Giza aka Douglas Muth&#39;s LiveJournal

> Remember kids, if you are comparing a clock speed from a G4 chip versus a P4 chip (which is entirely different), then please stop making a fool of yourself, okay?

Using floating point operations per second is just as much bullshit as using clock speed, though. Which is what the article's writer seems to rely on heavily, when he's not half admitting the question has no valid answer and mumbling about compiler conspiracies. *g*
Basically, "this is where Macs are better, so I'll declare THIS the strongest virtue!". In the same way that humans consider language far more intelligence proving than maze navigation, because rats and other critters are so much better at it than us [and they are animals, so they must be stupider by definition right?].

And using brand name prebuild PCs for price comparisons? What the heck is he on? Do you declare pants to be more expensive than skirts because Levi is expensive? Someone's trying far too hard.

From:

furahi.livejournal.com

Not to mention the author is all biased...

I mean, a 2.4 Ghz Celeron is only good to run Windows 98?
Windows 98 won't even take advantage of all the capabilities of that processor or its motherboard :P

I mean, I don't necesarily disagree with the whole thing, but c'mon ^^()

From:

>Windows 98 won't even take advantage of all the capabilities of that processor or its motherboard :P

Thankyou for giving me another reason to not run Windows. :-)

Random thing: that's another thing I like about my Powerbook. The fans on it do not turn on until the processor reached 140 degrees fahrenheit. And that only happens when I play a game. So normally, I can just barely hear the hard drive and that's it!

From:

furahi.livejournal.com

Heh, I don't run Windows either.. But it's like saying the eMac (or iMac? :P) can only run something like MacOS 9

I'm not done reading those, actually I was reading when I got the email about this LJ comment ^^

From:

>Using floating point operations per second is just as much bullshit as using clock speed, though.

How do you figure? I'm not much of a hardware guy, but I would think that say, Doom 3 for the PC and Doom 3 for OS X are going to have floating point operations done at the application level, and the underlying OS isn't going to make much of a difference on how many FLOPS are required for a particular frame rate.

I don't disagree with you that the writer of that article is probably biased, though.

>And using brand name prebuild PCs for price comparisons?

As opposed to... what, exactly? I assure you that big companies (or even small ones, actually) don't build their own whitebox machines, they buy them all prebuilt from one or more vendors. The current state of technology for desktops is such that it is cheaper for companies to order prebuilt machines rather than try to spend the labor to build them on their own.

From:

Floating point operations don't represent more than a small fraction of operations used in real life. Even doubling their speed will likely not yield a 10% speed increase for your average user and even with math-heavy calculations it's doubtful they would every even be the majority, let alone a strong majority.
Most of the time, computers move data around, not calculate it. And even when they do, it's usually not in the floating point (you're a programmer, which variable do you work with more, int or float?). It's silly, but it's pretty much universally true. :)

As opposed to vendors building you a box given your specs. When our library bought new machines we discussed what we needed then sent those specs to all the local vendors. They put together a profile computer that matched our specs and gave us a price offer of their componenets. We picked the offer we found best price versus component wise and they constructed the 15 computers with said specs and delivered them. We paid component cost, not Premanufactured PC cost. Standard practice here.

From:

[floating point stuff]

Okay, I'll buy that.

>They put together a profile computer that matched our specs and gave us a price offer of their componenets.

Ah, but you're a library and presumably on a budget, so that makes sense.

Just about any business problem can be solved with one of two things: labor, or money. When you're a small organization (or a non-profit), money tends to be tight, so it's easier to solve the problem by throwing labor at it (i.e. doing more research and analysis to get custom machines so you save $50-100 per machine. The savings in turn probably add up to a significant part of your budget.). When you're in a large, publicly traded company, it's labor that becomes the scarcer resource (and the money saved becomes a fraction of your budget).

For example, if I worked in IT at my company, I could try to save a few thousand dollars on some machines, OR I could just spend the money on some pre-ordered systems, and use my labor instead on say, helping the sales guy access his e-mail so that he can finalize the details on a $5 million dollar contract, which will bring in much more revenue than me saving money by not ordering pre-built PCs.

There's also something to be said being able to easily order new machines that are identical to currently used ones. In $BIGCORP where I work, there is a small number of different (pre-built) machines that everyone has, which means that IT only has to maintain a very small number of disk images for setting up new machines. So spending money in some ways can cause potentially larger savings of money and labor in other ways. Funny how that works out.

From:

That's a very wonky example. You could also say that you could either invest 15 minutes coming up with a sensible system and save your company $300US per one of the 100 new machines or sit around browsing HomeStarRunner. Seriously, it took us a grand total of 15 minutes to come up with the specs we wanted, maybe an hour to decide which offer was best. WHAT cost? :p

From:

And custom-spec computers are a perfect way to maintain hardware unity. I mean, what the hell do you do when Dell AssMuncher4000 model breaks but they discontinued it 2 years ago? How will you replace it with an identical system? No problem with custom specs, you just order the same or the closes hardware to what was ordered the last time, the other side does all the work for you.

From:

>I mean, what the hell do you do when Dell AssMuncher4000 model breaks but they discontinued it 2 years ago?

That risk can be eliminated if you have a properly worded sales contract. You could negotiate something that says, "You agree to sell the Stinkpad 69a laptop for the next 3 years AND we agree to buy at least 100 of those machines per quarter". This is really good for the vendor, too,
since they now have a guaranteed source of income for the next 3 years.

From:

That doesn't really say anything for or against custom built machines though. Why not say "with this configuration" to that vendor and get the exact same thing?

From:

That risk can be eliminated if you have a properly worded sales contract. You could negotiate something that says, "You agree to sell the Stinkpad 69a laptop for the next 3 years AND we agree to buy at least 100 of those machines per quarter".

On what planet does any customer have enough clout to dictate product line lifetimes to vendors? If it was 100,000 machines per quarter, then maybe.

From:

On what planet does any customer have enough clout to dictate product line lifetimes to vendors?

Who said anything about *dictating* lifetimes? The conversation could concievably go like this:

"We want a contract where we can buy the same machine for the next 3 years. Think you could do that with Model A?"
"No, we plan on discontinuing that this year"
"How about model B?"
"Sure, we can sell that for 3 years. How many units do will you buy each quarter?"
"100"
"Okay, when we stop selling it to the general public in 2 years, we'll set aside 400 for your purchasing in the following year"
"Done!"

Make sense? From a business standpoint, you can get a contract for darn near anything. So long as each party gets something out of it, there's no reason why such an arrangement wouldn't work.

From:

Floating point operations don't represent more than a small fraction of operations used in real life.

Oh?

If you're running a word processor, sure. But a word processor doesn't care what CPU your system has.

If you're running a scientific workload, they're everything.

If you're running DOOM 3, they're a large and significant fraction of the workload.

If you're running Photoshop, you're actually probably memory-bound, not CPU-bound (though "memory bound" can include the cache hierarchy, which PowerN and 80x86 flavours all implement differently).

If you're running a scientific workload or a 3D game, compare SPEC_FP benchmarks. If you're running something that does mostly integer computations (which Photoshop might, depending on implementation and filters used), look at SPEC_Int. Look at the "rates" benchmark if it's a streaming workload instead of an easily "chunked" workload. If you're running certain classes of AI routine or nasty filter that spend a lot of time with conditional statements, you're screwed no matter what the processor architecture. If you're working on a streaming task that's easily chunked, you're memory-bound no matter what your processor, and the processor dick-size war is doubly pointless.

For the kind of workloads I have, FP matters most, when I'm stressing the processor at all.

For the kind of workloads Joe Gamer has, FP matters a lot.

For the kind of workloads Joe Office-Worker has, they could be running on a Pentium 1 and not notice much difference, as long as their system was free of bloatware (coughXP/Officecough).

From:

And which of these uses represents the largest portion of the PC market?

From:

And which of these uses represents the largest portion of the PC market?

Probably an even split between business, where purchasing decisions are dictated by politics as opposed to requirements ("nobody ever got fired for buying a Dell loaded with XP and Office"), gamers, who will buy a PC with the most expensive graphics card they can afford, and clueless home users, who believe that getting a P4 will make their internet go faster, and use their machines for word processing and email.

Of these, the only one with any significant workload at all is the gamer market, which is FP-bound when it's CPU bound at all. None of these are bound by memory bandwidth.

From:

So on average, superior FP operations (as opposed to simply good) don't represent a major part of the market demand? :)

From:

So on average, superior FP operations (as opposed to simply good) don't represent a major part of the market demand?

When benchmarks matter at all, the benchmark with the largest impact to the market is FP performance.

Saying that FP performance doesn't matter is therefore a bit odd.

I'd also hesitate to accept a claim that the gaming market isn't "a major part of the market demand".

From:

Circular reasoning. "FP matters most, because people think FP benchmarks matter most."

I never said FP performance doesn't matter, I said that most computer users (and applications at that) only use it to a minimal degree. Even with heavy calculation, moving data fast is still at least as important [you need to do something with those results, no?], and most users don't require that sort of algorithmic power. Plus, these days, games rely on the graphics card far more than the CPU; buy an Audigy and even that is reduced significantly.
Though using games as a pro-FP argument for Apples is rather silly. :3

And the gaming market is insignificant, at least the one you're pointing out. People who play Doom 3 and other games represent a minor part of the game market. The games that make the most profit are not first person shooters nor any other game that requires a monster PC to run, they're The Sims, Rollercoaster Tycoon and various puzzle games. So, FP operations are needed by large research facilities and a section of a market that is already small compared to the business bit.

From:

Circular reasoning. "FP matters most, because people think FP benchmarks matter most."

Now I'm almost certain you're just trolling.

Case 1: CPU performance matters at all.
If this is the case, then FP is the benchmark we care about, because it's the metric that has most impact on the cpu-heavy tasks performed by users.

Case 2: CPU performance doesn't matter.
If this is the only case you're considering important, why the hell are you debating about whether x86 or power4/5 is the better architecture?

If gaming is any significant part of the market at all, Case 1 is relevant.

Even with heavy calculation, moving data fast is still at least as important [you need to do something with those results, no?]

This is, frankly, not the case for the vast majority of workloads put on computers. Neither office-work machines nor unclued-user home machines stress the memory subsystem at all. Gaming stresses the FPU a lot more than it stresses the memory subsystem - memory subsystem isn't the bottleneck.

The only application for which the memory subsystem _is_ the bottleneck is media editing, and even that is debatable (2D image manipulation algorithms can use chunking algorithms to reduce the load to main memory, and video manipulation goes through enough material that disk access is likely the bottleneck). Either way, this represents a far smaller slice of the market than gaming machines, making memory subsystem performance a distant second in the "metrics that matter" list.

Does this clear things up for you?

Plus, these days, games rely on the graphics card far more than the CPU

Physics isn't done on the graphics card, and it requires detailed manipulation of the world model's geometry. It's the primary load on the CPU, for games.

I'm including here also visibility calculations for AI ("can this critter see me?"), and decisions on what to light, and where to put shadow masks, for lighting dynamic enough to require this (the T&L engines on the graphics card aren't magical).

Though using games as a pro-FP argument for Apples is rather silly. :3

Where did you get the impression I was claiming this? Certainly not from any of my statements about the market breakdown. The people who use Macs in industry are graphic designers, as the software is more mature on that platform, and the people who use Macs at home are the people who like them (either people who use them at work, or people whose first experiences were on a Mac). Gamers generally use PCs, for a wider game publisher base.

What I've been taking issue with are your statements about benchmarking.

From:

Case 2: CPU performance doesn't matter.
If this is the only case you're considering important, why the hell are you debating about whether x86 or power4/5 is the better architecture?

Yes, thanks for rewording my argument and ignoring the context. If you hadn't gotten so involved in the same little holy war you were chiding but one post below you'd have noticed that I wasn't comparing FP performance at all. I never said which is better, I never claimed it is useless. What I did claim however is that floating point operations do not present an important benchmark for a majority of computer users. It doesn't matter what research labs or specialized institutions need them for, they are NOT a major market share. THIS is why FP as a benchmark is utter nonsense.

The other bits are frankly a derailment on the point of the argument. No matter how much CPU power Havoc ragdoll calculation, quasi advanced AI algorithms or complex Photoshop filters take up, the majority of the market doesn't use them, making the benchmark equally irrelevant.

From:

I wasn't comparing FP performance at all. I never said which is better, I never claimed it is useless. What I did claim however is that floating point operations do not present an important benchmark for a majority of computer users.

Ok, this pretty much proves you're trolling. Goodbye.

From:

Thanks for playing! Don't let the cop out hit you on the way out.

From: