Analysis: x86 Vs PPC

Gentu writes "Nicholas Blachford (engineer of the PPC-based PEGASOS Platform) wrote a long and detailed article, comparing the PPC and the x86 architectures on a number of levels: performance, Vector processing and Power Consumption differences, architectural differences, RISC Vs CISC and more. The article is up-to-date and so it takes the G5 into account too."

Comments

UNIX was way before the X86:

[AtariAmarok]

You said "UNIX technology was created for the x86 architecture"

First x86: "The 8086 blasted away at amazing speeds of 4.77 and eventually 8 MHz -- hardly a calculator by today's standards. All this started in 1978."

(check here)

UNIX invented: "An
interactive time-sharing operating system invented in 1969
by Ken Thompson after Bell Labs left the Multics"

(click here)

Hackers

[Anonymous Coward]

RISC is going to change *everything*
<z3r0-c00l> Yeah, RISC is good

Now you can be as smart as they were almost a decade ago.

These arguments are so tired

[uradu]

This isn't the '80s anymore where performance is the most critical issue and we jump platforms every time a faster architecture comes out, since we don't have a large software base anyway. Nowaways software IS the more important aspect, and only relatively few well-heeled, game-addicted geeks are going to jump on the PPC just because it's a fews ticks faster this week, and Jobs winked at them with that very special smile. Given the way this industry goes, IBM/Motorola will sit back again, wipe the sweat off their foreheads and take a breather, and before you know it, Intel/AMD will have a faster processor again.

If you have x-platform software that will compile painlessly on either architecture, go for it, switch with each faster chip. But for most others, I doubt performance rants like these will make much of a difference. After all, how many Mac users switch to the PC just for the performance during those stretches when the PC has the upper hand?

Re:These arguments are so tired

[ctr2sprt]

only relatively few well-heeled, game-addicted geeks are going to jump on the PPC just because it's a fews ticks faster this week

Believe me, no hard-core gamers are going to "jump on the PPC" this week or any other week. They wouldn't even consider switching until they had good reason to believe that all future games were going to have full-featured Mac versions released at the same time as other versions. I think it's very unlikely this is going to happen.

No, what this really does is give occasional gamers something to brag about to their friends with PCs. While the aforementioned friend is playing, say, Half-Life 2 and the Mac guy is burning with envy, the Mac guy can at least say "Well, my computer's faster on these benchmarks!"

Re:These arguments are so tired

[AT]

software IS the more important aspect

Yes, and software is becoming more and more portable.

As the article observes, Linux (and open-source software in general) is not locked into the x86 architecture like Windows is. The use of server-side Java is growing, also architecture agnostic. Additionally, the web and web-based applications have shifted much of the work custom client applications used to do into the browser. Once again, architecture is doesn't matter.

The trend is that CPU architecture as a means of lock-in is declining, due those factors and many others. At some point, the cost of moving to another architecture will decline to near-zero, and the CPU business will shift to more of a commodity market, where people will buy on merit alone.

The only question is when it will happen; people have been predicting it for years (remember when NT ran on PPC, MIPS and Alpha?).

Right now, the PPC seems to win in some areas (power consumption, die size) and, barring architectural lock-in, it would probably take a large chunk of the Intel/AMD market.

Re:These arguments are so tired

[Hard_Code]

Did you even read the article? It is not about how PPC is faster than x86. It's about how PPC is more *efficient* than x86 which leads in the long term to lower power usage, whereas x86 gets diminishing returns on ramping up their clockspeed and playing games shuffling registers, etc. He specifically mentions that CPU speed is not really as critical as the companies make it out to seem because there are diminishing returns due to other system components. He mentions that x86 is up against a thermal wall by 2004 although I don't know where he got that data (it may be in a footnote but I not going to go back just to check). Speaking as a gamer who runs a pretty loud machine that overheats in summer, I am VERY interested in chips becoming cooler, moreso than them getting faster (the hard work is typically shoved off onto a graphics card).

Re:PPC comes out on top!

[Alrescha]

"UNIX technology was created for the x86 architecture"

In other news, it has been discovered that the current crop of teenagers has invented sex.

A.

Truly suprising colnclusion, OR NOT!

[pbox]

Nicholas Blachford (engineer of the PPC-based PEGASOS Platform) says that the PPC is better than x86.

What an unbiased opinion. Maybe we should really hear the other side too. I like the article for the wealth of info, and we all know the shortcomings of the x86 platform, but the conclusion seems to be biased.

Or is it just me?

Re:Truly suprising colnclusion, OR NOT!

[stevew]

Yep, and he is still stuck back in the 80's with his RISC vs CISC arguments. He says that internally they pretty much look the same (which they do) but they're some how different because RISC is easier to make happen.

Well - today's RISC's aren't very RISCy anymore. ;-) Todays CISC's have the same aspect. The machines have all migrated to simpler cores running VERY fast, but then tagging on features like predictive branching, out-of-order execution, etc.

An example of where the guy goes wrong is in his discussion of the compilers. What he fails to understand is that one BIG reason that the Intel compiler is better than GCC is that the same kinds of compiler optimization that accounts for how the hardware schedules things work for both the PPC and the Intel architecture. This has been true since the original entry of the MIPs architecture for goodness sake. Intel KNOWS what the hardware is going to do, and built those smarts into the compiler! You can do the same thing for the PowerPC by the way..not saying you can't.

Nuff said - it was an interesting article but bowed to much towards RISC is Great - All Hail RISC bunch.

Re:Truly suprising colnclusion, OR NOT!

[geirt]

... but he does misses one of the major problems with RISC architectures, the fact that RISC executables are larger that CISC programs (since RISC usually have simpler instructions and fixed instruction length). Today CPUs are fast, but memory are not. Because of this modern computers have large caches, 800MHz FSB, dual DDR memory busses, etc, but still the memory is slow compared to the raw computing power of the CPUs. But since a CISC program is smaller, the memory pressure is lower on a CISC system, and that's one of the reasons way the RISCs don't have the (on paper) large advantage compared to the CISCs.

This was not true 10 years ago, since the memory timing back then was in the 25MHz range, and the CPUs where running 20MHz. Today we have 3.2GHz CPUs and memory at 800 MHz, so program size matters.

Modern ARM RISC CPUs have worked around this problem by adding an extra instruction set called arm thumb, to make the program smaller. Smaller programs = faster execution on the same memory system

dang, somebody better tell intel

[cheezus]

um.... the PDP-11 was an x86?

"Desktop" weasel word. Where's the Opteron?

[Anonymous Coward]

The current desktop PowerPC and x86 CPUs are the following:
x86
AMD Athlon XP
Intel Pentium 4

PowerPC
IBM 750xx (G3)
Motorola 74xx (G4)
IBM 970 (G5)

I don't care if it's marketed for servers, just look at the cost: If you can afford a P4, you can probably afford an Opteron on your desk right now. If you can afford a G5 on your desk, you can definitely afford an Opteron on your desk.

Saying the Pentium 4 and Athlon XP are the current x86 chips, is just plain wrong. Those chips are obsolete except for very low-end (i.e. under $1k) systems. If you're building a x86 machine and your budget is approximately the same as the budget of a guy building a mid-range PPC system, then you have to be crazy to not get an Opteron, desktop or not. Thus, Opteron is the chip this author should be comparing to.

how long can x86 go?

[cheezus]

I remember years ago there being talk of the x86 never being able to keep up because it would just get hotter and bigger.... but now they're over 3ghz... was that all just hooey, or will there be a point where the x86 is dead and the RISC processors that replace them just have a CISC compatibility later?

Re:how long can x86 go?

[Blob Pet]

Which dies first, BSD or x86?

Re:how long can x86 go?

[Elwood P Dowd]

You could say that right now, "The x86 is not dead, because the RISC processors that replace them have a CISC compatibility layer".

The P4 decodes the larger, more complex x86 instructions into smaller chunks for use inside the processor, which is more or less RISC in its core. The CISC vs. RISC debate is kindof over, because both CISC and RISC chips have been adapted to gain the advantages of each others' design principles. Even the PPC 970 has to decode some of its "RISC" instructions into separate micro-instructions for execution.

The only chip design methodology that still has its original meaning is VLIW. That original meaning is "bankruptcy."

Law of diminishing returns

[Anonymous Coward]

The Law of diminishing returns is not exactly a new phenomenon, it was originally noticed in parallel computers by IBM engineer Gene Amdahl, one of creators of the IBM System 360 Architecture.

As opposed to economists, thousands of years ago.

A good OS...

[svenjob]

...makes all the difference. The thing that made me switch to PPC was, without an effing doubt, MacOS X. I went from an Athlon 2400+ with 768MB RAM to a home-made PowerMac 800 with 512MB RAM. I cut my processor by a 3rd and lowered my RAM. What did I gain? An amazing OS. If RISC processors continue to get more and more into the same processing spectrum as x86's, I think that OS X will help draw in the masses. Another thing that would help would be increased yields. That would lower prices and increase market share. Anyways, if x86 had OS X, I probably would have stayed with x86. But since it doesn't, I didn't.

An interesting viewpoint

[downix]

From my experience with RISC CPU's is that rating them by Mhz is often times the way to not understand what makes a RISC a RISC and a CISC a CISC.

Let me explain by example.

My MIPS R4400, running at around 120Mhz, I believe, runs circles around my Duron 750Mhz machine here. This is while the R4400 uses sDRAM vs DDR-RAM in the Duron, and the R4400 uses older plain-jane IDE while my Duron runs ATA-100.

I find it nice to boot up my old Indigo2 and play around, it responds so nicely, and renders quite well.

Re:An interesting viewpoint

[uradu]

> rating them by Mhz is often times the way to
> not understand what makes a RISC a RISC

What you mean is that you can't compare RISC MHz to CISC MHz--or any design's MHz to any other design's MHz, for that matter. Your statement in fact reveals that YOU don't understand RISC, because MHz are a much more reliable metric for RISC than for CISC CPUs. That is because by the very definition RISC CPUs tend to take a constant amount of ticks per instruction, which is not the case for CISC. So yes, comparing two RISC CPUs that both execute one instruction every two cycles on a MHz basis will give you a pretty good comparison of their relative performance.

Re:An interesting viewpoint

[Octorian]

Ok, that statement is just plain wrong, unless you're comparing something other than CPUs.

First, that Indigo2 is not "plain-jane IDE" (unless you're using some weird adapter board), but rather "plain-jane SCSI-2" (10MB/s).

Second, one big factor you notice when comparing CPUs, especially when some are "budget models" is that magic thing known as cache. Ever wonder what feature they're cutting to lower cost? I'll bet the R4400 has plenty of cache, while the Duron cuts cache (so does the Celeron, and some of Sun's older and slower microSPARC CPUs)

Third, even with those factors, there's no way in hell that the MIPS R4400 (at 120MHz) CPU could ever come close to touching the performance of an AMD Duron (750MHz). You have to be comparing graphics cards.

Now, one of the features of the Indigo2 that you might be using, is the "Impact" line of graphics cards. The Solid (no texturing) and High Impact (texturing) versions have about 450 MFLOPS of performance on the card itself, and the Max Impact has double that. I will believe that your Indigo2 whoops the crap out of the Duron on graphics, if you're comparing one of those fine GIO64 graphics cards to some POS card you threw in the PC.

But I will NOT believe you're comparing CPU performance.

How do I know this? Well, let's just say I've got an R10000 (195MHz) SGI Indigo2 High Impact sitting next to me.

[Q] Small & Expensive = CISCRISC?

[4of12]

When Microprocessors such as x86 were first developed during the 1970s memories were very low capacity and highly expensive. Consequently keeping the size of software down was important and the instruction sets in CPUs at the time reflected this.

So I'm puzzled. Perhaps someone can enlighten me on this.

If CISC is particularly appropriate for memory that is

1. low capacity, and
2. highly expensive

why doesn't the same argument apply to CPU's with no main memory per se, but just a good sized L3 cache?

Modern cache memories are, guess what,

1. low capacity, and
2. highly expensive

so it would seem to follow that higher performance could be got by using a CISC model.

Since main memory latency and BW are pretty limiting, I half expect that there's good argument to make very high performance systems live completely inside a large cache.

Re:[Q] Small & Expensive = CISCRISC?

[makapuf]

yup, microcode, that is several processor-specific RISC instructions put into a CISC instruction.

Then, RISC perormance and CISC compactness. If you want not to use CISC "huffman compression" , well use cheap instructions.

PPC

[pmz]

Given that electricity is not free, the fact that a PPC-based computer (or almost any non-x86 computer, for that matter) draws significantly less electricity is, well, significant.

If a company spends extra money on a set of gorgeous G5s or whatever, a non-trivial amount of that money is made back on the utility bills for very similar performance.

Other RISC vendors can be a win, also. For example, my old UltraSPARC workstations are not the space-heaters they might be stereotyped as (USII draws less than 20W). UltraSPARC III tops out at 65 watts, which although not as good as the PPC 970 is still much better than P4 or Itanic.

Re:PPC

[RzUpAnmsCwrds]

There are low power x86 processors. The Opteron, for example, draws around 55W for the high-end model. The Athlons aren't so bad either - around 65W depending on the model. P4 ranges from 60 to 100W, also depending on model.

Remember, electricity is pennies a KWh. Now, there are cooling considerations too, but even those are managable. In general, the highest operating expense of a company is not cooling or electricity but other factors like the facility, staff, or bandwidth.

Around here (Colorado), electricity is seven cents a kilowatt hour. Say a P4 @ 90W does the same work as a G5 @ 30W. That's a savings of 60W. Imagine the computer is on 10 hours a day, five days a week. That's a savings of 3000Whr (3KWh) or 21 cents a week. You save a total of $10.92 per year. Or, say the computers are on 24/7. That's a total of $36.69 per year.

Say an HP XW4100 system (P4 3.2CGhz) system does the same work in a CAD app as a dual 1.6GHz G5 system (remember, most CAD apps are not dual-processor optimized). The XW4100 is around $1500; the "low-end" G5 is $2000. At $36.69 per year (running 24/7), the G5 will pay for itself in 13.62 years.

Re:PPC

[pmz]

There are low power x86 processors.

Generally, they do not perform like the POWER4, UltraSPARC III, etc., for comparable power consumption. The Opteron is the closest bet for x86.

Remember, electricity is pennies a KWh.

Although $37 looks small, the savings scales with the company and can amount to thousands of dollars saved. Imagine an 8-way server ($300/year saved) or 32-way server ($1,200/year saved) or an office with 50 workstations ($2,000/year saved). That savings just might replace a broken photocopier or other budget-constrained items.

Power costs aren't something to laugh at, and conservation should be practiced in all aspects of a company (lighting, insulation, etc.). For self-employed people, it can mean an extra week's gasoline, for a large corporation, it can mean not laying someone off. These are real tangible benefits to buying low-consumption devices.

Re:PPC

[pmz]

At $36.69 per year (running 24/7), the G5 will pay for itself in 13.62 years.

I forgot to address this one. I think the payoff is faster than that, considering that there is added HVAC load from hotter computers, though I don't know how to estimate that.

Also, I don't mean to troll, but there is also the added savings of not dealing with Microsoft Windows every day (financial as well as psychological).

The break-even point is probably more like five or six years, which is a fair replacement interval for non-PC workstations. And after six years, the performance of a new workstation would be justified.

This means, at worst, a PowerMac G5 costs absolutely no more than a PC over time, and most likely (counting administraction costs) will be a net savings all around.

Re:PPC

[ichimunki]

I think you missed his/her point. If the cost of the more energy-efficient processor exceeds the amount of the money saved on the power bills, the company or household is worse off for buying the more efficient model. In the example, the $37 was no match for the $500 extra expense of the system.

Imagine buying a G5 iMac desktop will save me $50/year in electricity bills, but the system costs $200 more than a comparable x86 machine. Then it takes four years for the energy savings to pay for the added equipment expense. Multiplied over 50 workstations, the effect is the same, only the numbers get bigger on both sides of the equation. Just because those 50 machines will save me $2500 annually, doesn't mean they're necessarily worth $10,000 more up front.

However, the energy assumption is a difficult one to make. Energy costs are volatile, generally only increase, and are not an insignificant variable expense for most businesses-- minimizing that expense is not a bad move.

Re:PPC

[JaguarCro]

Around here (Colorado), electricity is seven cents a kilowatt hour.

Well welcome to Northern California where our electricity is currently 24 cents per kilowatt hour! Now here we are talking ($36.69 x (24/7) or $126.14 per year per machine. Apple doesn't sell a dual 1.6 Ghz machine, but if you still use your comparison numbers and prices we get a payoff in less than 4 years. (and if you really were just doing CAD you wouldn't need to Superdrive or Modem which cuts the price difference now down to $270 or almost 2 years!!)

In otherwards, before a/c costs are taken into account the machine will pay for itself in just over 2 years and for the remaining 1 to 2 years it is used it will be saving the company even more money.

By these numbers alone, I would say that buying a PC in expensive electricty areas of the country is a short sided mistake that will hurt your company or institution.

On the Tclk myth

[curious.corn]

Let's all remember that the MHz jump by intel was quite a marketing op. Consumers need an easy metric to evaluate goods (Hp in cars... btw, I wonder why people don't use Watts; must sound dull, dimensioning a car on a lightbulb unit) and intel chose to give one. They went as far as re-designing their machines around the pre-condition of high clock freqs. Take a P4 and clock it to 300 MHz (assuming it would run at those speeds and not bleed all charge out of it's gates), I don't think it would perform anything decent.

Re:On the Tclk myth

[Moridineas]

You make that sound so much worse than it actually is.

We know that as chips get more complicated they get harder to scale to faster speeds. The P4 was a chipdesign that, from the beginning, was designed to scale--huge pipelines, etc (and the pipes are getting bigger too).

Now what's wrong with making a chip that is easy to scale?

970 is a real superscaler

[norwoodites]

the 970 can have more than 200 instructions in flight at the same time, it can finish up to 5 instructions each clock (4 if there is no branches).

low power? not even close

[JohnZed]

Yup, g4s and g3s use substantially less power than their x86 foes, but the g5 is a different story altogether.

Each g5 dissipates a whopping 97 watts (see http://www.eet.com/sys/news/OEG20030623S0092, which is why the new powermacs have such absurd cooling systems and massive, mostly empty cases. The high-end powermacs actually come with an OUTRAGEOUS 600 watt power supply (http://developer.apple.com/documentation/Hardware /Developer_Notes/Macintosh_CPUs-G5/PowerMacG5/Powe rMacG5.pdf.
Let's be clear, this power supply is not for peripherals: the g5 powermac only supports 3 drive bays and 3 pci slots.

The numbers cited by the author come from an early projection of power consumption for lower-spec ppc970 processors.

Ummm, Sun?

[Wiz]

In the high end markets, RISC CPUs from HP, SGI, IBM and Sun still dominate. x86 has never been able to reach these performance levels even though they are sometimes a process generation or two ahead. RISC vendors will always be able to make a faster, smaller CPUs. Intel however can make many more CPUs for less.

Lets see. HP and SGI have sold themselves to the Itanium 2 which is ok but is EPIC, not RISC. IBM have the Power4+ and Power5 on the way which are pretty damned good. Sun have the US3 and US3i. Sun certainly don't have a performance lead. I've benchmarked an Opteron against the US3 and US3i and it isn't pretty for Sun. The Opteron is actually MORE efficient clock for clock in 95% of tests I ran. And yes, all the tests in question were real world programs running real world data. So I disagree. Just because it is RISC, doesn't mean it is faster. Chips like the Alpha prove it is possible, but with the rate of developlment of x86 and it's compilers it is becoming more difficult for them to keep up.

Check your facts, please; G5 IS low power

[LionMage]

Each g5 dissipates a whopping 97 watts

No, two G5 (PowerPC 970) processors together dissipate 97 Watts. Each individual processor dissipates about half that.

Don't believe me? Check out this chart on ArsTechnica. (The heading for the chart reads "Preliminaries: die size, power consumption, and clock speed.") A single 1.8 GHz PowerPC 970 dissipates 42 Watts. So a single 2.0 GHz PowerPC 970 dissipates a little more than that; therefore, it's reasonable that two of them would dissipate somewhere between 90 and 100 Watts, total.

The EE Times article you cited is highly inaccurate. They only look at the total number of fans in the G5 machine, and forget the fact that these are low-RPM fans and are software controlled per-zone to regulate temperature. Low RPM means less volume of air moved per unit time. So the design tradeoff that was made, clearly, is to have more fans running slower in order to keep noise levels down and to target cooling for each zone appropriately.

This is why it's a good idea to check multiple sources for your facts. Then again, if your goal was to present a very distorted version of reality to fit your goal of painting the G5 as a power hungry monster, you would very carefully choose your source of information so that it seems to support your assertion.