IBM's 90nm PowerPC 970FX

[ffakr]

Originally posted by ktlx
Not true. The memory is clocked at basically the same rate at L3 cache but the latency is tremendous in comparison. L3 cache 800Mhz will always be faster than main memory at 800Mhz because of latency.

Wow, There is a lot of FUD in this thread. I'm not picking on ktlx but a lot of people here need to actually look a few things up and learn something before they start teaching others.
The P4EE has an ON-DIE L3 cache.. Just like the ON-DIE L2 cache that normal Pentium4s have (P4EE is actually a Xeon). On-die means... anyone? anyone? It runs at the same clock speed as the rest of the die.. 3.2GHz.

It's also not fair to make blanket statements about the processor buses.
The P4c and P4EE have an 800MHz 64bit wide processor to system bus. Bandwidth is 6.4 GB/sec
The PPC 970 has two uni-directional 32bit HT-like buses running at up to 1GHz. That's 4GB/sec up and 4GB/sec for a total bandwidth of 8GB/sec, but you don't get to use more than half of that in either direction. The bus on the 970 moves data in packets too.. so you have to subtract a routing/packet overhead from max theoretical bandwidth. The 1GHz G5 bus is closer to 7GB/sec.

Dual Channel DDR PC 400 (PC3200), for comparison, has 6.4 GB/sec of bandwidth (for both channels).
The P4c doesn't have a 'lame' bus.

Opteron beats them both though, as it's on die memory controller has half the latency of the P4 (and presumably half the 970 in a G5).

 

[amir12]

Hey guys, i attached the Heat Dissipation for the Pentium-M.

The best processor for Mobile PC's. They run very cool, quiet, and battery life is over 3 hours running full power all the time. IE: games. I get 4.5 hours doing normal stuff like internet browsing, music, and office if I limit the CPU to 600mhz (max) using the max battery setting in XP.

Note: the P-M is 130nm. Intel actually did a very good job. The 1.4 GHz P-M is "equivalent in speed to a Desktop P4 ~2.2 GHz." I've used the P4-M's and they are terrible for mobility.

Comparing these numbers it seems that Apple should be able to design very nice G5 PowerBook @ 1.4 GHz using the 970FX (perhaps even higher GHz). Although they must also account for any other heat dissipation from other components on the mother board that have not yet been specified by IBM or apple, but they have been discussed in this thread.

Although using the P-M wattage as a comparison, things seem optimistic for apple!

I also have a question for the more technical: Does anyone know how the current g4/g5's (and a prediction for the FX) scale processor wise. I mean in comparison to Intel's SpeedStep technology. Also how great are the power settings in OSX, or other third party battery management apps.

As an example I can limit my Mobile PC to 600mhz to get the max battery. When the computer is idle, it will drop down to under 100mhz. This obviously increases battery life tremendously when CPU usage is small or idle.

I may be completely wrong, but I remember reading somewhere that Linux/*inx operating systems and some random others can take advantage of a CPU instruction when the CPU is idle so that battery life is dramatically increased if the processor/board support it. Some software makers have attempted to create similar programs for windows (I think it was called "CPU Idle").

Thanks everyone and I hope the Intel Chart leads to a better discussion.

-Amir

EDIT: Also in case anyone is interested in the full pdf that includes all of the technical information on the P-M the link is:

ftp://download.intel.com/design/mobile/datashts/25261202.pdf

 

[tortoise]

Originally posted by nighthawk
Of course, this means that the 970 has a L3 cache up to 8GB running at the "same" speed as the P4 Extreme 2MB cache.

Unfortunately, not even close. Latency matters a LOT for memory performance, and L3 cache will run rings around main memory for latency. To add more injury, the latency to main memory for the PPC970 is actually pretty mediocre as such things go, fully TWICE the latency of the AMD64 chips to main memory, and possibly worse than the P4EE (I don't know the latency numbers for this off the top of my head).

So no, the higher bandwidth is really no substitute. If the latency was comparable your comparison would be valid, but in practice they aren't even close.

 

[Telomar]

Originally posted by tortoise
Unfortunately, not even close. Latency matters a LOT for memory performance, and L3 cache will run rings around main memory for latency. To add more injury, the latency to main memory for the PPC970 is actually pretty mediocre as such things go, fully TWICE the latency of the AMD64 chips to main memory, and possibly worse than the P4EE (I don't know the latency numbers for this off the top of my head).

Latency for ordinary PIVs is about 50% higher than Opterons in terms of processor cycles. When you actually work it out in terms of time they are actually pretty similar but it's worth noting the PIVs have an unusually low latency bus. The PPC 970s is dreadful but I'd also expect it to be one of the first things that gets fixed...

Originally posted by me_17213
the MIPS on te older 970 is alot higher than the FX version, I'm wondering why the difference would be so much.

970 1.8 7584 MIPS

970FX 2.0 5800 MIPS

Anybody care to fill me in?

Typo. Dhrystones are a near useless benchmark on modern microprocessors since the whole thing can be stored in cache. It ends up just scaling with frequency and in fact you can do a lot of things to tamper with the test if you were so inclined. Not really sure why anybody still quotes it, even the maker doesn't see the worth in the test anymore.

 

[tortoise]

Originally posted by stingerman
The P4's Integer units are double clocked but that is out of necessity.

Errmm... Double pumped ALUs are bleeding edge technology, and increase the performance of the CPU at relatively little cost. Everyone could benefit from this technology. And it makes sense if you consider it; the ALU is simpler than some of the other pipelines, and if you can manage to drive it twice as hard as the other pipelines you have a nice performance improvement on your hands on the integer side with only marginally more chip complexity.

You make it sound like it is a bad thing, but it is actually a cool thing. Mind you, I don't think Intel's cores are all that efficient in many ways, but they have found a way to exploit a design inefficiency that currently exists in both the PPC970 and AMD64 CPUs. I seem to vaguely remember reading somewhere that AMD was working on double-pumped ALUs, and they have no difficulty performing as well or better than the PPC970 clock for clock in their current incarnations.

Expect to see double-pumped ALUs in future AMD and IBM chips. It makes too much sense for them not to do it too.

 

[splashman]

Splashy thoughts . . .

Regarding G5s in Powerbooks, I'm under the impression (perhaps mistakenly) that the supporting chipset for the G5 will also consume more power, so Apple's engineers have more to think about than just processor wattage.

Regarding the suggestion for a re-designed G5 tower (smaller, or more drive bays) to reflect the lower-wattage 90nm procs, I'd say Apple is NOT going to dump a tower design that quickly. Whether you like the current design or not, Apple invested big bucks in it, and they are not going to invest in a step backwards by effectively LOWERING the tower's ability to dissipate heat. Instead, the current tower's capacity for heat dissipation will be useful for quite some time -- i.e., procs that are much faster than 3ghz. And with the first gen of 90nm procs (<3ghz), the fans will run that much less, reducing fan noise even more. I'm happy with my G4, but I'd be even happier if it were as quiet at the G5.

And regarding the need for more drive bays, etc. . . . Up until August, my trusty 9600 was my main work machine. I got it just as the first G3's were released, as I hated the fact that the G3 tower had so much less expandability. In the five years since I got the 9600, I made good use of all those PCI slots and RAM slots, and liked to gloat over the poor souls who were lured into the tiny new towers. But now that I'm getting along very nicely with my new G4, it seems to me that only a very small niche has need for boxes with multiple optical drives, more than 2 HDs, and more PCI and RAM slots. Sure, it would be nice, but I can't fault Apple's marketing or engineering strategy.

 

[JW Pepper]

The heat problem does't exist, but will it fit?

It is clear that the new G5 generates less heat than the present G4's used in the powerbooks, however what are the dimensions of the new G5's. Will they actually physically fit inside a 1" thick enclosure with a heat sink?

 

[tiselday]

G5 Cube, please

Forget the G5 Powerbook! for a moment at least

But, come on, Apple, give us a silent G5 Cube!!! 12W@1.4 Ghz would make it feasible.

Something as cute as this would take my $$$ (or rather, my kr) inmediately!!!

http://www.conf.co.jp/new_folder/making/cube_9.html



T

 

[jderman]

Guys,

It's not gonna happen tomorrow, or even day after tomorrow, but it will happen soon. They got over the big heat/power hurdel. Now all they have to do is redesign the motherboard and make it function in different power modes (Not that this is as simple as chopped liver). The big challenge is over. G5PBs are on their way. I'm banking on an introduction at WWDC. It makes sense. 5 months is enoguh R+D time for this (Thats not including the previous 6+ months that apple has probably already known about this and pre-empted it's development). Also consider apple's revenus. They have recreational R+D cash to develop this (And for god sakes it's the powerbook which equals big $$$ for them!).

It's coming

 

[Kermit]

Now, I do not want to ruin everybodys day here, but I still have to point it out. That 24+ W at 2.0 GHz is typical heat dissipation, meaning that the the number for maximum h.d. might be as much as 5-10 W more.

 

[Sun Baked]

Originally posted by Kermit
Now, I do not want to ruin everybodys day here, but I still have to point it out. That 24+ W at 2.0 GHz is typical heat dissipation, meaning that the the number for maximum h.d. might be as much as 5-10 W more.

Maximum -- when looking at the Apple XServe Max Power chart -- seems to be a LOT more than that.

The 2.0GHz 90nm G5 tipped the scales at 55W Max Power in the Apple XServe G5 docs.

 

[Head Wound]

Re: The heat problem does't exist, but will it fit?

Originally posted by JW Pepper
It is clear that the new G5 generates less heat than the present G4's used in the powerbooks, however what are the dimensions of the new G5's. Will they actually physically fit inside a 1" thick enclosure with a heat sink?

If they generate less heat, wouldn't they need a SMALLER heat sink? Also, wouldn't the new 90nm manufacturing process mean a smaller chip as well?

 

[l008com]

1.1 GHz Bus. If Apple comes out with 'new' G5 towers with a new Max speed of only 2.2 GHz.... There will be a **** storm of bitching on every computer forum on the planet!!!! But IBM has kept things secret for Apple in the past, I hope they are doing it again here.

 

[Kermit]

Originally posted by Sun Baked
Maximum -- when looking at the Apple XServe Max Power chart -- seems to be a LOT more than that.

The 2.0GHz 90nm G5 tipped the scales at 55W Max Power in the Apple XServe G5 docs.

I was basing my estimates on the information available regarding the G4. If what you say is true, then a G5 Powerbook might still be a long way off. Or do you disagree?

 

[jderman]

I don't know about you guys but i know that when I'm rendering a movie in FCP, watching a DVD, and working on a photoshop project all at once, the PB is usually plugged into the wall. On that note, who cares wether or not the max power consumption is at 30 or 40. The reality is that a very small portion of the PB market actually does that sort of thing without AC. if typical powerconsumption is in the 20-watt area (which mind you typical power consumption on a g5 probably equates to processing power much greater than the current g4 running at max power) then why not?

 

[Sun Baked]

Originally posted by Kermit
I was basing my estimates on the information available regarding the G4. If what you say is true, then a G5 Powerbook might still be a long way off. Or do you disagree?

Hard to tell, they may be similar in Power Usage (but they may be using different test software to get typical usage).

So a 1.4GHz PPC970FX isn't too big a stretch for an iMac.

But the UniNorth 3.x still creates quite a bit of heat -- look at the heatsink for the thing.

Remember the iMacs and PowerBooks of old, they used to use the PowerMac chipsets (like the eMac still does) -- so we "may" be waiting for the UniNorth 3.x to go 90nm, or Apple to make a integrated iMac/PowerBook chipset like the Pangea/Intrepid.

 

[Rincewind42]

Originally posted by Sun Baked
Maximum -- when looking at the Apple XServe Max Power chart -- seems to be a LOT more than that.

The 2.0GHz 90nm G5 tipped the scales at 55W Max Power in the Apple XServe G5 docs.

This may be counting the dual processor setup. I think Apple tends to count both processors together when quoting power usage (if you can even get a quote out of them). I recall back in the WWDC time frame being told that the PowerMac G5 put off a typical of around 100W - which seems ludicrous and can only be accounted for by totaling both CPUs.

 

[Rincewind42]

Re: Performance numbers

Originally posted by Booga
If you compare the performance numbers, the 970FX is a bit slower per MHz for SPEC numbers than the original 970. Although this is usually the case as the processors get faster clocks, it's kind of unfortunate to see.

It lists a 1.8GHz 970 as 828 SPECint2K and 1036 SPECfp2K. The 970FX at 2.0GHz (which is 1.11x faster in clock) gets 890 (1.07x) and 1100 (1.06x). It makes me wonder how it will scale into the 2.4+ GHz range.

The reason for this is actually incredibly simple - overhead. Regardless of the CPU frequency there is almost always a fixed amount of overhead that simply doesn't scale as well as the CPU's clock frequency, thus resulting in less than the expected gains (here 7/6% actual vs 11% expected).

So now that you know, don't be so surprised .

 

[123]

Originally posted by stingerman
The P4's Integer units are double clocked but that is out of necessity. They have do do a heck of a lot of work and need it just to get by. .....

WTF are you talking about?

Simple integer instructions have 0.5 cycles latency and 0.5 cycles throughput (P4), which is fast. Double-pumping is a good thing (tm) and has nothing to do with the RISC/CISC debate or whatever you are talking about, end of discussion. "out of necesssity"... tsss.

 

[Rincewind42]

Originally posted by Samurai980
Well it is nice to see IBM try to improve things, but why still such a low amount of L1, L2, and L3 cache?

Now the PowerPC 970 has a measily 32K L1 and a tiny 512K L2. Why doesn't IBM/Apple upgrade this tiny thing? I mean there would be a significant speed and performance boost as a result.

As others have mentioned, the 970s caches (L1&L2) aren't particularly small - they are in fact pretty average for the desktop market. The lack of L3 is presumably due to the bandwidth abilities of the FSB - the designers may not have seen it as necessary for their performance targets, and given how expensive L3 is that isn't surprising (just look at the cost of a P4 vs a P4EE).

Another interesting difference is the transistor count of the three processors. P4EE is 178m, AMD64 is 106m, and Apple is 58m.

The reasons for the difference is many, so I'll just breeze over the top ones.

1) P4EE has a HUGE amount of L3 cache on die, which accounts for the added transistor count over the P4, which I think was just south of 100m

2) AMD64 has a large transistor budget for x86-64 that the P4 doesn't.

3) The P4 & AMD64 have a large transistor budget for x86 to muOps conversion (that is, the conversion from external to internal instruction set).

4) AMD64 has, as you said, more L1 & L2 cache than the 970. The P4 has less L1 & L2, but makes up for it (in spades) with L3.

 

[Rincewind42]

Originally posted by 123
WTF are you talking about?

Simple integer instructions have 0.5 cycles latency and 0.5 cycles throughput (P4), which is fast. Double-pumping is a good thing (tm) and has nothing to do with the RISC/CISC debate or whatever you are talking about, end of discussion. "out of necesssity"... tsss.

There is no debating that the P4's integer throughput is good, but the necessity comes from the fact that at least half of the instructions going through a P4 have to go through the integer units, if only to do address calculations.

Just compare spec numbers - if the P4's integer units were able to go full out on only programmed operations then it'd smoke any other processor anywhere on integer operations (6.4 Ghz is quite a force to be reckoned with for integer) but the reality is that it doesn't - there is simply too much overhead in the CPU for that.

So while the P4 is an impressive integer machine, the x86 instruction set and the CPU design put together club it in the kneecaps.

 

[Sun Baked]

Originally posted by Rincewind42
This may be counting the dual processor setup. I think Apple tends to count both processors together when quoting power usage (if you can even get a quote out of them). I recall back in the WWDC time frame being told that the PowerMac G5 put off a typical of around 100W - which seems ludicrous and can only be accounted for by totaling both CPUs.

Not really the G4 at 1.33GHz is running max at 30-45W depending on the chip 74x5/74x7, version number, and the core voltage. Plus the rest of the numbers seem to be single CPUs.

So the single 130nm PPC970 may very well have been cranking out 97W @ 2.0GHz -- just seems quite high when Motorola was always talking about their 10W G4s.

I_am_Andrew posted this MAXIMUM POWER chart at ARS... and if you follow the discussion, Apple may have swapped a couple numbers.

 

[Rincewind42]

Originally posted by stingerman
As far as W usage, all I can say is WOW. IBM has pulled off a technological miracle. According to Motorola's charts the 970FX dissipates as mush W at 2GHz as the G4 7447 at 1.3GHZ! Now we know that the G5 architecture is a lot more efficient heat and energy wise than the G4 architecture. And, the 970FX can evidently dynamically scale not only frequency but also voltage. The chart implies that the FX will dynamically scale from 1.4-2GHz and 1-1.3V.

Watch how you generalize . That chart has a fatal flaw in it - it states the dissipation when running at a core voltage of 1.6v. However all of the PowerBooks run at either 1.3v or 1.1v (don't recall which) so those power quotes are (assuming 1.3v) over 50% to high! And if the PowerBooks run at 1.1v, then they are over twice as high!

Given all that, I would bet that only the 12" runs at 1.1v, and thus the 17" draws about 14w@1.3Ghz. So while that may be a feather in the cap of PowerBook G5 speculators, it remains to be seen how long it will be until we actually see one of these puppies and at what speed they will be rated at.

Also of note is that the G5 @ 1.4Ghz is running on 1.1v logic (not specified in the paper, but it is consistent with the power output) so that along may require some redesign of the PowerBook logic boards.

 

[mkaake]

me, i'm just happy

i figure, these are going to show up in some product, be it a cube, laptop, or desktop... which means refurbs, old machines and used machines suddenly get cheaper

call me cheap (or poor - that would be more accurate), but i love it when new products come out so i can get what everyone used to think was great at a good price...

matt

 

[dongmin]

Originally posted by l008com
1.1 GHz Bus. If Apple comes out with 'new' G5 towers with a new Max speed of only 2.2 GHz.... There will be a **** storm of bitching on every computer forum on the planet!!!! But IBM has kept things secret for Apple in the past, I hope they are doing it again here.

My guess is that the Xserves zere originally planned to top out at 2.2 ghz and 1.1 ghz bus but Apple scaled it back for some reason. IBM will not play spoiler for Apple, meaning expect higher (much higher, I'm guessing) numbers for the Power Macs when they're revised.

As others have noted, the big issue facing Apple now is how to make rest of the motherboard more efficient to fit into the PB form factor. The CPU is no longer an issue.