G5 = IBM Powe4

[Marianco]

Server CPUs are expensive

Server CPUs like the Power4 and Pentium4 Xeon chips are power hungry and very expensive. The Power4 itself costs about $2500. Try buying a Dell Server. The single CPU server seems reasonable. But adding another Pentium Xeon CPU to create a dual CPU server will cost you another $2000. These CPUs clearly are not consumer models. The latest AMD Opteron server procesors are likewise going to be very expensive.

 

[alex_ant]

Re: Server CPUs are expensive

Originally posted by Marianco
Server CPUs like the Power4 and Pentium4 Xeon chips are power hungry and very expensive. The Power4 itself costs about $2500. Try buying a Dell Server. The single CPU server seems reasonable. But adding another Pentium Xeon CPU to create a dual CPU server will cost you another $2000. These CPUs clearly are not consumer models. The latest AMD Opteron server procesors are likewise going to be very expensive.

Server chips/systems have massive margins. For instance, in 1998, SGI sold MIPS R5000 CPU modules for its O2 at a price in excess of $2000, when the R5K itself cost $50 per unit in quantity. It's fun to look at the price lists of IBM/Sun/SGI/HP etc. "24X CD-ROM drive: $399. 1GB parity SDRAM: $999. 36.4GB Ultra160 hard drive w/ proprietary mounting sled: $1299."

The big Unix players need high margins to fund their R&D. Intel sets the Xeon prices high because they can get away with it - not because the Xeon costs anywhere near as much to produce as what it sells for. So, I'm sure that with a scaled-down and redesigned Power4-derived CPU put in place inside a consumer item like a Macintosh, the price would be able to come way, way down.

Alex

 

[T'hain Esh Kelch]

...

Originally posted by Sun Baked
The really interesting choice should be what type of bus Apple decides on for the next gen of CPU. I'd almost expect something safe like Hypertransport, but Apple could leapfrog to something much faster.

Back to the real world.. 166 mhz DDR...

 

[Chryx]

Re: ...

Originally posted by T'hain Esh Kelch

Back to the real world.. 166 mhz DDR...

FYI, AMD's Hammer processors use hypertransport for AGP/PCI <> processor links.. and guess what kind of memory they use?.. 166Mhz DDR Sdram....

 

[type_r503]

IBM

1. For the Idiot that wrote Altivec is the same as MMX. The are not even in the same boat. MMX which means Multimedia Instructions puts common MM tasks into silicon. Altivec is a 128-bit wide number cruncher. Intel has nothing like altivec and probably don't care to ever implement it. Their "speed" is from super high clock rates on the CPU.

2. Motorola is hurting bad. This rumor might indicate that IBM will be purchasing the PPC unit from MOT which should have happened long ago.

3. For altivec to even be used software companies must optimize their code to take advantage of the features. OSX is not optimized for the the altivec engine.

4. A 64bit bus is not just for memory addresses but also for data lines. This allows for 64bits of data to enter the CPU on a given clock cycle. The G4 only allows for 32bits. Once the data is in the CPU anything can happen. This greatly reduces the bottle neck caused by the memory bus.

type r503

 

[Chryx]

Re: IBM

Originally posted by type_r503
1. For the Idiot that wrote Altivec is the same as MMX. The are not even in the same boat. MMX which means Multimedia Instructions puts common MM tasks into silicon. Altivec is a 128-bit wide number cruncher. Intel has nothing like altivec and probably don't care to ever implement it. Their "speed" is from super high clock rates on the CPU.

Guess what SSE2 is...
a double precision int/fp number cruncher.....

3. For altivec to even be used software companies must optimize their code to take advantage of the features. OSX is not optimized for the the altivec engine.

I was under the impression that a lot of OSX code IS Altivec?

 

[gbojim]

Re: IBM

Originally posted by type_r503
1. For the Idiot that wrote Altivec is the same as MMX. The are not even in the same boat. MMX which means Multimedia Instructions puts common MM tasks into silicon. Altivec is a 128-bit wide number cruncher. Intel has nothing like altivec and probably don't care to ever implement it. Their "speed" is from super high clock rates on the CPU.

2. Motorola is hurting bad. This rumor might indicate that IBM will be purchasing the PPC unit from MOT which should have happened long ago.

3. For altivec to even be used software companies must optimize their code to take advantage of the features. OSX is not optimized for the the altivec engine.

4. A 64bit bus is not just for memory addresses but also for data lines. This allows for 64bits of data to enter the CPU on a given clock cycle. The G4 only allows for 32bits. Once the data is in the CPU anything can happen. This greatly reduces the bottle neck caused by the memory bus.

type r503

Just a couple of corrections...

The common belief is MMX stands for MultiMedia eXtensions. However, MMX officially does not stand for anything. Info courtesy of Intel docs. BTW, the newest incarnation of MMX known as SSE2 is closer in concept to Altivec but cannot hold a candle to it for performance.

OS X - especially with the upcoming release of 10.2 is heavily optimized for Altivec.

The G4 already reads 64 bits of external data per clock. Internally it processes 32 bit integer, 64 bit FP and 128 bit vector quantities per clock.

 

[Chryx]

g{The common belief is MMX stands for MultiMedia eXtensions. However, MMX officially does not stand for anything. Info courtesy of Intel docs. BTW, the newest incarnation of MMX known as SSE2 is closer in concept to Altivec but cannot hold a candle to it for performance.}g


SSE2 has nothing to do with MMX, totally seperate, MMX was integer only and couldn't be run in parallel with the floating point unit for starters... (modern x86 chips do still support it, although I believe the Athlon can still use the FP hardware whilst running MMX code?)

SSE2 OTOH, can do (IIRC) 2 double precision FP ops per cycle. (or 4 single precision)

then you have to factor in the bus/clockspeed advantage the P4 holds over the G4

Altivec is VERY nice, but Intel have something of a brute force advantage ATM.

 

[gbojim]

SSE2 has nothing to do with MMX, totally seperate...

That is not exactly true. SSE2 is quite different than MMX which I stated in my earlier post when I said the design is much closer in concept to Altivec. However, all Intel SIMD processors including SSE2 derived from MMX, are fully backward compatible to MMX and the design intent remains the same.

 

[Chryx]

Originally posted by gbojim


That is not exactly true. SSE2 is quite different than MMX which I stated in my earlier post when I said the design is much closer in concept to Altivec. However, all Intel SIMD processors including SSE2 derived from MMX, are fully backward compatible to MMX and the design intent remains the same.

I tend to look at SSE2 as a superset of MMX (and at the functions it shares it's much faster at than MMX was) I seem to have a vague recollection of the MMX and SSE2 execution hardware being seperate on the P4 though?.. might just be my frazzled brain though..


as for design intent?.. well, yeah.. making repetetive (instructionally) tasks faster was the design intent, so yeah

 

[Marianco]

Brute Force

The G4 Altivec parallel processing unit and the Pentium 4 Deep Pipelines are both brute force methods of processing data. The Altivec unit processes chunks of data in parallel, the Pentium 4 processes chunks of data serially but at high speed.

The Pentium 4 copied the G4 by adding an SIMD/Altivec-like unit. It one-upped the G4 by having a better bus to transfer data and instructions. It then upped the G4 by having deeper pipelines to process the data at higher speed.

The lower clockspeed G4 can keep up with the Pentium 4 because the Altivec unit is so much better than the Pentium 4's SIMD unit and because complex instructions to the CPU cause frequent dumps of the pipeline contents, slowing down the advantage of the long pipeline. However, the higher clockspeed of the Pentium 4, when coupled with less complex intructions (i.e. few branchpoints), allow the Pentium 4's Integer units and Floating point units to seriously beat the G4 in common tasks and games. Most common programs such as databases or word processors or games do not use the CPU's SIMD unit. Games use the graphic processors SIMD units instead.

Thus the G4 is at a disadvantage running at lower clockspeeds, for the many programs which need only integer processing. This alone makes me want to get a fast Pentium 4 computer to run Filemaker Pro 6.

The latest G4+ incarnations do add more pipeline depth, thus bringing more balance into the equation. Hopefully Apple unveils the 10 deep pipeline G4+, with a much faster bus (such as hypertransport, or whatever). This should allow the G4 to match or overrun the Pentium 4, maybe even the Athlon. However, the Athlon chip has deep pipelines, a great floating point unit, and and an SIMD unit which can match the G4. It's bottleneck is it's fairly slow bus.

 

[Chryx]

Re: Brute Force

Originally posted by Marianco
However, the Athlon chip has deep pipelines, a great floating point unit, and and an SIMD unit which can match the G4. It's bottleneck is it's fairly slow bus.

The AthlonXP already has a much better bus than the G4 though (although it can't use L3 cache)

133Mhz DDR Is significantly less of a bottleneck than 133Mhz SDR

(Clawhammer should be frickin' impressive though)

 

[PyroTurtle]

i still find it funny as to how much mac users seem to know (or, at least think they know as they themselves say) about the PC world...
i can sum all these treads up...

Better performance now, and give us those chips in high quantity, and very cheap! and preferably not from moto...maybe keep altivec if it really does work...

did i miss anything?

 

[Chryx]

Originally posted by PyroTurtle
i still find it funny as to how much mac users seem to know (or, at least think they know as they themselves say) about the PC world...
i can sum all these treads up...

Better performance now, and give us those chips in high quantity, and very cheap! and preferably not from moto...maybe keep altivec if it really does work...

did i miss anything?

I think you're supposed to say something snide about using PC133 on the top of the line machines in the year 2002?

 

[DavidRavenMoon]

Re: ...

Originally posted by T'hain Esh Kelch

Back to the real world.. 166 mhz DDR...

That would be Xserve Actually it has 266MHz PC2100 DDR SDRAM

I'd imagine we will see DDR in the new G4s.. when ever they come out.

 

[DavidRavenMoon]

Re: Re: IBM

Originally posted by Chryx
Guess what SSE2 is...
a double precision int/fp number cruncher.....

Except as in the case of the P-IV, it can't do FP and vector ops [SSE] at the same time, since the one unit handles both.

The G4 has a double precision FP unit, and four AltiVec vector processors. Altivec doesn't do double precision FP however. The "G5" is supposed to fix that.

BTW, I have read several times that IBM did indeed licence AltiVec from Moto, but just doesn't seem to be interested in using it. But I guess they could if Apple wanted them too.

Also someone was wondering... yes the POWER4 is a PowerPC chip.

 

[Chryx]

Re: Re: ...

Originally posted by DavidRavenMoon

That would be Xserve Actually it has 266MHz PC2100 DDR SDRAM

Actually, it has 133Mhz DDR PC2100 SDRAM, the 166Mhz DDR he was referring to would be PC2700

the "266Mhz" is very much a marchitecture term

 

[DavidRavenMoon]

Re: Re: Re: ...

Originally posted by Chryx


Actually, it has 133Mhz DDR PC2100 SDRAM, the 166Mhz DDR he was referring to would be PC2700

the "266Mhz" is very much a marchitecture term

I copy and pasted that off the Apple Xserve page:

Memory
*256MB or 512MB of 266MHz PC2100 DDR SDRAM with up to 2.1GB/s throughput
*Four DIMM slots supporting up to 2GB of DDR SDRAM using the following:
— 128MB or 256MB DIMMs (64-bit-wide, 128Mb technology)
— 512MB DIMMs (64-bit-wide, 256Mb technology)

http://www.apple.com/xserve/specs.html

If you are clocking the chip at 133Mhz, and it's reading on the rise and fall of the clock cycle, that would be 266DDR (Double Data Rate). 133 X 2.

PC2700 is ~166MHz (333DDR)

" PC2100 DDR is designed for use in systems and motherboards which require a 133MHz front-side bus. While the system or motherboard may operate a 133MHz front-side bus, its effective front-side bus speed is 266MHz in DDR operation. To get the effective front-side bus speed of a DDR system, double the operating frequency — which in this case is 133 MHz — to get the 266 MHz.

(PC2100 is also backward compatible for systems with a 100MHz front-side bus.)

PC2700 DDR is designed for use in systems and motherboards which require a 167MHz front-side bus, with an effective front-side bus speed of 333MHz. "

http://www.crucial.com/crucial/pvtcontent/topdef.asp?device=motherboard_extended

 

[Chryx]

Re: Re: Re: Re: ...

Originally posted by DavidRavenMoon

PC2700 DDR is designed for use in systems and motherboards which require a 167MHz front-side bus, with an effective front-side bus speed of 333MHz. "

That's my point exactly, "effective 333Mhz" isn't the same thing as "actual 333Mhz"

and you can't buy "Actual 266Mhz" DDR ram (well, not for system memory, it's available for videocards)


Like I said, it's marchitecture, it isn't actually running at 266mhz, but they claim it is because it "sorta is kinda".. bah

 

[TMay]

clairvoyance

One thing that I would bet on. Apple won't go with a different architecture than PPC.

two reasons:

1) PPC architecture is scalable from imbedded, low power consumer devices to PowerX server processors (or massively multiprocessor supercomputer architectures). There is no other architecture in the world that will do that. This portends OSX derived consumer devices, which will be a further boon to developers. Read this anyway that you want, but I wouldn't rule out a devices with PDA functionality.

http://www-3.ibm.com/chips/products/powerpc/newsletter/

2) It is cheaper for Apple to buy out Moto's desktop PPC line and develop the design in house and send it to a foundry (even AMD) than to switch. Either way, I believe that Moto's interest is considerably higher now that the economy sucks. Motorola also has some significant GaS manufacturing technology that will come on line in 2005 that may scale to 70Ghz. Moreover, the 8500 is substantially the same manufacturing technology as the 7500 (G5), so there really isn't a reason that we won't see a G5 from Moto. And of course, IBM would be interested, though they still aren't all that keen on Altivec.

My own opinion is that G5 development is advancing quite well, but the impetus for Apple to market such a device today (or even January) has been killed by the state of the economy. Moto will naturally be encouraged by Apple to keep these developments quiet.

Note that Apple has had to put in substantial resources to OSX so the hardware designs have had to take a back seat. With the immenent release of Jaguar, I believe that more resources will be available to hardware and consumer device development.

 

[Marianco]

Better CPUs from Apple

I believe that there is a huge amount of pressure on Apple to come up with faster computers. Thus there is a lot of pressure to come up with the G5.

Apple, remember, has been gobbling up high end video firms and recently eMagic. THere was also that closed door meeting with the Hollywood big-wigs, where Steve asked the leaders what will it take to make the switch to Mac. He was told, it is rumored, that power was king. Dual and Quad CPU PowerMacs were necessary. The server is one result of that meeting. XServe can be used in render farms.

 

[Marianco]

IBM's Role

Although IBM has been traditionally a server company and thus has not been interested in Altivec because it isn't needed when it comes to pushing data with server CPUs like the Power 4, it should definitely be eyeing the addition of an SIMD unit like Altivec on future CPUs.

What IBM missed is that SIMD units spectacularly speed up scientific programming tasks. Witness how the Dual 1 GHz G4 can do gene sequence searches 5 times faster using the program, Blast, then using Blast on a Pentium 4. Witness how much faster a 1 GHz G4 is in searching for encryption keys than the fastest Athlon or Pentium 4. Since IBM also makes massively parallel super computers for scientific research, it is shooting itself in the foot when it tries to use the Power 4 CPU in them compared to a super computer with multiple P4s or Athlons with their SIMD units. Again, the Power 4 is great at pushing data and doing server work (which is the bulk of what IBM makes profits on), but for scientific work (the other area where servers are also used), it acts like two G3s with a very fast bus. The AMD Hammers and Pentium 4 Xeons will nail the Power 4 on these tasks.

In any case, there are only two more days left before Apple unveils their next computers. woo hooo!

All this speculation about the future means nothing since you can't use it today.

I'll buy the latest greatest when Apple unveils them. When the G5 comes out, I'll buy that one too. I'll also buy a Pentium 4 computer since they are cheap enough from Wal-Mart - without Windows installed, to use as a second computer or peripheral to the PowerMac. Accounting software, unfortunately, is more available on the Windoze side. But then again, I do cross platform development. The AMD Athlon is great, but not entirely compatible with Windows XP like the Intel chips. Also, the P4's clockspeed has risen high enough to outstrip the Athlon in most things. Don't underestimate Intel. The Pentium 4 will keep getting faster even if it has to copy features of the G4 or other CPUs.

 

[Chryx]

Re: IBM's Role

Originally posted by Marianco
The AMD Athlon is great, but not entirely compatible with Windows XP like the Intel chips.

Do you have any evidence to substantiate this or are you just spreading FUD? (or blaming AMD for VIA's inability to make chipsets that work?)

 

[tjwett]

Originally posted by Megaquad
whats with all those programers who were working really hard to optimize apps for altivec? especially big players like adobe etc.

by the looks of things most developers have not been putting too much effort into AltiVec. in the begining they were but now everyone just seems to be Carbonising. perhaps it's because they know the new processors won't need it.

 

[Marianco]

Regarding Athlon compatability with Windows

Athlon-based motherboards are not totally compatible with certain hardware and software packages like the Intel Pentium-based motherboards.

For example, I do a lot of videoconferencing. The videoconferencing hardware/software for a 3-ISDN line 380 kbps setup I have specifically excludes Athlon compatibility. I thus have to use a Pentium-based motherboard, despite how great the Athlon is.

Yes, perhaps it is the fault of the motherboard makers, but for this niche, the hardware surrounding the Athlon is different enough to make a big difference.