Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
G4 7448 and 7457 swap
- Thread starter JoyBed
- Start date
- Sort by reaction score
Mostly to satisfy my own curiosity I timed how long it took to compile TFF from source. There is a thread on doing that here:
forums.macrumors.com/threads/the-unofficial-tenfourfox-development-toolkit.2291789/
I was wondering how much of a difference the L2/L3 cache sizes would make for this task, so I'm repeating the build on the same computer after changing out the CPU for each test. So far, a 1.6GHz 7457 with 2MB of L3 cache is faster than a 2GHz 7447. I'm running the 7448 right now, and I want to also try a dual CPU. Each test takes a while to run but I should have some numbers to update this thread with before too long.
I've been trying to setup the build to run two processes on a dual CPU and a single process on a single CPU, but it's hard coded to 2 somewhere that I haven't figured out yet. So these results are for two simultaneous processes on a quicksilver with 1.5GB of RAM and the original 80GB IDE disc:
4.1h 1.6GHz 7457 2MB L3 Dual CPU
8.1h 1.6GHz 7457 2MB L3
9.4h 2.0GHz 7447
So at least for gcc software compiling, the L3 cache is a big factor. This may also show the advantage of L3 on a busy machine with lots of different CPU actvity (because of the 2 processes sharing a single CPU in this test). I suspect that running a sinlge process may reduce this advantage.
I ran the 7448 at the same time on a different machine, but it has a SSD that I think may be an unfair advantage for this task because there is a lot of disc activity. So now I'm running it again on the same machine so that the results should show only the differences due to the CPU.
4.1h 1.6GHz 7457 2MB L3 Dual CPU
8.1h 1.6GHz 7457 2MB L3
9.4h 2.0GHz 7447
So at least for gcc software compiling, the L3 cache is a big factor. This may also show the advantage of L3 on a busy machine with lots of different CPU actvity (because of the 2 processes sharing a single CPU in this test). I suspect that running a sinlge process may reduce this advantage.
I ran the 7448 at the same time on a different machine, but it has a SSD that I think may be an unfair advantage for this task because there is a lot of disc activity. So now I'm running it again on the same machine so that the results should show only the differences due to the CPU.
Here's an update with the 7448 result. The cache seems to make a big difference on an "overloaded" system:
4.1h 1.6GHz 7457 2MB L3 Dual CPU
7.5h 2.0GHz 7448
8.1h 1.6GHz 7457 2MB L3
9.4h 2.0GHz 7447
The extra L2 cache on the 7448 shaves almost 2 hours off the build time.
4.1h 1.6GHz 7457 2MB L3 Dual CPU
7.5h 2.0GHz 7448
8.1h 1.6GHz 7457 2MB L3
9.4h 2.0GHz 7447
The extra L2 cache on the 7448 shaves almost 2 hours off the build time.
Thanks for taking the time to do these real world benchmarks, I have to say, this is pretty much what I expected to see.
800Mhz QS without the L3 cache?I put the original 800MHz CPU back in and it's grinding away... any guesses?
I'm going to say between 17-22 hours.
Just for reference my QS Dual 1ghz with 1MB L3 cache built the Linux kernel under Debian in 17 hours, my Powerbook 12" 1.5 Ghz 7447a with 512k L2 cache built the same kernel in 27 hours, while my x86_64 AMD Ryzen 1700 8 cores running @ 3.75 GHZ with 16 threads 16MB L3 cache built the same kernel in about 45 min.
Last edited:
Seeing these tests, I'm debating on whether or not it's worth creating a 745x-744x interposer board, allowing you to install 7448 CPUs on machines such as a PMG4 MDD, or just installing 7457s. The only issue I've ran into with the latter option is finding faster L3 cache chips, allowing you to clock the CPU faster (the L3 cache chips run at a ratio based on the CPU clock). If I could find some ~350 MHz rated L3 cache, I could probably get the 7457s running at around 1.6 GHz or so. I did make a video about my most recent upgrade here if anyone cares to take a look:
a Potential workaround for the L3 cache issue that I would try is to see if you can reprogram the SPD type EEPROM that are on these cards, with appropriate firmware you can increase the clock to cache ratio which would mean for a given clock speed the L3 cache Chips are run at a lower speed which would then allow you to clock the CPUs faster without hitting the L3's own clock limit so quickly
a 1.42Ghz Firmware would give you a clock ratio of 6:1 and allow you to clock further with existing chips
here is a link to where you can find a dump of a 1.42Ghz MDD cards "SPD" EEPROM
https://68kmla.org/bb/index.php?threads/building-the-fastest-power-mac-9600.31076/#post-331668although I do think a BGA483 745x to BGA360 744x interposes would be very neat! I do know that theres 1 person who has managed to make such an interposer but i dont think any details have ever been shared about it sadly
it would be really fun to see if one could install a 7448 via one of these interposer cards onto a Sonnet 7455 PowerMac 730-9600 Upgrade card
and make a 7448 Power Macintosh 9600
Last edited:
Oh, yes! I had a 1.25 GHz card I dumped the SPD EEPROM from, but not a 1.42. I'll have to try that with my 250 MHz L3 chips. However I noticed on the 1.25 GHz card, it has 300MHz-rated chips.a Potential workaround for the L3 cache issue that I would try is to see if you can reprogram the SPD type EEPROM that are on these cards, with appropriate firmware you can increase the clock to cache ratio which would mean for a given clock speed the L3 cache Chips are run at a lower speed which would then allow you to clock the CPUs faster without hitting the L3's own clock limit so quickly
View attachment 1921614
a 1.42Ghz Firmware would give you a clock ratio of 6:1 and allow you to clock further with existing chips
here is a link to where you can find a dump of a 1.42Ghz MDD cards "SPD" EEPROM
although I do think a some BGA483 745x to BGA360 744x interposes would be very neat! I do know that theres 1 person who has managed to make such an interposer but i dont think any details have ever been shared about it sadly
it would be really fun to see if one could install a 7448 via one of these interposer cards onto a Sonnet 7455 PowerMac 730-9600 Upgrade card
and make a 7448 Power Macintosh 9600
Ooh 300Mhz chips thats nice should give you even more potential main CPU clock speed
just one thing to keep in mind of course when swapping SPD EEPROMs/Firmwares about is to that some CPU cards have 1MB of L3 per CPU while others have 2MB
and even of the same speed they can be differently configured for maximum confusion LOL, namely the Firewire 400 Dual 1.25's had 2MB of L3 but the Firewire 800 1.25's had 1MB (because the expensive 2MB config was reserved for the high end 1.42)
Found some 300 MHz 1MB chips. I'll install them onto my card, flash that SPD EEPROM dump, and see how fast I can get it to run.Ooh 300Mhz chips thats nice should give you even more potential main CPU clock speed
just one thing to keep in mind of course when swapping SPD EEPROMs/Firmwares about is to that some CPU cards have 1MB of L3 per CPU while others have 2MB
and even of the same speed they can be differently configured for maximum confusion LOL, namely the Firewire 400 Dual 1.25's had 2MB of L3 but the Firewire 800 1.25's had 1MB (because the expensive 2MB config was reserved for the high end 1.42)
Do you happen to know how many processes were used in each case? I suspect trying to run a bunch of simultaneous jobs on a single CPU bogs it down with context switches that make the cache less effective. Also, was the powerbook at half speed (DFS)?
One more comment, which CPU does your quicksilver have? The stock dual 1GHz Quicksilvers usually had 7455 chips with 2MB of L3 cache.
Linux has it's own code for DFS mode, the machine was DFS_HI when I compiled the kernel and I set it to one process, as I have before experimented with running more than one process one single cpu units building the kernel and it slows things down a bit.
On the Dual 1Ghz I ran it with two processes and 16 on the Ryzen.
It was a stock Dual 1Ghz QS so it must have been 2MB L3.
The 2MB L3 cache upgrade was successful, and I was able to get the 7457s on my card running at 1.5 GHz. Made a video about it here:
Updated list with the stock 800MHz CPU:
4.1h 1.6GHz 7457 2MB L3 Dual CPU
7.5h 2.0GHz 7448
8.1h 1.6GHz 7457 2MB L3
9.4h 2.0GHz 7447
19.4h .8GHz 7455
4.1h 1.6GHz 7457 2MB L3 Dual CPU
7.5h 2.0GHz 7448
8.1h 1.6GHz 7457 2MB L3
9.4h 2.0GHz 7447
19.4h .8GHz 7455
How did you get your 7457s running at 1.6 GHz? I couldn't manage to get mine running well at that speed (which may be due to the older VRM design on the older card I was using).
I'm not sure why exactly some CPUs will overclock better than others. I just watched your video about changing the chips on the MDD and I'm curious where you got the mc7457. I'm guessing the sc7457 came from that china ebay seller who was pulling parts from scrap boards. I suspect that if you try, the mc7457 will clock to 1.6 (or higher). I had my dual quicksilver with mc7457 chips running at 1.73 for a while, but at 1.6 I can use a lower voltage and a much quieter fan.
So I would suggest trying to set the speed higher on that one mc7457 chip and see if it works. Also, once you start upping the voltage cooling becomes important. If you're at the speed limit for a given voltage, then you can up the voltage to try for more speed. This also increases the heat/power, and thus temperature... which can then defeat your attempt at higher speed. On the MDD, the CPU is sort of floating between the logic board and heat sink. If you put in shorter chips then you might need to add copper shims to maintain good contact between the silicon and heat sink.
So I would suggest trying to set the speed higher on that one mc7457 chip and see if it works. Also, once you start upping the voltage cooling becomes important. If you're at the speed limit for a given voltage, then you can up the voltage to try for more speed. This also increases the heat/power, and thus temperature... which can then defeat your attempt at higher speed. On the MDD, the CPU is sort of floating between the logic board and heat sink. If you put in shorter chips then you might need to add copper shims to maintain good contact between the silicon and heat sink.
I added a couple more results and made a pretty graph. The dual 500MHz 7400 actually does quite well with gcc. Note that this was a different computer with a slower bus speed, but it was the same physical hard drive and also 1.5GB of RAM as the Quicksilver results. Now I'm curious how a dual 533 would do with its faster bus speed, but maybe I'm getting off topic here.
I saw another powerbook got a 7448. Nice work dosdude1 !
Mate, great videos, great work all round. I envy your skill.
Do you know of anyone who is willing and able to perform a dual 1.25ghz to 2ghz conversion for an MDD?
Doing this requires an interposer board, to allow 744x series CPUs to be installed in place of 745x. I tried making one, but it didn't work. You also lose L3 cache completely. Best thing to do is just stick to a 7457 at 1.5 GHz or so, and upgrade the L3 cache to 2MB if necessary.