IBM PowerPC 970 Architecture

riclewis writes "Hannibal from Ars Technica offers an explanation of some of the internals of the new IBM chip. It's certainly more powerful than anything on the desktop now, but by the time it's released a year from now, it looks to be middle-of-the-pack (which could still be a step up for Apple...) This excitement over the early release of hardware specs kinda reminds me of all the hype surrounding the Sony's Emotion Engine when it was introduced a couple years ago. In fact, some are suggesting the PPC 970 chip might be closely related to the PS3's 'Cell' processor..."

Comments

Apple Chips

[cryptorella]

Middle of the Pack is not a step up for Apple... The G4 chips outperform Intel and there microinstruction intuperted to Risc instructions.... alot more goes into a processor than it's MHZ... Take a read of Hennessy and Patterson's book Computer Architecture A Quantitative Approach

Re:Apple Chips

[WittyName]

The PowerPC 970 triples the length of the PowerPC pipeline

This will give it the same issues the P4 has. Namely a large penalty for branch mispredicts, etc. Instructions per clock will decrease.

OTOH, they should be able to crank the speed!

Re:Apple Chips

[jcupitt65]

for general int code, the 800MHz G4 in my mac is about twice as fast as the 450MHz PII in my old work machine ... it only gets faster if you altivec stuff, which no one does (except some clever peeps in apple)

Re:Apple Chips

[Junks Jerzey]

Middle of the Pack is not a step up for Apple... The G4 chips outperform Intel and there microinstruction intuperted to Risc instructions.... alot more goes into a processor than it's MHZ... Take a read of Hennessy and Patterson's book Computer Architecture A Quantitative Approach

True, but there's still no denying that current Pentium 4's are faster. For the sake of argument, let's say that an 800MHz G4 is roughly equivalent to a 1.4GHz Pentium 4. Well, now a bottom-end $500 Dell is shipping with a 1.8GHz processor, the norm is 2-2.4GHz, and you can buy up to 2.8GHz, if you really want to throw your money way.

Bottom line: Yes, the G4 is faster than most people claim, but it is still measurably slower than what Intel is currently offering.

Re:Apple Chips

[Visigothe]

> Instructions per clock will decrease.

Actually, IPC is *increased* from the current G4. It will now fetch 8 instructions per clock, and retire 5 per clock.

The current G4 IIRC fetches either 3 or 4 per clock. I have no idea how many it can retire at once.

This coupled with a quick move to a .09 process shows me that this 970 chip has legs. Another thing... IBM has *always* been conservative about what not-quite-ready chips will do as far as clock, and benchmarks. I expect "Real World" [no relation to Peter Gabriel] performance to be quite good. [although I expect Peter Gabriel's performances to be fantastic =)]

Re:Apple Chips

[be-fan]

Actually, the DDR thing is a little misguided. The real reason DDR had no effect was because the 2.1 GB of memory bandwidth was feeding into 1.3 GB/sec of processor bus bandwidth.

Re:Apple Chips

[Mocenigo]

> Actually, IPC is *increased* from the current G4. It will now fetch 8 instructions per clock,
> and retire 5 per clock.

This for the branch/integer/fp core only. Which is borrowed from the Power4 one. This does not count
altivec, which is a separate unit on the same chip. Further, the two fp units of the core
can work in parallel with the altivec unit, which the P4 cannot do, because its vector unit uses the normal fpu pipelines...

>The current G4 IIRC fetches either 3 or 4 per clock. I have no idea how
> many it can retire at once.

fetche 3, retire 2 (IIRC, recent iterations may also retire 3)

Re:Apple Chips

[Shanep]

PPC chips can only work on one swing of the computing "cycle", not on the up and down like an Athlon can for example

It's called positive and negative edge triggering. It's not a new technology either. I was dealing with it in the 80's at the discrete logic level.

AGP 2x uses this and 4x uses positive, negative, high and low triggering. Certain UDMA modes make use of this clocking technique also.

Your argument doesn't hold water.

His arguement DOES hold water. PPC CPU's DO outperform Intel x86 CPU's by a good margin when compared clock for clock (showing the MHz Myth for what it is). Especially the G4 and boy when AltiVec can and is exploited... Wow. There IS more to CPU design than smaller die and deeper piplining for higher MHz.

As far as I can tell, Apple seem to be in a position where they have to make the best of what they can get, due to Motorolla dropping the ball pretty baddly.

I hope IBM comes to their rescue. How ironic.

Re:Apple Chips

[be-fan]

Depends, the G4e has a 7 stage pipeline, so tripling it would make it 21 stages.

Specmarks similar to Pentium 4 2.8 GHz

[Anonymous Coward]

ok, so it's SPEC INT and SPEC FP numbers are 937 and 1051 respectively. From www.spec.org, 2002 q3: dell Precision WorkStation 340 (2.8 GHz P4), specint base is 970, peak 1010; specfp base is 938, peak 947. When it's actually released, if they make 2003 Q2, it won't be particularly impressive. But the current apple G4 specmarks are about 35% of the 970, so it'll look good compared to that.

Apples and oranges... and dollars

[phorm]

That really depends if you're doing a cost-per-performance comparison though. Mac is still often expensive.
A lot of windows people I know build gaming machines though, so I suppose if there were a comparison there (if Mac could run all my games) then the cost of expensive video accelerators, etc could be factored in. While I suppose Mac would factor in such costs as well, most of the Mac people I know didn't buy their systems to run Doom3 and the newest UT.

In a surprise move ...

[mfago]

Apple starts shipping these in January. Hey, I can hope damn you! ;-)

At least IBM is pretty good at manufacturing microprocessors, while Moto is certainly not. IBM already has a 0.10 micron (not 0.09) fab in testing, so perhaps the 970 will get to >2GHz "soon."

In a related story: Moto is supposedly selling their chip business. I guess they finally realized they have no idea what they are doing.

Re:2004?

[Mononoke]

I wonder if my old 600 MHz G3 iMac can hold out that long?

Probably. My o'clocked 233 G3/DT (beige) is still going strong, and running Jaguar just fine, thank you.

64-bit processor benefits

[m11533]

Back in the "good-old-days", a primary benefit of the "newer", larger "bit" processors were the larger instructions. An 8-bit processor had small 8-bit instructions, with maybe some double-"word" instructions that were much slower to execute, along with an 8-bit integer math unit. Floating point, when you had it, was also constrained by the 8-bit size, though a bit less tightly. Thus, moving up in size, meant increases in performance on many fronts, but instruction width, integer math width, and addressing were the big ones.

I am wondering how this applies to these latest 64-bit processors. In the days of RISC, one would think that a reduced instruction set would easily fit in 32-bit instructions (those are rather huge and comfy compared to the old 8-bit days), though I would guess that a 64-bit instruction can include an opcode, register specification AND 32-bits of memory address, which would mean fewer multi-word instructions, which by old measures means faster execution. A 64-bit integer unit would have some real benefit. I find more and more cases where 32-bit integers are not sufficiently large to cover the range of values needed for problems, and that is without addressing over 32-bits of data.

I am curious if someone can compare these attributes of the current Pentium 4/Athlon XP processors with this PowerPC 970, the current SPARC from Sun (Ultra is it?), and the current HP/PA processor (though isn't that being dropped in favor of Itanium?)?

Re:64-bit processor benefits

[stux]

The 64 bit PPC uses 32 bit instructions.

Basically the only real difference is in the details of some instructions, and the 64bit registers.

Since you're using 64 bit integer registers, you can now use 64 bit addressing (pointers), which means you can calculate addresses for 64bit address spaces, which yes, means more RAM.

Macs are currently limited to below 4GB of ram, which is actually a limit... I think the most significant reason to move to 64bit PPC is to go beyond 4GB of physical ram.

The other benefit will be the ability to handle 64bit integers fast. As used by databases ;)

Another benefit will be 64bit load/stores which can happen in 1 cycle, rather than 2.

Of course, the Altivec unit has allowed 128bit load/stores for a while now (and the fpu allowed 64bit load/stores before)

Anywho, the big points of PPC64 are increased integer size and larger address space.

PPC does not use segment hacks like x86

if IBM makes it... i would think so

[johnpaul191]

supposedly the issue with Apple's chips over the last few years was Moto's manufacturing process. rumors say that IBM was always able to make more chips of higher speeds than Moto. the story is that because of the contract between the 3, IBM chips did not go in Apple boxes (upgrades and whatnot), and they could not outclock Moto.

yes, that's from the rumor mill, but everyone knows Moto has been going through a lot of corporate restructuring and who knows where they will be focusing in the next 5 years. IBM is going to make these chips (where ever they are going to be used) at a brand new plant in NY state. they have a great rep for quality control.

i kind of creepy thing is that the articles say they will probably debut 2nd half of next year (Macworld NYC? one last hurah! before MW moves back to Boston?) or not till January 2004. the articles also inply that they will debut at 1.4GHz. Apple is now selling 2 x 1.25 GHz G4 chips.

will Apple stall at or below 1.4 GHz till these new chips come out? the general upgrade of Apple machines is 5 or 6 months right now. that leave 2 possible revisions to the G4 towers before these babies are set. now i know that these chips will come with a super motherboard and 64 bit vs 32 and bla bla bla but Apple fights the megahertx myth even to somewhat educated comsumers. how will they be able to spin it when they have to explain it in terms of Apples vs Apples?

i guess it's a minor problem if these chips are as zippy as they say... a few benchmark tests and bar graphs should convey some message? maybe instead of having a 12 y.o. kid set up his iMac and go online in 5 minutes, they will have a 12 y.o. kid clone his dog or something. i would be impressed.

Re:Oh look

[frankie]

if you can't tell the difference between chip architectures by now, there's no point in trying to explain.

I have no idea who you are, Mononoke, but I'd wager $1000 that Hannibal Stokes knows more about chip architecture than you do. The PPC 970 will have a hard fight (both in marketing and benchmarks) against the 4+GHz x86 chips also due a year from now.

p.s. How the heck did that get rated as Insightful? I'm as rabid a Mac addict as any of you, but it's just plain wrong to mod someone up for spouting false evangelism.

Re:This is gobbledegook to me...

[Visigothe]

Well, I'll try.

rendering apps like Lightwave, Maya, etc will benefit from this for several reasons:

The 64bit architecture:
Lightwave [if rewritten to be 64bit] will be able to use bigger numbers, and use more memory. Bigger numbers means that calculations that would involve making a 64bit word out of 2 32bit words [as it currently stands] needn't be done. Being able to address more memory is *always* a good thing.

Really good Floating Point Performance:
3D rendering apps love FP. bigger/faster/more Fp units are a good thing.

Memory Bandwidth:
The 900MHz bus will allow a *huge* amount of memory to be shuttled back and forth from the processor *very* quickly. This means your huge scenes will be rendered faster.

Altivec/Vector Processing unit:
Because the VPU doesn't do double precision FP, it doesn't help in the final rendering [much]. It *will* help in things like realtime previews, where the math is simplified. Imagine *big* previews of scenes in realtime.

Multiprocessing:
This chip is [as implied] MERSI compliant. This means that it is a perfect candidate for multiprocessing, like the current G4.... but the 970 can go many more "ways" than the G4 [the G4 was in an "optimal" multiprocessing stage with 2 procs]. The 970 can go up to 16, IIRC.

This seems like it'll be a winner.

.

64-bit provides many advantages.

[Inoshiro]

Once you move beyond a 4.5billion, into the realm of 18.5 (two orders of magnitude past trillion), you can address anything for the forseable future (since you can count each year until the heatdeath of the universe this way, for example).

For vector operations, 64bit words make for some fast math operations, since you can pack more 32-bit integer components into each bus transfer.

For floating point, it means you have greater precision in hardware (allowing things like real physics and shapes to be modelled without noticable issues caused by subtle number creep). Since most systems use IEE-784 (64bit double precision floating point), it means a speedup to that software since you're not working with it as 2 32-bit operations.

In terms of storage space, it means you can address more than 2,199,023,255,552 bytes (~2 terabytes) of disk space (assuming a 512-byte sector). This is important for people with big RAID arrays today, and people with ludicrously big Maxtor drives 3-4 years from now.

For RAM, it means you don't have to worry about your server topping out at 4 gigabytes of RAM. It also means that your VM space has no effective limitation for the forseable future (very useful for people working on large projects, trying memory-intensive algorithmic approachs to traditionally NP-hard problems, or distributed computing problems).

I'm sure I missed a lot of the benefits even with this list. As you can see, 64-bit is not just a number game. It is 32 orders of magnitude larger than 2^32, meaning our grandchildren will probably still be using 64bit machines with no limitions being apparent (unlike 16-bit to 32-bit, which only moved from 65k to 4.5 billion in terms of addressable amounts of something).

Sounds kinda like the Athlon...

[fastpathguru]

Decodes/breaks down the native ISA, repackages them in bundles, then issues them to the execution units... A point-to-point FSB... Will have higher IPC than Athlon, but has all the same scalability limits. Hammer has the integrated memory controller and multiple hypertransport interfaces for fast IO and glueless MP. In short, PPC is similar to 7th generation x86 along with P4 and Athlon. Hammer is much more like Power4, but more highly integrated/cost-reduced. fpg

To all the x86 haters...

[Erich]

I see people here on Slashdot a lot who dislike the x86 processors because they do translation from the x86 ISA into internal opcodes.

Note that your new IBM chip is doing exactly that.

Intel and AMD have repeatedly shown that they can do whatever they like to implement top-notch internal architectures, and lopping on a translation unit only adds 10-20% die area and typically a very small performance hit over a traditional sequential RISC architecture. And they're free to change the internal architecture between revisions. And both Intel and AMD sell enough chips that they can spend a lot of money on designs and make them very good and still turn a profit.

Show me the money

[Andy Dodd]

http://www.heise.de/ct/english/02/05/182/

SPEC benchmarks for the G4 processors. (Not a synthetic benchmark issued by Apple, but by an unbiased third party, SPEC)

G4 1 GHz SPECs at 306 integer 187 floating-point
Interestingly, the 1 GHz G4 was almost neck-and-neck with a 1 GHz PIII (http://www.heise.de/ct/english/02/05/182/qpic02.j pg)

http://www.spec.org/osg/cpu2000/results/cpu2000. ht ml
A large archive of SPEC results for many CPUs, including x86.

A few choice results:
1.2 GHz Athlon (Ancient by today's standards) - 443 integer, 387 FP
Athlon XP 1700+ on an Epox EP-8KHA (Happens to be my mobo - Slowst Athlon XP listed for this mobo):
633 integer, 561 FP
Dell Precision Workstation 330, 1.3 GHz P4 - 474 integer, 502 FP (The P4 doesn't seem to be taking too much of a branch misprediction hit here)

So in the case of G4s, while they may be a bit more efficient MHz for MHz (And the P3 vs. G4 benchmarks so that this isn't even necessarily the case), the fact that they're so far behind on the clock speed curve hurts them badly.

If you want to see a good example of MHz not being everything, check out the benchmarks of Alpha systems - The 750 MHz ones chew even 1.2 GHz Athlons for lunch. But don't look at Apple...

Also interesting in the case of the SPEC benchmarks run by Heise - MS C pays a 10-15% performance hit over GCC in the SPEC benchmarks.

64-bit FP in hardware is already the norm

[coult]

For floating point, it means you have greater
precision in hardware (allowing things like real physics and shapes to be modelled without noticable issues caused by subtle number creep). Since most systems use IEE-784 (64bit double precision floating point), it means a speedup to that software since you're not working with it as 2 32-bit operations.

Actually, most CPUs today (including G4 and P4) do double-precision in hardware. The G4 does 64-bit FP multiply-add with a throughput of one operation per cycle (I'm pretty sure the P4 does too). Even the loads and stores are operating on 64-bit chunks. Going to a 64-bit processor won't change any of that. The only thing different for FP operations will be (1) you can hold a heck of a lot more numbers in memory! and (2) it might be possible for extended precision (128-bit) to be done easily in hardware.

Re:Selling at a loss?

[Frobozz0]

"Selling hardware at a loss is typical of the console market."

No, actually it's not. Only Microsoft loses money on it's boxes. No matter what Sony is selling their boxes for, they make a profit on every one. It's the difference between a profit oriented, well thought out plan, and a slapped together Microsoft 1.0.

Re:This is gobbledegook to me...

[Anonymous Coward]

With 64-bit ints, you'd have a maximum of 1.845 * 10^19 bytes of RAM. That's 16,384 exabytes.

Basically this is because you need to use pointers to address your memory. 32-bit ints can only go up to 4.29 billion, so they can only address 4.29 billion bytes - 4 GB. Even if you could install more memory, it would be useless, because you wouldn't be able to point to it.

Re:Power4 is not PowerPC except when it is :)

[lweinmunson]

Found the article www.iseriesnetwork.com. Verry good history of the IBM architecure and some info on the upcomming POWER5 and POWER6