This page is a wiki. Please login or create an account to begin editing.


36 posts / 0 new
Last post
SkyCapt's picture
Offline
Joined: 2017 Jan 11
SATA-II (2nd generation speed increase) PCI card in a G4

I've tried doing this and it worked, at around 180 MB/sec flow thru an individual SATA cable, but it was usable only in short bursts of just seconds or a few minutes, prior to crashing the system due to the device driver not playing nice with my G4. Any successes?

SATA-II card is said to be "PCI-X" which carries the minimum requirement of G5 motherboard, but my particular G4 MDD-2003 motherboard (secretly?) is very like the G5. I suspect my motherboard supports the "PCI-X" protocol tho probably at no more than 33 MHz so that my PCI-X bus bandwidth is 33MHz X 64bits = 2133.33 megabits per second aka 266.66 MegaBytes per second. I've used OSX to successfully read and write files at over 180 MB/sec but driver bugs (those belonging to the card?) are crashing me within seconds - bear in mind a 1GB file transfer (from ram) completes in under 6 seconds, so the success isn't trivial. I think older motherboards, that are "PCI" and so know not of "PCI-X", might not post the card's existence or if so, wouldn't begin to operate?

How about SSD 'blade' PCI cards? I know they're popular in PCIe cards but PCI/PCI-X ?

Comments

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

The PCI slots in a MDD are 64-bit (like PCI-X), but they are not fully-cached enough to do this reliably. When I was using a PCI-X raid-controller in mine (for drives bigger than 2TB), it would work, but only if NOTHING ELSE WAS RUNNING (even screen-saver). Also, had to disable one CPU with C.H.U.D. tools. Any deviation from this would lockup Finder and I'd have to force-reboot (and even in some cases had to when doing all of this).

Used it for 2-years like this until discovering External FW800 enclosure. Is likely these slots would have been eventually made more PCI-X compliant, but even with the Sonnet/Firmtek cards that are designed to be both PCI/X compliant, the cards are not being used in 64-bit (according to the company).

Also, since there is no back-side bus on these CPUs, your data path is effectively halved thus greatly limiting transmission rates (some have rumored that the 7448 G4s actually have a backside Bus as well integrated in the CPU card - and this explains their radical performance increases over 7450/7447s, but now that Freescale has moved on and so few of them can be found in the wild this is next to impossible to verify at this point).

SkyCapt's picture
Offline
Joined: 2017 Jan 11

There's definitely a fundamental conflict taking place, but I'm not convinced it can't be solved, maybe by properly-made firmware in the card/computer.

Now I remember the startup chime was either non existent or garbled, but I immediately saw this is a Trick ploy, because the card was indeed Posting and could work briefly. The computer (via firmware) was genuinely trying to chime but some of the card-conflict is with its audio system. Other conflict was with the USB system, not certain the details. I think i had to stop using USB1 and plug my keyboard and mouse into the USB2 card and by that I mean keys+mouse plugged into a USB2 hub first and the hub plugged into the USB2 card, because USB1 devices plugged directly into the USB2 card would still dance with the USB1 system. Maybe I disabled USB1 extension too, letting USB2 software take total control. Then the audio in OSX was still garbled and it'd crash, but now I think I did NOT try disabling the audio extensions. I assumed there was a 3rd or more conflict but maybe the audio system was the thing crashing even though I wasn't using it.

I still got the SATA-II card, FirmTek SeriTek. This is one I'd enjoy re-testing and might start doing that. I'm under the impression it should work with proper drivers and the missing/garbled startup chime was a demented attempt to have me and most everyone give up right away.

The 64bit PCI dates way back to the beginning of G4 era and has nothing to do with 64bit chips. It uses 66MHz X 32bits to operate the equivalent of 33MHz X 64bits.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

It's not a conflict, it's simply not able to send that data as fast as the card it timed to receive it. It wants G5 Hypertransport Bus and you are giving it 1/4 of that. As for the sound problems, when I was shopping around for a replacement for my first raid card I tried two different replacements (one Sil3114 and the other Sil3112) and they caused the same symptoms, but it had nothing to do with drivers. It was borking the open-firmware. I know this because I couldn't get to the choose screen by option-booting or the OF screen with Command+Option+O+F. I had to remove it, Blank the PMU and zap-Pram about 6-times before I could get the chime and other OF systems back. You should be careful - there is a real risk of corrupting your firmware and you could well need a new logic board. If speed is so important, just get a newer machine. Pushing these old critters so far out of their wheel house is just asking for trouble.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

It won't be a hardware timing issue so basic, not one that couldn't be avoided. PCI communications incompatibility would manifest almost instantly and not allow a completed 1GB file transfer. You said the card maker says/admits their PCI G4 mode drops out of 64bits, and that cuts the PCI bandwidth in half to 125 MB/sec, this is the limitation I am finding with my SATA-I cards, an individual SATA-I port does upto 95 MB/sec but a RAID0 pair hits a wall at 125 MB/sec when obviously could go much faster. OTOH the SATA-II card is doing 180 MB/sec with just one SATA port and therefore also 180 MB/s on PCI, it must be utilizing 64bit PCI, and then I believe conflicting with a couple of "landmines" that had been workloaded by designers onto the 64bit half of my PCI bus bandwidth.

SATA-II card is now again installed and being tested. First I'm finding it conflicts with my SATA-I cards in which either the -I or the -II runs, not both at the same time. Don't know yet how to control which cardtype runs, seems random but in favor of the -I. Retesting has only just started so results aren't in. I haven't had any firmware screens go missing nor the need for pmu/pram reset, not in my previous tests nor right now.

Thanks, and I appreciate your inputs. You're right, the firmware could get a "borking" but I think the risk is just in runtime, not endangering hardware permanently.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

One thing that does matter - this comes from Seritek themselves - if you have two PCI cards with the same chipset - they must be using the same Firmware, and often cannot be seated side-by-side. When I got my eSATA card it conflicted initially with the internal SATA that I've been using by Sonnet (same chip), and the rep sent me and updater and I had to update each card in turn and then put one in slot I and the other in slot IV and now they play nice.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

Instructions for flashing all of FirmTek's cards say it can be done in any G4/G5. Trying to use it however, the SATA-II card eventually stopped Posting, that's when I had to run the other tools on the driver CD(s) of which I have two versions.

I keep Slot 2 empty, it's the one atop the graphics card. This helps cooling it since my upsidedown graphics card has no fan; and the slot 2 coverstrip I keep there is one of those swiss cheese types. [MDD PCI slots are named 2 thru 5, slot 1 is what they call the AGP graphics.]

SkyCapt's picture
Offline
Joined: 2017 Jan 11

Done with this. The SATA-II card isn't performing PCI "bus arbitration", the job of drivers. My original description "does not play nice" stands. The way to make it sorta-work is to give it the entire bus, meaning no other PCI cards not even USB2, and I used a USB1 boot volume ugh. Then it mowes down every misc thing trying to use PCI which is a lot of little timers etc, crashing unavoidable. When it is working, the speed is addictive and it is a shame 64bit PCI cards weren't built.

And you were right, the firmware got 'borked', very scary. And then I was right, the damage wasn't permanent. Mine had ceased to complete the booting of any volume, it'd KP on the Grey Apple screen right when the spinner was supposed to appear, it'd even KP trying to boot Tiger Install using physical PATA DVD. The one thing that I think got it back in working order was removing the 3.6V battery from the motherboard - because pmu reset and pram reset were taking but not helping.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

Glad it came back - these old logic boards are really tough thank goodness.

There are a few people who are doing writeups on using PPC hardware for Linux servers and revealed a lot of PCI recocnition comes down to OpenFirmware and that that is one of the reasons G4s aren't quite fully Rohs compliant (like Acard for instance and some Sonnect PCI/USB2 cards that just won't work even with drivers).

Further that is one big advance on G5s, especially the later ones with PCIe. If the drivers exist, they work a LOT better.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

Let me propose a model which describes what's taking place. Certain PCI cards like SATA being discussed can relatively slowly corrupt the computer's runtime state, be that state OpenFirmware and/or Mac OS. Properly made or updated drivers in the computer-end might remedy this, but there exist two sets of drivers. Firmware has built-in basic quality drivers whereas OSX has larger more recent drivers. @Maximum R.I.S.C you say you've used otherwise functional cards which slowly corrupt the firmware processing. Since the computer lives in a world of billions of calculations per second, and, the computer is only briefly in the firmware mode when booting begins, before execution is handed-off to OSX and its drivers, the firmware corruption appears to take effect slowly. Having working drivers in OSX prevents similar corruption from happening during the products use, and firmware will continuously corrupt slowly because G4 firmware updates aren't available.

Apply this understanding to the SeriTek 2SE4 SATA-II I'm investigating, the corrupting influence looks like activation of 64bit channel PCI, which is what I'm looking for, coupled with drivers I don't believe are playing nice. If I yelled to FirmTek get off yer a$s and create a driver that works right, it would become (if they obeyed me) a functioning card with still the tendency to twist in the firmware mode, like some cards you've said you've already observed.

It looks a little bit now like the thing I *WANT* is a card which tends to corrupt my G4 firmware. That could be the trick to knowing it's opening up 64bit PCI and doubling its bandwidth over cards that ignore 64bit PCI. If pram reset is the only necessary workaround, I can do that from the keyboard. If pmu reset is required, it's a bit more of a bother but doable to say the least.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

I think it is exactly the opposite of that. The firmware on the card is doing most all transactions with the main system bus, and drivers intercede for certain functions (like Raids, where the OS simulates the record-keeping a chip work normally do on the $1,000+ Raid cards). I've only ever had to install ONE driver for PCI in 9 Macs I've owned (not counting AGP Graphics cards).

Ironically that driver was the one internal raid-card that worked for the two years before I got the FW800 housing. It was a SIL3224 PCI/PCI-X 2-SATA/2-eSATA card. The good news for you is even though the Raid function was the part that kept messing up, when not raided, it worked pretty stably. They had OSX Tiger Drivers for it. The Bad part is that I suspect what little success I had was at least in part due to firmware, because when I bought it it explicitly stated G4/G5 compatibility - which is the part that is going to vary like crazy - with each vendor having custom firmware mods. The majority of the other cards I tried failed, and messed up the Logic-Board within 2-3 boots or on 2-occasions during big file transfers, so it isn't so much a time factor as much as when certain functions are required the card's firmware wasn't properly setup for.

Even the Hallowed Sonnet TSATA/Seritek cards with the perfect compatibility are just SIL3112 chipsets with GREAT firmware. A guy selling 2 of the cards I tried that didn't work/ or didn't work for long, admitted he was just force-flashing PC cards with the Mac-compatible firmware - but that seems hit and miss and he was honest it was really tough to guaranty bootability.

So there seems to be 4 things at work. The chipset, the firmware, certain hardware-jumpers/or board mods, and finally software drivers, and as you go up that line you have less direct influence on the binary-level functions of the card.

The fastest communications I've heard of for MDDs seem to be from attaching multiple USB-connected drives to a PCI USB2.0 card and then software raiding them together (a few discussions on youtube about this), but it comes with a few issues - like taking a long time to get going and using sometimes ridiculous amounts of CPU to work - so not practical for most functions.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

Interesting stuff, I know how complicated it gets. There's even two firmwares involved, I've used the word firmware only to denote that which is in the G4 motherboard, then I see you're using the word firmware to denote that which is in a PCI card's "flash" programming.

I don't necessarily need or want a working SATA-II card. If I can make a SATA-I card take on 64bit PCI, it'll be a big improvement over the lesser PCI SATA-I cards I'm using now. A single SATA-I port won't talk faster, it's already maxing at 95 MB/s, but SATA-RAID0 and copying files from one SATA port to another SATA port will go a lot faster. Right now, file copy drops to 62.5 MB/s ~ on src and dst ssds that can both do 95, it's because the 125 MB/s PCI-reduced bandwidth is being evenly divided between src drive & dst drive in my controlled tests.

Ya, i wouldn't waste my time doing RAID0 on the USB2 interface. It's a Clunky slow bootup already without RAID, then with RAID it only gets a lot worse. A SATA-III SSD using a SATA-I port is still the fastest in OS 9&X and better than USB2-RAID for a hundred other reasons.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

I noticed that too. Rather than having a transfer between two drive on the same SATA half the speed they actually go faster than single-direction traffic off the card, and the FirmTek guy told me it's because their cards are setup to take most of the data transfer job away from the System Bus and only report package-transfer success reports to Finder's status update function. I remember when I bought one I had the choice between the two-port and a four-port and decided not to get the four because, at the time, it seemed like I would be only getting 1/4 the speed, and I know know it would have been faster then having a dedicated SATA and dedicated eSATA cards both installed.

I've since talked to a few Apple system's guys that explained a fair bit of this. I used to think that the 167MHz bus was available to each PCI, but in fact they all share it. So if one slot is using 40% of the bandwidth, the other 3 each only get 20%. The only bus not draining from this pool being the AGP bus that is completely separate. So it actually makes sense to let one PCI slot handle as much as possible, if it has the bus-boost, execration functions the Seritek/Firmtek/Sonnet cards do.

This is a big contrast to the way bridges multiplex in PC motherboards, that actually allow much more complex traffic. PPC Macs, in truth are not nearly as powerful, they are just much more efficient.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

PCI slots (not PCIe) are on a shared bus, meaning the slots compete against each other for PCI bandwidth. But total PCI is in G4 only a small fraction of the "FSB"/CPU bus. In the 167MHz MDD, PCI maxing amounts to only 20% of the FSB so it doesn't interfere with CPU activity. 33 MHz PCI began around in the days of G3 when the combination of PCI and FSB was like its own shared bus, the same bus. Those early PCI systems were very performance limited, however, the last PowerMac G4, MDD-2003 has no problem running circles around the 33 MHz x 32bit PCI , to offer some challenge it needs the full 33 MHz x 64bit PCI interface that's found on all PowerMac G4 dating back to 1999. We agree the AGP bus has its own unique pathway. PCI-X in the G5 increased the PCI speed to 100/133 MHz by introducing a dedicated PCI controller built using G5-era chip technology (high bandwidth).

So why do card makers/apple abandon the G4's 64bit PCI? They seriously want me to believe the G4 can't handle it?, can't handle something it had been outfitted with. The way the 64bit mode PCI card I tested went and gave my G4's firmware a seizure looks like one of those aforementioned "landmines" by-design, or design-revision. They seem to be playing up the superstition that a G5 PCI card capable of 100/133 MHz and 64bit interface might Harm the 33 MHz system or require 64bit chip tech to drive the 64bit interface, but I know better. It is some relief to find the G5 card officially can be "flashed" when installed in a G4.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

Actually I think if you ad-up the math it makes perfect sense. The G4s have only one pipe into the CPU (Front-side Bus), so in the case of a 1.42GHz, that multiplexes to half that value on average (input+Output) to a stream of 710MHz. Half of that goes to RAM (333MHz in this case), 25% to AGP (4x speed, or equivalent to 4 normal PCI slots) and that leaves about 25% to the rest of PCI/ATA. Now on incoming/outgoing to RAM it can go higher, but only till that 'charge' of RAM is exhausted, then it must refill before it can output again. You can see this on any traffic going across system bus in System Monitor. It isn't constant, it's pulsed as RAM fills and discharges. Now if the signal (like a video conversion) isn't RAM heavy, it can input/output simultaneously, but for big file-transfers you really feel this.

The exception is in the transmissions going within those SATA cards or via the FW800 card because they have their own cache and transmission gates and while they still pulse, it's much faster and very much gives the appearance of steady read.

Now imagine that on a G5 or Intel with a Backside bus (faster of course) where you can read and read/write simultaneously to a MUCH faster bus, and in the case of the Intels 8 or 12MB of L3 cache to buffer any holes in the stream.

I'm not saying G4s are not still awesome (hell I'm typing this on one), and they are nothing short of AMAZING compared to any Intel of the same clock-speed because of the R.I.S.C. performance advantage (now emulated in Intel chips B.T.W.), BUT they are not going to be able to beat a modern machine, and you'll find the edges much faster.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

That's not how it works. The "one pipe" FSB on G4 is huge, 167MHz X 64bits = 1.333 GigaBytes per second, and the new bidirectional traffic feature added to the CPU-FSB on G5 is actually a small optimization, the increased FSB speed in G5 is its bigger deal. The G4 PCI when driven to maximum is taxing just 20% of the G4 MDD2003 FSB and that's if it's 64bit PCI. If avoiding 64bit PCI and dropping down into 32bit interface mode, the maximum drain that the G4 MDD2003 FSB suffers from PCI is only 10% of the overall FSB bandwidth, 133 MB/sec. Subtract the typical "overhead loss", then this is the 125 MB/sec limit being measured.

CPU core speed isn't part of the hardware peripheral I/O. CPU core is a device attached to the FSB, same as the PCI in a G4/G5 (pre-PCIe) is also a bus device attached to the FSB. CPU and PCI are competing for FSB and in turn RAM, but they aren't evenly matched they are David vs Goliath.

What matters most is whether any bus is Synchronous vs Asynchronous (anti-synchronous). These are synchronous, the G4 and G5 pre-PCIe devote all their asynchronous Northbridge activity to RAM, AGP, and HyperTransport between the North-South bridges, so the PCI must attach with a synchronous relationship to the synchronous FSB because there is no where else it, relatively high bandwidth, can go. The 64bit PCI in the MDD2003 can get in the way of the CPU-on-FSB by only as much as 20%. The most prominent bottleneck is the 32bit PCI interface. A better 64bit PCI interface doubling the bandwidth that PCI carries was already implemented - & gets left unactivated it seems. 50% of PCI bandwidth is being hijacked from my G4, producing a real bottleneck more pronounced than the G4 FSB allegedly is, and I'm indignant about this.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

Sorry man have been writing Apple code for a decade and dealing with some very bright minds on the hardware side and none of what you just wrote is even close to the models I've been shown while optimizing. G4s are very dumb chips this is why they were so fast proportionately - because it's just like a saw-blade with data coming in, that can, in some cases be fed at twice the width (or more for things like color-data in Photoshop), but there is no direction of traffic, the way a C.I.S.C. system does - outside of pool/run/write operations.

P.S. Bandwidth is only an advantage on data that can be vectored with Altivec and only between the CPU and RAM. It is NOT a 64-bit system outside of that.

Have fun trying to squeeze blood out of a turnip - you'll need it!

SkyCapt's picture
Offline
Joined: 2017 Jan 11

The G4 CPU cache drinks in at 64 bits per cycle, filling the caches faster, benefiting arithmetic CPU instructions too, not just altivec. The pre-PCI-X 64bit PCI hardware interface, limited to 33 MHz, was designed to be driven by 32bit(ALU) support chips cycling at a minimum 66MHz cores, successfully started in 1999 with the then-latest chip technology. I thought you would say I'm in conflict with everything else. But on the wikipedia page for PCI (Peripheral Component Interconnect), they state PCI was Synchronous and the PCI protocol matches that of FSBs in use at the time, so it is real that the PCI is a FSB device. In fact it is the lack of dedicated PCI controller that is limiting the G4 FSB to 167 MHz and limiting the PCI to 33 MHz. Insert the PCI controller a la G5 and bingo PCI faster than 33 MHz plus : faster FSB because ~ with the PCI controller, signals are no longer synchronous all the way from the FSB travelling out to the farthest PCI slot intersecting every PCI slot to and from. The G5 PCI controller separates the physical slots from the FSB, even tho the controller itself is still an FSB device like one CPU. It isn't until "PCIe" that the whole thing becomes point-to-point within the northbridge, and asynchronous to the FSB.

Oops, I see there are PCI 33 MHz versions of G5 that have no PCI-X controller, so this wasn't necessary to achieve the G5 FSB speed. Instead the big change was the introduction of asynchronous RAM, substitute "RAM slots" for "PCI slots" in the previous paragraph to explain why the G5 FSB is so much faster than G4.

I've already shown my MDD2003 can do the 266.66 MB/s PCI with its 64bit interface, shown when a 1GB file is written from RAM to SSD in under 6 seconds. It is Drivers not allowing my full PCI and I'll assume this is affecting all G4 series. Who's drivers, Apple, the card maker, both?? Without the right drivers, this project ends, period. But you really think a PCI card can't be made that talks to drives using SATA-II and talks to the host using PCI. There's no law that the drives must be allowed to communicate their maximum.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

The last thing I can relate on this is that the Bit-Width of far less important than the timing. The signal is "End to End" not "Side by Side". You'll find that the clock of Main RAM, adds up to the sum of frequency of all components attached to the Bus X 2 for input and output. The ATA/PCI section is given 1/4 the time of each cycle, AGP 1/4 and RAM half.

To achieve the 64-bit transfers you seek most likely requires a driver rather than simple firmware because it is being packeted into smaller chunks that the CPU would need to reassemble (much like how a G5 had two 32-Bit matched RAM stick side-by-side to provide a 64-bit register).

It's likely that the RAM/caching overhead for this process will over-time diminish the speed returns anyway, as RAM will need to cycle a lot with the CPU to achieve this. This is one of the main reasons why newer systems are focusing so much on processing directly from Data-Storage using error-correction rather than caching for safety because the speed advantage can be as much as 10 to 1 (not to mention energy savings).

Even if you can get this to work, it would most likely be necessary to copy files greater than 2GB (to fully-engage Virtual Memory) to get a final averaged speed.

Wikipedia's info on PPC systems is extremely limited. Sadly most of the tech guides are long gone, and my old pals on the subject have moved onto greener pastures, but there is a lot of high-quality info at http://barefeats.com/ and http://www.xlr8yourmac.com/ (especially when they did early comparisons between G4 and G5 systems. They conclude that G4's through-puts are often half of theoretical limits and most interestingly that the CPU is shocking efficient, (especially when armed with larger L3 caches).

BTW: L3 Caches in a G4 only partially buffers the System Bus. It is also a Buffer for the RISC algorithm systems created by code strings, because such chips don't have dedicated functions as CISC chips do, and effectively rewire their structure to emulate each function required by apps as loaded. If L3 starts having to carry too much Bus traffic, the function-structures get pushed to main RAM (half the speed of L3) and that's why if you have a program sitting in the background it can take several seconds to come back to life, as the algorithmic structure spools back into L3, where it waits to be called on by L2. The 7400 G4s had very big L2 and often no L3, and a dual 533 MHz CPU Digital Audio could outperform a newer 7450 G4 at 1.25GHz on Altivec powered apps, but as copying bigger files (like from Video cameras) become more important, L2 was shrunk, L3 got BIG and RAM went from 133MHz to 333Mhz to provide backstop for juggling not only CPU operations, but also Big Finder/File-Moving tasks, But these newer arrangements are in some ways a fair bit less efficient than doing one or two things at the same time, as we were doing with OS9 when G4s first came out.

In short Multitasking has a big cost. Newer systems are made for it, but G4s had to learn it as a second language.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

(reseting this forum's page indentation...)

SATA-II PCI card, even if using 133 MB/sec, will offer improved drive performance on all G4 - it won't be a whopping two fold increase like they hope the name SATA-II vs SATA-I correctly conveys in most cases. And then to correctly utilize 64bit 266 MB/s PCI would make it work a lot closer to how intended, about 180 MB/s per port, but RAID0 and file-copy would run below their SATA-II maximums the same way my 133 MB/s SATA-I cards do. Because "-I" can do all that, "-II" shouldn't be any trouble.

Anyway, I love the talk about G4 vs G5 architecture. In my Original Post (OP) this page, I mentioned my particular MDD2003 motherboard is very like the PowerMac G5 made that same year. I have the G5 northbridge/RAM controller chip and its newly introduced two channel asynchronous memory system. My motherboard is, in the shape of G4 MDD, similar to the 1.6GHz Single CPU 2003 PowerMac G5 in which both are without the PCI-X controller chip therefore 33 MHz PCI / PCI-X? and in which both are one CPU socket and 4 memory slots not 8. I'm like a G5 motherboard with a G4 CPU socket and I don't believe my RAM slots have any matched pair relationships. I got numbers to prove this too.

You know the first G5 introduced asynchronous RAM because the model with 800 MHz FSB & 333 MHz RAM has no synchronous relationship between the two parts. A synchronous relationship would be a ratio that is whole number, n.5 second best, etc and 800:333.3333 is 2.4 which if synchronous would cause lots of wasted cycles while the two ends of the bus keep Waiting for alignment. The fact that the G5 FSB speed can change independent of RAM should also be a strong clue. 400 MHz RAM is running at exactly that speed in both the 900 MHz FSB and 1000 MHz FSB G5s, while on the other hand everyone knows the very synchronous MDD-2002 will make a RAM speed change to accommodate a change in FSB speed. They made sure everyone knows the S in SDRAM/DDRSDRAM stands for Synchronous, I remember TV commercials with a little boy telling dad he must get 533MHz SDRAM and it made an impression on me. Then when the architecture changed to become Asynchronous they told nobody and it's still being called SDRAM today.

The two channel isn't hard to prove. Wikipedia, I know, lowest common denominator a lot, states the PowerMac G4 MDD CPU has access to only 50% of total RAM bandwidth and "the rest is available to the graphics card" meaning AGP. 64bit wide DDR modules are really supportive of two channels 32bit each. One channel synchronous systems simply convert 167 MHz X 64 bits into 333 MHz X 32 bits and ignore the 2nd channel. Dogs don't know it's not bacon. But working two channel asynchronous RAM began being made at this time. You know the AGP is a 32bit asynchronously interpreted bus right, there is an extremely large chunk of silicon the AGP interface, used to be on its own chip before integration into the northbridge. The 32bit AGP interface is as big as a CPU, and the addition of the channel two 32bit interface sitting between RAM and FSB turns the northbridge into something like dual-core 32bit CPUs. The 2nd large core in the northbridge is the "DDR interface" which everyone said needed to be there but was left out on G4. It required a leap in northbridge chip making technology (a G5). While channel one that which I'd call FSB is talking to RAM 32bits per cycle, channel 2 AGP can also talk to RAM 32bits per cycle simultaneously. No longer does the GPU wait for the CPU to release control of RAM, no longer does the CPU wait for the GPU to release control of RAM. It can cause "DDR" RAM to look twice-as-fast as ordinary RAM pre-DDR. And in case they had to try extra hard to hide the truth, the first DDR was 266MHz 2x faster than the final pre-DDR RAM of 133MHz. So (ahem) in answer to the question what makes DDR RAM with a proper DDR interface get the computer going two times faster than that old undesirable RAM? 266 is 2x 133, isn't it obvious (/end of sarcasm).

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

You'd be far better at arguing in a court of law (where empirical facts matter less then perception) than dealing with computers ;0p, but I see how it might seem this way, without better info.

You can find more on the entry-level G5s here (and yes they ALL use matched RAM, even the SP 1.6 GHz) http://lowendmac.com/2003/power-mac-g5-2003/.

You are right that the base 1.6GHz model is similar to the MDD in that it uses 33MHz PCI, but its Bus is still 5X as fast and has Front + Backside bus - effectively doubling bus-efficiency over G4s (but ironically in this case, most of that Bus-Speed truly is Wasted on on PCI and would most likely only translate to better Firewire speeds, since the lack of CPU power would even choke USB 2.0 - the SP 1.6 G5 is often termed a "Road Apple" because of how poorly-designed it was). It's Bench value is about 30% slower on in all areas, than the DP 1.42 G4 because of reduced Altivec efficiency.

The DP 1.8 G5, with True PCI-X is MUCH better at file-transfers (this is actually what you are looking for) but is only 36% faster than a DP 1.42 G4 at CPU operations, because processor-vectoring has next to nothing to do with Bus Functions and 32 vs 64-bit system bandwidths.

PPC systems ultimately failed to survive because of the overall passive strategy in how RISC threading is handled. In tests between the most powerful G5s (the DP 2.7 and Quad 2.5 and Intel Macs of the same speed-range for functions like QuickTime Rendering), at most - a particular task can only use about half of the CPUs' power compared to an Intel that can focus upwards of 90%. What this translates to in real-world terms is that the best Powermacs can change-gears very fast and multitask pretty well, but never were able to regain their tight application focus achieved initially with G4s in OS9. So basically Apple was chasing a goal-post that Intel was always moving.

But if you are looking for basically a PPC server, one of the more reliable models like the DP 2.3GHz models, would be a good fit because they have full-utilization of that faster bus for shoving data around, and enough CPU to do some actual work should you get around to needing that.

P.S. G4s don't technically even have a Northbridge controller. If anything the board controllers on G3/G4 systems are much closer to being a Southbridge (completely passive).

SkyCapt's picture
Offline
Joined: 2017 Jan 11

Looks like this discussion will stay all-or-nothing between us. G4 no northbridge? It's on the underneath side of the MDD motherboard and has its own heatsink bigger than the GPUs. Doesn't "heatsink bigger than the GPU" convey its importance? RAM controller is its other name. It was named "Grackle" on early G4s when the AGP interface was a separate chip. When the northbridge and AGP interface were combined into one chip, it became know as the "uni-n", "uni-north", meaning Unified Northbridge chip. Apple also calls this the "System Controller Chip". It physically sits directly between the CPU socket and the RAM slots, because its most important function is that of RAM controller to the FSB.

I change my MDD2003 RAM speed and her FSB speed independent of one another, like G5 can. This is nothing like the Internet tutorials for MDD2002 switching between 133 and 167 MHz FSB/RAM. They (MDD2002) must match both FSB and RAM speeds by switching resistors on both the CPU card and the motherboard. I don't even have that exact switch on my motherboard, I can change speeds without modifying the motherboard! And the speeds I can change into are G5 ! I can do 400 MHz RAM (usually reported wrong) like G5 can do, and I can do AGP-8x like G5 ~ the AGP-8x graphics card required that i physically unlocked it before it would even power on. I measure both these improvements taking effect in spite of the G4 CPU bus being the G4 CPU bus.

I said my mobo is "like" the 1.6 SP G5, but, I know my System Controller Chip is actually like the one they put on 2.0 DP G5 motherboards. The MDD-2002 required a 400-500 MHz SCC like remember when this was a 32bit CPU limit built on the 250 nanometer process, it requires a newer process to both go faster and fit more transistors on the one chip with workable power consumption. So when the 180 nanometer G4 first published, it came in a range 867, 933, and 1000 MHz speeds. The 180nm chip without defects can really do 1067 MHz, Apple wouldn't at the time push it to the edge. Minor defects, like a speck of dust gets in the die during fab, can require the voltage to be raised and the speed reduced. So the 180nm G4 first appeared available in a range 867 to 1067 MHz. This is the G5 System Controller Chip, the speed range of its introduction is responsible for the corresponding 800 MHz FSB limit, 900 MHz FSB limit, and 1000 MHz FSB. And the 180nm chip holds exactly 2x many more transistors at a similar power level. This is because 180nm2 is half of 250nm2, 180x180 is half 250x250, so while the number 180 doesn't look half the magnitude of 250, the 180nm chip holds 2x the number of transistors than 250nm. The 250nm 400-533 MHz G4 MDD2002 System Controller was like one 32bit processor (for AGP) and the 180nm 800-1067 MHz G5 System Controller has 2x as many transistors, dual core 32bit processors (for AGP+FSB to interface with RAM). My MDD2003/G4 has this most important G5 chip, and, has this chip in the then-best 1067MHz version of it. I have the "DDR Interface" and then some. The way I say it works couldn't be more different than you and other sources, but my way not only makes sense it can't be done any other way except for using more modern chips to pretend being old. The official description of "DDR" makes it sound Like Magic, no shit, Like Magic vs obvious common sense.

The RAM I run is FASTER than 400 MHz. My Southbridge activity is a lot less than the actual G5's, so my Northbridge devotes less energy dealing with Southbridge and spends more energy dealing with RAM : 467 MHz RAM on my MDD2003. This RAM increase is out of the range of not only the CPU but the GPU too, HOWEVER the RAM controller itself can use the extra RAM bandwidth, and the RAM controller has blitter instructions built in. When I increase my RAM speed from 400 to 467 MHz, I see the improvement on-screen as the result of blitter operations completing faster, a lot faster since there becomes so much new RAM bandwidth dedicate to blitter-only.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

I've already contributed this concerning the 467 MHz RAM speed in MDD-2003:

http://macintoshgarden.org/forum/audio-gain-level-line-in-mirrordoor2003...

and, Another MDD2003 RAM Tool is its "AHT" v2.0.2 and v2.1 Apple Hardware Test.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

While "The Math" on all of this just doesn't add-up to me (after being a Mac Systems-Admin for 5 firms, and Software Dev for the better part of a decade), for the sake of argument, even if your upgrades are actually working I just don't see any way this could be stable in extended usage.

I've setup dozens of machines (16 at one place alone) to run for weeks at a time, at very high-level and not crash. My own MDD will be at 100% CPU for sometimes 12-hours straight and not crash (although with an SSD often needs block-recovery time).

I just don't see the value of a momentary speed bump that leads to cascades of kernel panics and instability. Crap, friends of mine who upgraded their Powerbooks to the 2GHz G4 CPU had copious increases in panics. With what you are doing, it has got to be running your machine ragged. If not, if I'm wrong, you should be posting Youtube Videos of what you are doing because it would catch-on like "Cold Fusion" if provably functional.

Moving on back to my 'slow and steady' workflow.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

yeh I'm not a video-making guy, but I doubt I'm the only one with this model, everything is so hidden in plain sight that I'm sure it is by design.

Here is what I think: engineers would not begin to sell AGP-8x to the public until they had working the equivalent of what they'd call "AGP-16x" in their laboratories. While the MDD2003 was made and sold slightly after the G5 debuted, I believe the story of MDD2003 begins before G5, it's mostly the MDD2002 with next-generation Northbridge chip. The MDD2003 with 1067MHz northbridge is really a Mass Production Run of the superprivate "AGP-16x test rig" all conveniently disguised as identical to MDD2002. The dedicated blitter bandwidth/467 MHz RAM being the key secret ingredient in the "AGP-16x" recipe. It was made to do this and run AGP-8x cards, these aren't overclocks, they're normal ways of programming the northbridge, they're normal clocks whereas Apple began UNDERCLOCKING how they sold stuff.

The 467 MHz RAM speed isn't crashing. New DDR1 RAM is labeled for compatibility among all 266, 333, and 400 MHz systems. They're omitting the next step up, they simply need to become Honest about which modules are limited to 400 vs can run all speeds 266, 333, 400, 467. In fact all new RAM can easily do the 467 speed, but they seem to want to flood the market (50/50) with RAMs actually limited/rigged to 400 like they say it is vs RAMs that can step up to 467 without you suspecting it.

I do demanding things like convert an entire iTunes AAC library into MP3 for use with OS9, this can take all day and sends the Altivec as well as everything else to max. I do get occasional crashes requiring reboot, not often. Likely it's the RAM being CL3 and not outright certified as CL2.5 which the MDD has been programmed to demand. CL2.5 is faster than CL3, so it is kind of nice to be required to have CL2.5 RAM - when it can't crash.

But it's equally likely my 1.583 GHz CPU core speed is causing the crashes because this is a serious overclock for real. My front panel switches that reduce speed are great for helping stability, I can slow the CPU to 1425 MHz or 1266 MHz if I get paranoid about what might crash. Here's what I think about the CPU: it is a 155nm chip no-defects with a process limit of 1533 MHz, it certainly cannot do 1600 MHz which would be a "polite" resonance frequency. Most people didn't get their CPU clockups past 1500 MHz, and mine was at first prone to crashing when set for 1583 MHz exceeding the "limit" of 1533, but I believe the 467 MHz RAM speed is helping to stabilize the 1583 MHz CPU speed. Think about it: RAM doing 333 or 400 MHz and CPU is limited to 1500/1533 MHz. Introduce an "odd" askew looking RAM speed of 467 MHz and presto the "odd" askew looking 1583 MHz CPU core begins working via a "magnetic burn-in", the 467 MHz RAM "teaches" the electrons in the 1583 MHz how to flow more correctly. I haven't run slower RAM long enough to see if my 1583 crashes more often but that would be an interesting test to perform.

All MDD2003 CPU set for 1.25 GHz are really pure 1.533 GHz chips which Apple underclocked. This is because of the 1.55Volt core, which cannot go any lower without crashing. The 1.55V core dictates the 155nm chip must have no defects. The 1500 MHz speed isn't an overclock, it is a normal clock, sorry to burst your bubble if anyone thought their 1.50 MDD CPU alone had put them in league with overclockers. What the intended design roadmap had been was a FSB speed increase to accompany the new northbridge chip, like seen in G5. Existing G4s could increase their FSB from 167 MHz to 200 MHz easily. But Apple engineers wanted this to be secret, wanted MDD2003 to have a secret highend mode in which the FSB jumps to 200 MHz (done thru programming! since the FSB can now change independent of everything except the synchronous CPU itself), the 200 MHz FSB would immediately cause the 1250/167 MHz CPU to run at 1500/200 MHz, and these CPU chips really ALL can do this 1500 MHz speed. The 200 MHz FSB would also eliminate the CPU bottleneck which AGP-8x experiences on the 167 MHz CPU-FSB. The 1.25 GHz MDD2003 CPU speed is an outright Underclock, for the trick purpose of allowing a 1500/200 MHz "mode" where the user knows no truth about what's really happening.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

Well as nice as all that sounds, so do derivative security and credit default swaps.

It's safe to say no-one will take your word for it. Kinda' like perpetual motion machines.
If you can prove it (works and is stable), you probably could slip right into an I.T. job with it.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

oops I reviewed my previous post and see I told about spending all day transcoding with altivec, then transitioned to how I get the occasional system freeze - but I didn't mean to relate those two sentences. I transcode all day some days without reducing the 1583 MHz CPU, and, those rare system freezes haven't formed a pattern yet. And this, Not to be confused with a nasty system freeze bug I found in Tiger 10.4.9 thru 10.4.11- switching my 10.4.11 from at-first "Client" to the "Server" OSX software package has fixed the nasty Tiger bug! Bless you macintoshgarden for offering Tiger Server!

-----

"Mirror Door" computers, once the top of Apple's line, were sold with the side-fan facing backwards and the Internet hasn't admitted it 16 years past as of now. I know the Internet is phony, and real life has become disease-infected from internet proliferation.

There is more than enough proof. What did they say the speed of AGP-8x is, 533 MHz, whether that's true or not, what would they say the speed of AGP-16x is ... 1067 MHz of course. My AGP-16x has a 1067 MHz Northbridge chip housing the AGP interface. Can a computer at least briefly simulate a 1067MHz synchronous AGP bus without a northbridge doing at least 1067, No. Should we conclude the debut AGP-16x controller was greater than 1067, No. The 1067 MHz Northbridge in the G4 is the prototype AGP-16x design. G5 (first edition) was handicapped with increased Southbridge activity that prevents asynchronous 467MHz RAM, G5 also handicapped with slower CL3 RAM, the 800 MHz G5 has 333 MHz RAM and can't do 400 MHz RAM, the 800-900 MHz G5s have slower northbridges, slower blitters. The entire G5 series was a great belly-flop, go away IBM.

I can keep contributing, maybe you noticed I always save some reserve ammunition (ideas, numbers, proofs). I'm not desperate to be believed.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

The easiest way it seems to determine if the Mirror Door model has the G4 northbridge or the G5 northbridge is by whether RAM slots are limited to 512MB capacity (G4 northbridge) vs 1GB per RAM module with the G5 northbridge.

I'm bringing this up again because I also want to point out this difference is explicitly mentioned in their separate MDD2002 and MDD2003 "Setup Guides" published by Apple. MDD2002's booklet/PDF details Memory on page 71, makes no mention of 1GB DIMMs. The corresponding MDD2003's page is number 40 but 1GB DIMMs still aren't mentioned, as if one hand didn't know what the other hand was doing at Apple when it seems they want to omit the memory change from the manual, but, then they do go and outline this change to the hardware when the manual mentions it a 2nd time in the summary section. MDD2002's is page 101 - says DIMMs can be 128MB, 256MB, and 512MB. MDD2003's is page 67 - says DIMMs can be 256MB, 512MB, and 1GB.

The difference in RAM capacity per slot is a generational change in the RAM controller (northbridge chip) and by being described in the MDD "Setup Guides" it means the MDD2003 was Mass Produced this way with the G5 chip. MDD2003 with this new "DDR Interface" can do what they say it does, 333 MHz RAM and 266 MHz AGP-4x, using just an 800 MHz northbridge to also cover the 200MHzX32bits Southbridge artery. Nothing short of the 1067MHz northbridge creates the 467 MHz RAM and therefore the "AGP-16x" mode.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

That's circumstantial. You can also use dual 1GB sticks in a MDD, but have to leave two empty. I know people who've done it with no problems. The limit to 2GB (as opposed to the theoretical of what is it 3.6GB) is because the G4 can only address 2GB - the limit is within the chip itself.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

1GB Memory Sticks only work in MDD2003, the MDD2002 lets 512MB half of one be operated. Capacity per slot is in the design of the northbridge, not CPU, but since the CPU has been moved into the northbridge on new intel models, who could argue that. Getting back to G4-G5, the model G4 "QuickSilver" only has 3 RAM slots of 512MB each, so, that model can't even have the 2GB - it's max RAM capacity is 1.5GB even tho the previous model "Digital Audio" holds 2GB.

G4 CPU can address 64GB of RAM even with 32bit software, see wikipedia "PowerPC G4". The 2GB limit we are familiar with is all software design OSX's "fault". My MDD2003 takes 4GB onboard, reports 4GB total in "AHT", and AHT probably tests the entire 4GB for accuracy, not sure if I confirmed it does. I don't "have to leave two [RAM slots] empty" - OSX will say my 4GB is only 2GB, and OS9 says its limit is 1.5GB .

The CPU is not the god or rockstar you think it is. The northbridge is the beating heart of the entire computer. The CPU is just a barnacle clinging to the FSB. The CPU doesn't create it's own bus speed, it rides the bus created by the northbridge chip. The CPU is unrelated to hardware i/o, peripheral busses, etc.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

Think what you like. Simplifying Assumptions without Empirical Data = Wishful Thinking.

Just look at economics. Those guys make a good sounding case, but couldn't find their butt with both hands in the real world, because of preconceived notions that amount to little more than confirmation-bias. True Scientific Method is deductive reasoning where you work to prove yourself wrong, and only when you can't, after extensive testing, can it then be considered a working theory.

I've met some very bright minds on these topics, and if anything they see the stability thresholds of G4 systems even more conservatively than I do. This was the primary reason Mac was the choice it was for so long is focus high-quality reliability, not 'Hale Mary' computing.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

What's this about now, we're back to "all or nothing" aka "winner takes all". You've made a deluge of false statements. Info on Internet is worse now than it was years ago. Gets worse by the minute it seems.

Tried looking up the 64GB RAM addressing of the G4 with little success, even tho I just covered this a little while ago and now can't find the same articles that described the "32 plus 4" pin assignment I'd just saw. I found 'pinouts' of late G4s showing the 36 total memory address pins/bits. 2 to the 36 is 64GB. It seems maybe early G4s didn't have this yet (32 bits for 4GB instead), but the MDD model is past that. I cite printed material: PC Magazine (parent-company name, the emblazoned name is ExtremeTech) "Spring 2005" Special Issue "CPU BIBLE" page 99 chart - PowerPC G4 : "Maximum memory addressability" its written "64GB" specifically naming the 7455 chip, the cpu chip that's in the Mirror Doors.

Maximum R.I.S.C.'s picture
Joined: 2017 Oct 18

This is the last comment I'm going to make on hardware with you - saw your post a while back about OS9 drivers for newer Radeon cards - and its the same thing.

Just because 64-bit is handled one way on a CISC machine like say a Core 2 Duo, doesn't mean it's the same on another machine (heck this Specter bug proved that - it varies wildly on how much access it can gain from say Intel vs AMD (eventhough both share the same root compatibility of AMD64). Implimentation is everything. In PowerPC 64-bit (so called BigPower) there are a few things that work better but just as many that don't - read the TenFourFox developer blog sometime - Cameron Kaiser is constantly running into optimizing-bugs with G5s.

In retro-gaming the Sega Saturn was actually 60% more powerful than the Sony Playstation (Saturn 50 Million instructions per second / Playsation 30 Million instructions per second), but the Playstation was drastically better optimized for 3D and thus won that war (eventhough Saturn was still quite superior for 2D).

Early 7450 G4 chips were faster than the previous 7400s but added a huge overhead in processing steps (much longer pipe) that ate up all that advantage. A 533MHz/7400 is faster than a 733MHz/7450 for many tasks, but the strides the 7450 made - paved the way for later G4s (especially the ones used in PowerBooks and eMacs that have often only a half-meg of L3 cache, but running at half the CPU-speed, so as long as you are not juggling too many apps it actually performs pretty well.

Did the G4 have some forward-thinking possibilities - YES, but they were not fully-realized and many served about as much use as a human appendix.

The only all or nothing in all these conversations is you expecting everyone to take your claims at face value with no proof and in the midst of constantly admitting bugs you are running into trying to get the improbable out of a very old machine.

Since speed seems to be the only thing you care about, just a buy newer machine. Used PCs are cheap as dirt this days and a good Linux distro like Mint Mate can outperform an Windows10/i7 running on an i3 (have seen this first hand).

Best Wishes.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

My post a while back about OS9/Radeon ? You mean this one ?:
http://macintoshgarden.org/apps/ati-radeon-9250-mac-os-9-drivers

I didn't start that, and what's wrong with my post? The 9600 card is a year 2003 model and OS9 was still being tinkered with by Apple in 2003. You didn't like my words "the 9600 [an AGP-8x card] is twice as fast as the 9000 [an AGP-4x card]"? I said "twice" which is prose/opinionated, not "2x" which implies scientific precision. My post posed a valid question, the card works in my machine for OSX (10.2 even) but has no 3D/DVD in OS9.

I'll stay with using my overclocked AGP-16x Mirror Door, everyone who sees it loves it.

-----

Edit: that's not the post of mine you're referring to, you mean the one about the speed ceilings for the smallest possible chipfab process-sizes? Which, is why you respond in saying CISC might behave different than "other machines" (RISC). If CISC is what's allowing higher clock rates, and I don't believe so, then CISC has faster cycles by getting less done per cycle.

How would I build the 2nd edition of PowerMac G5? I'd move from the 180nm northbridge, played out in G5's 1st edition, to the next-best chipfab, 155nm capable of 1533 MHz. Then I could do FSB speeds up to 1533, such as 1150 - 1250 - 1350 MHz. Why isn't this new northbridge internal bandwidth being assigned to anything. It's not enough to convert PCI to become asynchronous. The 130nm northbridge (PCIe) established 64bits per cycle (in CPU), X 2000MHz that's more bandwidth than I've dreamed of. If I sum all the PCIe slots (and eliminate the AGP slot), there's even more bandwidth left over for new things like a 2nd Ethernet port and a 4th USB2 port. But technology probably doesn't work in this manner, right? It's got to be magic.

davisdelo's picture
Offline
Joined: 2016 Oct 31

Nobody cares about any of the nonsense you keep babbling on about. Youre being obtuse and rather annoying. Take the hint and move on.

SkyCapt's picture
Offline
Joined: 2017 Jan 11

260nm - 533 MHz speed ceiling
220nm - 800 MHz
180nm - 1133MHz (1066 without liquid cooling)
155nm - 1533MHz
130nm - 2000MHz
110nm - 2400MHz
090nm - 2733MHz
065nm - 3200MHz
045nm - 3733MHz
032nm - 4066MHz
022nm - 4333MHz
011nm - 4800MHz
001nm - 5333MHz

Topic locked