Testing 333MHz IBM 604ev 288in PGA module with G3 carrier for a PCI Mac

[herd]

The G3 750G chips might work well in these computers, and they're available to buy. They have 1MB of L2 cache built into the chip and reportedly go up to 1GHz+. They might be the best you could do without level shifters, external cache circuitry and other complications.

 

[Powerbase]

The 604ev has no integrated cache-interface and needs a separate L2-cache controller. The additional circuitry on the Mach5-cards is exactly that: a cache-controller with clock-doubling feature.

The beige G3 Gossamer board has an unused onboard cache-slot (66MHz 64Bit) but that was never populated or used. I have compared the G3 cache-slot to IBM cache modules and there is a match with "75H5463" 1Mb L2 cache for RS6000 p-series. The slot itself is quite elusive since IBM used a bigger slot that would also take these smaller cache modules. I have not found anything compatible on my search.
View attachment 75132

The Gossamer board was originally meant for use with 603ev and 604ev CPUs but since the G3 was much more performant than anticipated, they switched gears and went with that instead. Despite having only had the integer units, the G3 is nearly as fast as the 604e with double the amount of integer units. The true power of the G3 shows with is its integrated cache-controller. The G3 CPU+Cache module is a cobbled together stopgap since Apple needed something to compete against the Pentium II. The G3 itself is not much more performant than the 604e, on the contrary, it is a bit slower. But adding a high-performance, high-speed cache into the mix and suddenly the G3 ist a performance monster. The low-power nature of the G3 (since it was originally meant as a 603ev replacement) allowed more speed without the need for a massive heatsink.

As the previous benchmarks might have indicated, the PPC platform has a big gain when the bus-speed gets higher. The PPC CPUs seem to need a lot of memory access cycles when calculations are performed. This in turn might hint at a deficiency of L1 cache stemming from the Power-platform ancestry these had much much bigger L1 caches. A fast L2 cache can compensate a bit for the lack of L1 cache and if we look at the later G4 generation and the final G3 generations, there was a clear trend to move the L2 cache ever more closer to the CPU and therefore making it mesh more seamlessly into the architecture as an extension to the L1 cache.

IBM was very aware of this flaw and added a cache-controller with onboard cache onto their CPU-cards, long before Apple designed the Mach5-card:
View attachment 75133
I am, on the other hand, quite amazed, that Apple struggled so much to implement a similar system in the Power Macs. The technology was clearly available but Apple chose to reinvent the wheel themselves.... Especially with their many ASICs and custom solutions, it seems that Apple was a bit too prideful to get 3rd party hardware into their machines as it was already common in the PC market. Only with the Gossamer board, they had swallowed their pride and began to use 3rd party parts like the Motorola 604-bus system controller that was capable of up to 100MHz bus speeds and had a much more advanced PCI interface integrated.

I think Steve Jobs actually calls out that weird need to make their own Apple-version of commonly available hardware/software in one of his keynotes. You'd think it would have been cheaper too, since Apple was pretty strapped for cash in the 90s.

 

[Arbee]

The MPC106 "Grackle" PCI controller used in Gossamer was an upgraded revision done at Apple's request as well (but available to everyone once Moto/Freescale started producing it). It was used because it was PReP compliant and supported Windows NT, which was Gossamer's original design goal. (That changed rapidly when Jobs came back though).

 

[stynx]

Here is a comparison between G3 and 604ev at nearly the same clock and with several configurations.
Based on these results, it can be said that the 750/G3 design was very capable since it could get 604-level performance out of a much simpler design. In combination with the fast cache, the G3 is on another level, nearly doubling the "processor" performance. I suspect that this is not as noticeable in calculation-heavy situations. The "processor"-test of MacBench 3.0 is most likely intended to simulate typical tasks on a desktop-style computer. The floating-point test is possible more geared towards typical number crunching where a fast and big L2 has no big advantage.

 

[stynx]

i have found some more infos about the ibm 604e

Scirocco: ..... IBM PPC604eBB133, 148 mm2, 0.35um process
Helmwind: ... IBM PPC604e2BB200, 96 mm2, 0.27 / 0.29um process
Mach5: ........ IBM PPC604e3BB300, 47 mm2, 0.25um process

Prozessorübersicht

The Helmwind was a bit of a disappointment in terms of speed and just replaced the Scirocco for legacy support since it was cheaper to produce and was pin and voltage compatible to the Scirocco. It made it the perfect replacement. It was seemingly originally meant to reach up to 300MHz but never went beyond 240MHz.
The Mach5 was produced in a more advanced process than the early G3s and the logic was that the Mach5 was SMP capable and panned for 2 and 4 cpu systems. A multicore Mach5 was seemingly planned but never prototyped. I suspect that the G3 was just too good a design for the 604 to keep up.

The Helmwind is a strange thing. It is a lot smaller and may produce a bit less heat but still needs 2.5V core voltage...

 

[Powerbase]

i have found some more infos about the ibm 604e

Scirocco: ..... IBM PPC604eBB133, 148 mm2, 0.35um process
Helmwind: ... IBM PPC604e2BB200, 96 mm2, 0.27 / 0.29um process
Mach5: ........ IBM PPC604e3BB300, 47 mm2, 0.25um process

Prozessorübersicht

The Helmwind was a bit of a disappointment in terms of speed and just replaced the Scirocco for legacy support since it was cheaper to produce and came. Hemlwind was pin and voltage compatible to the Scirocco and made it the perfect replacement. It was seemingly originally meant to reach up to 300MHz but never went beyond 240MHz.
The Mach5 was produced in a more advanced process than the early G3s and the logic was that the Mach5 was SMP capable and panned for 2 and 4 cpu systems. A multicore Mach5 was seemingly planned but never prototyped. I suspect that the G3 was just too good a design for the 604 to keep up.

The Helmwind is a strange thing. It is a lot smaller and may produce a bit less heat but still needs 2.5V core voltage...

I wonder if that processor card in the PowerExpress Rom jpg was a dual 604ev card.

 

[stynx]

I wonder if that processor card in the PowerExpress Rom jpg was a dual 604ev card.

I have not seen that picture, could you elaborate?

The Mach5 is about as small as a G3 and can be easily mixed up if you look at the size only. The Mach5 did not have a cache-management unit and was dependent on external logic (a cache controller) to have access to a L2 cache. If there is any cache-chip on a CPU card and no other big logic IC apart from the CPU, it is most likely a G3. I suspect that the G3 is actually dual cpu capable but it may need (extensive) external logic for memory coherency. This function is integrated in the 604e for SMP support and allows multiple cpus to use the same memory/cache/bus.

The 604ev was available in 1997. If the EVT/DVT was built in that timeframe, Apple could have built a dual 604ev-card for internal testing. I have seen a dual 604e card for the ANS but i could not look at the chips. That ANS-card had a separate 1M L2-cache for each CPU on the card and another 1M on the mainboard. There were two additional uncooled white IBM chips on the dual ANS-card. It looks a bit like the Apple Mach5 card but maybe they used the Helmwind (2nd gen 604e) since it is implicated that the chip can handle 100MHz bus-speeds (im struggling to find sources for that, though).

 

[Powerbase]

 

[stynx]

View attachment 75522

There is cache and a big Chip besides the CPUs. I would say dual Mach5 confirmed.

 

[LaPorta]

What does dual-anything get you in these machines if you are running OS 9?

 

[herd]

Photoshop Speed Test (G3 vs dual 604e)

 

[stynx]

What does dual-anything get you in these machines if you are running OS 9?

There are some tasks that did use the 2nd CPU. But the actual speed increase in these tasks is typically 50-60% over a single cpu. A 2x604e200 is about as fast as a G3-300 in these special situations. If you imagine a dual 350 Mach5, you cold expect equivalent performance of 550MHz G3 in 1997 un her MacOS (8.6). That would have been a welcome speed increase for a graphics-workstation, especially since the G3 was only available at 233 and 266MHz at the start. The later G3 300 tower was still not even near the performance of that dual Mach5 card. The real advantage would only emerge with OSX, though. The Unix kernel was SMP capable from the start and may have supplied a 1.8-1.9x factor (equivalent to a 650MHz G3 for a dual 350MHz Mach5) instead of 1.6x in Mac OS 8.6. The dual 604e card will most likely not be supported by Mac OS X but I will have to test that.

 

[Powerbase]

What does dual-anything get you in these machines if you are running OS 9?

Pretty much nothing, really, except some edge cases. But ... You did get to say I have TWO processors in my machine. That was something back then.

 

[obsolete]

The dual 604e card will most likely not be supported by Mac OS X but I will have to test that.

It's not. I had a 9500/180MP years ago, and with OS X installed with XPostFacto, only a single 180MHz 604e was recognized

 

[LaPorta]

Photoshop Speed Test (G3 vs dual 604e)

Yes, but my point is outside that one, single thing....nothing effectively.

 

[stynx]

Yes, but my point is outside that one, single thing....nothing effectively.

Cinema 4d and other raytracing Apps did also support MP. There was some support in calculation intensive apps. The way MP is realized on the Apple/daystar card was not optimal.

 

[stynx]

Interesting site...
The dual 604 might not be compatible with all the software on the list...
Photoshop and Cinema 4D are a give but even Quake 3, wow
I will try to test Q3 in OS9 in the future (i don't have much free time at the moment)

Programs that use both CPUs of a G4 Dual Processor PPC

Programs that use both CPUs of a G4 Dual Processor PPC

macos9lives.com

Here is a list of programs that can use (without the help of Mac OS 9) both CPUs of a G4 Dual Processor Mac:

3ivx
Absoft Pro Fortran
Adobe AfterFX
Adobe GoLive
Adobe Illustrator
Adobe Photoshop
Adobe Premiere
Alias Maya
Apple DVD Studio Pro
Apple Final Cut Pro
Apple Quicktime
Apple iDVD
Apple iMovie
Apple iTunes
Avid Media Composer
Deneba Canvas
Discreet Cleaner
Discreet Combustion
EIAS Universe 5.0
Emagic Logic Audio (only Plugins)
Eovia Carrara Studio
Filemaker Pro
Heuris MPEG Power Professional
id Software Quake 3
MOTU Digital Performer
Macromedia Dreamweaver
Macromedia Fireworks
Macromedia Flash
Macromedia Freehand
Maxon Cinema 4D
Metrowerks CodeWarrior
Microsoft Virtual PC
Newtek Lightwave
Omnigroup Giants: Citizen Kabuto
Pinnacle Cinewave (incl. DV100)
Propellerhead Reason
Roxio Toast
Sorenson Video Codec
SoundJam MP
Steinberg Cubase VST/32 5.1
Steinberg Nuendo
Toast Video CD
Toon Boom Studio
U&I Artmatic
U&I V-Track
Wolfram Mathematica

 

[Daniël]

Photoshop and Cinema 4D are a give but even Quake 3, wow

I do know that Q3A was one of the first x86 PC games to support SMP, neat to know it also does on PPC.

 

[stynx]

I just tested Q3A on a 8600 with 200MP cpu-card.
I could not see any improvement with "set r_smp 1". The CPU was correctly used by CineBench 2000. I suspect that the SMP-support in Q3A is only usable with "real" dual CPU-systems like the G4/450MP for early 2000. As far as i have seen, SMP was not supported from the start in Q3A but later assed. The release of the first Dual G4 would be most likely the reason that ID-Software added that feature. Since Q3A was originally developed on the Mac and then ported to the PC (in an early development stage), the SMP feature would have been possible to run on a 9600/200MP but i think that this system is just too slow for Q3A. A test with a Permedia 3, 1MB L2 and 256Mb Ram at 512x384p/16bit resulted in 7.3fps (no difference between SMP on/off) compared to 16.4fps with a 400MHz 604ev single cpu in the same system. (tested with "four.dm")

 

[Powerbase]

Honestly, I'm not surprised they didn't try to implement the weird multi-processing of the multi-604s since they were such an edge case.