was touted as increasing performance by as much as 25%. I wasn't expecting much but can honestly say the performance increase I got would have been around the advertised 25% mark.

The classic OS benefited greatly from extra cache. Whether or not it recognizes L3 cache I could not tell.

The thing that annoyed me about the MDD was the slow system bus speed. At the time, they were advertised as being Pentium 4 killers. And I got similar performance from my MDD Dual 1.25ghz as I did my 2.4ghz P4. However the MDD only had the 133mhz FSB when the P4 was running at 800mhz FSB. I can only imagine the difference in performance if Apple had shipped it with even double the FSB speed that they did.

Offline

#6 2015-01-20 15:54:02

cc333
Member
From: North S.F. Bay Area, CA
Registered: 2014-05-23
Posts: 571

Re: Xserve Proc -> MDD/G4/1.33GHz and doubled L3 Cache . . .

You must've had the base model MDD; the higher end models had a 167MHz bus FSB. Still not much when compared to 800 MHz, but better than 133?

c


Main Macs: Early '09 Mac Pro, Mid '12 MacBook Pro 13"
Secondary Macs: Early '08 Mac Pro, Mid '12 MacBook Pro 15"
Playthings: Mac SE/30, 3.0 GHz Mavericks-based HackServe, Many others....
Desired: Lisa, Kanga PowerBook G3, Apple IIc, Apple II, Spare parts, etc.

Offline

#7 2015-01-20 15:54:44

ClassicHasClass
Member
From: Electron Alley
Registered: 2014-05-26
Posts: 1,099
Website

Re: Xserve Proc -> MDD/G4/1.33GHz and doubled L3 Cache . . .

It does (at least, it did with my MDD, before I put the Sonnet upgrade in it which has no L3). However, the dual 1.25GHz systems (of any generation) are 167MHz FSBs.

Offline

#8 2015-01-20 17:06:40

Eudimorphodon
Member
Registered: 2014-09-02
Posts: 525

Re: Xserve Proc -> MDD/G4/1.33GHz and doubled L3 Cache . . .

dogknight wrote:

The thing that annoyed me about the MDD was the slow system bus speed. At the time, they were advertised as being Pentium 4 killers.

The MDD era was the peak of Apple's advertising outright lying making rather optimistic performance claims based on a very specific set of application benchmarks, IE, mostly a small set of Photoshop filters, etc, that were *heavily* optimized for Altivec and didn't hit the RAM or disk I/O subsystems very hard. If you compare more mundane real-world tasks like crunching spreadsheets or compiling code G4s usually come off looking a lot worse.

(Okay, I lied, the MDD was the peak of advertising spin when it came to desktops; they kept pretending the G4 was the best thing since sliced bread for laptops right up until the bitter end. Which I suppose they had to but it's still pretty sad how much something like a Dell D600 from 2003 outperforms the very last PowerBook G4 model from late 2005. Apple's advertising solution to this problem was to pretty much ignore the existence of the Pentium M.)

Last edited by Eudimorphodon (2015-01-20 17:09:01)


Flap Different.

Offline

Board footer

About ThinkClassic

ThinkClassic specialises in the maintenance, repair, restoration and modification of Vintage Apple and Macintosh computers. Ask questions and find answers about classic Apple desktops, laptops, accessories and peripherals.