Slashdot Mirror
The Dual 1GHz Pentium III Myth
Sander Sassen writes: "HardwareCentral has the latest on the dual 1GHz Pentium III controversy. Here's a blurb: 'The 1 GHz Intel Pentium III seems to be the subject of much controversy, as many claims have been made about its inability to run in a dual CPU configuration. HardwareCentral has been following the discussion closely and decided to put an end to all the rumors and get a couple of GigaHertz Pentium IIIs and a dual CPU motherboard and find out what exactly is the truth of the matter.'"
Or not _wrong_, exactly, he's just got slightly incorrect expectations. 3dfx is NOT helping him understand these issues, because naturally they want him to produce demos that are way over-the-top, like effects sequences in The Matrix or something.
Think of it as _time_ antialiasing. You don't antialias by having some big shocking Gaussian Blur that goes out for 10 pixels: it's very local, and the effect is to make angled edges appear more clearly angled (less stair-stepped, and not fuzzed). I do have a page that gets into this w.r.t fonts: it's this page.
By the same token, if you can see visible blurring happening in action, the motion blur is _already_ too much. What 3dfx need is not to arrange for Quake to send 27 frames to make a smooth blurry contrail behind everything (no matter how nice that may look in screenshots)- what's needed is to buffer a _single_ frame and average it with the current frame at a variable blend ratio. 50/50 might even be overly strong though it would maximize the time-based antialiasing- 40/60 could be better. Screenshots might show a tiny video echo effect on the fastest moving objects, but the most significant effect at typically high frame rates would be the softening of edges only in the specific directions of movement. To demo this you'd want something like a guy dressed in black and white pinstripes- the desired effect would be that the guy wouldn't blur out, but you would have a much clearer subliminal sense of his exact movements- something it would take a gamer or multimedia geek to pick up on.
Running at sufficiently high framerates (say, over 80), the ideal density _would_ be 50/50, because as the speed is pushed, differences between frames become smaller. The end result would be strictly confined to the softening of just such edges that are moving the most, which would highlight exactly what types of motion are happening.
Serendipitously, this type of motion blur (being not a fancy cinematic 'blur' that you don't see ALL THE TIME anyhow, but time antialiasing) would be just as suited to the driver as the planned spatial antialiasing- and just as suitable for application to ALL extant games that can be played on a 3dfx card. Again, all that's needed is to buffer one frame and average it with the current frame. It's not supposed to be a lightsaber blur, and 3dfx is foolish to emphasize this concept.
Time-based antialiasing is just as effective as spatial antialiasing but people don't know what to look for, partly because nobody seems to be bothering to even try doing it properly! It'll be just the one frame buffered- or possibly two. For the purposes of 3dfx, clearly using a single extra buffer and setting a blend ratio is the way to go. For serious video, a better approach is this: think of the frames as one pixel _deep_ and treat the antialiasing as a sphere centering on each pixel. The most weight would be on the pixel being tested, the pixels that are directly left or right or up or down or earlier or later would have less of a weighting, and so on: a pixel that's one pixel up _and_ left _and_ a frame earlier would have the least weighting. I've had very good results experimenting with code that averages the frames directly before and after a frame, then doubles the size of the immediate frame and dithercopies it down to antialias it and averages it with the time shifted frames. The goal was to cut video noise, and this approach was very effective at doing so without causing lots of stupid blur. (anyone doing an open source video editing program might try this- I need to finish up my version if possible and gpl it ;) )
The long and the short of it is that 3dfx should offer this spatial antialiasing as an extra driver feature, and that Carmack shouldn't have to do anything to enable it- and the last thing you want is to dump lots of frames in for a nice pretty blur effect. As Carmack quite rightly says you can do that easily in the program anyhow and don't need special interaction with the card- the only thing he's not getting (and this is because 3dfx are pointedly not suggesting he try it) is that time based antialiasing really isn't his problem, and is an entirely undramatic effect, to be sensed rather than seen and marvelled at. And marvelling at effects might sell more 3D cards ;)
Again, the desired effects are practical rather than strictly visual in a 3D game. Racing and flying games would definitely be more exciting with such a driver feature, but in a 3D game it would be a matter of quickly sensing when a player you're chasing is turning or slowing- or whether you are dodging a flying projectile thoroughly enough. These things would be conveyed through the subtle softening of the lines perpendicular to motion- giving you more information than just the straight frames. (I really need to render up some demos of this...)
draw frame at 2X
dithercopy or otherwise antialias to a screen-size buffer
display blend of buffer and the last frame's saved buffer
copy current frame's buffer to last frame's buffer, repeat
This also uses less display RAM. However, it is inferior to this:
draw frame at 2X
make blend of buffer and the last frame's saved buffer, also at 2X
dithercopy or otherwise antialias the blend to a screen-size buffer
copy screen buffer to screen, repeat
The reason the second is so much better than the first (despite using a lot more RAM) is that subtle movements are a lot more likely to show up at the scaled-up level (this does assume that there is a buffer that gets drawn to and antialiased down, which I think is a safe assumption). Storing only the screen buffer means that for many pixels there would be no change from frame to frame- storing the larger 'aliased' buffer and averaging at that size makes it far more likely that there will be significant differences. These will be averaged, and then antialiased by either dithercopying to a smaller size, or some 3dfx approach that is effectively the same thing in practice (there are only so many ways to do this). The result would be the effect I describe, of a softening of lines perpendicular to movement, but it would take on a subtlety and delicacy quite comparable to film.
Which is of course what 3dfx really want...
Hmm, that's funny. I thought the "king of floating point" was what made *Intel* absolutely superior to everyone else.
I concede that the cache speed is a problem for AMD, although in most benchmarks it seems to affect things perhaps less than optimizing for Intel chips does, and maybe makes the Athlon comarable in speed to the PIII, I have run into situations on my K6 where programs run horribly because of the cache. However, people need to start writing code with less cache misses where possible! Smaller is still better, a lot of the time. But AMD is working on that anyhow, just like Intel is working on actually releasing 1Ghz chips in any measurable quantity.
Also, Intel has major problems with (guess what?) overclockability, high power consumption, not producing reliable chips in quantity, not selling them for reasonable (market?) prices, and high operating temps! When you're pushing the chips this hard, they're *all* going to have these problems. I admit that the Athlon is a beast, but it's also faster, clock-for-clock, than the PIII core, which explains the extra transistors.
As to the future: Intel will have their new (slower for x86!) next generation architecture, while AMD will have... copper interconnects? Faster cache speeds? Even faster 64bit x86-compatible chips? Well, we'll see what the future brings, but I know who I'm rooting for.
And if you really want overclockability, low power consumption, good operating temps, etc., etc., don't look to fast chips from Intel *or* AMD, but rather wait for Transmeta or get a PPC chip or something.
---
pb Reply or e-mail; don't vaguely moderate.
pb Reply or e-mail; don't vaguely moderate.
The link is just as mythical as the Dual 1Ghz Pentium III.
I'm also tired of people such nonsense. The fact of the matter is that the CPU bus is not a bottleneck (it was proven in multitude of tests running 3d games on 66 and 100 MHz bus systems -- performance difference was less than 1%) and the memory bandwidth is not a bottleneck (it was proven in benchmarks comparing RDRAM to regualar PC100 SDRAM). However, for the 3d games, the CPU is the bottleneck at low resolutions. At higher resolutions (1024x768 and up), the video card becomes the bottleneck. (However, at any resolution the 3d card speed is much more important than the CPU speed). That is why a lowly Celeron equipped with GeForce DDR will kick a 1GHz machine with RDRAM and say a ATI card any time.
___
___
If you think big enough, you'll never have to do it.
JEEZ! Did you even READ the original post? Let me quote the relevant portion for you; try to engage your brain before responding.
I got news for you, in most non-server environments the proc is still the biggest bottle neck
Hmm..."non-server". Let's say it again together..."non-server." Now, let's read the relevant part of YOUR post:
The computer in question serves files...
Tell me again how your response was relevant to the original post.
Pardon me while I mock this thread.Since your feeble mind isn't going to catch the analogy, I think I'll just sit back and enjoy my self-imposed feelings of superiority.
(/FLAME)
Now, where did I put those hot dogs?
> even on my dual 500 with 768MB RAM the system still cant do everything i want it to. like seamless audio/video in the background, compiling on seperate windows, java stuff etc etc. more ram will only take you so far.
While multiple processes exacerbate your CPU load, they exacerbate your ability to feed the CPU with data more or less in proportion. Your problem may not be the amount of RAM, but you certainly have a problem with memory bandwith on any x86 ever made.
And the above analysis is before factoring in cache. Any one of the things you name is likely to put a heavy load on your cache, and when you try to do several at a time, you effectively run (several-1) of them without a cache. Or maybe all of them, if you get into a cache thrashing mode.
And it's those cache misses that make a CPU stall out for multiple cycles at a time.
--
Sheesh, evil *and* a jerk. -- Jade
I've been sysadmin in a grapics department that has several dual CPU machines. The dual processing capability doesn't give one the option to multitask when rendering. Of course, I perhaps might have been able to set it up so that it would, but for the most part, 3DSMax rendering ate both processors for lunch. Granted, they were nearly twice as fast as the nearest single processor machine in the office, but during rendering, both procs maxed at 100% utilization.
That said, the average joe MS user is far more interested in faster bandwidth. Most of that computer time is spent online (chat rooms, net games, pr0n, etc.) Joe doesn't care about 1Ghz, except perhaps as "my processor is faster than yours".
:)
So who has a budget for these things and actually wants them if not the technical incompetants? High end servers? Would rather have a fatter pipe. These processors are only useful for high end work stations. Maybe you care, but most of us don't use or need that.
Of course, I wish my processor was faster than yours...
-Ted
It isn't like any normal consumers can afford
> one of those chips anyways.
Well I sure can't. But 1 GHz chips I can't afford means remaindered or price-slashed 700 MHz chips that maybe I can afford. Not that you need a fast machine to run Linux - my Linux box is a P5-100 - but if you're running an MS OS, damn buddy, think how fast Turbo C v.2 is gonna run on one of those! Go Intel & AMD!
Yours WDK - WKiernan@concentric.net
"just a little bit of rewiring needed. ;-) "
The rewiring is to make the motherboard support the chip. The chip supports SMP fine, no rewiring needed.
_
Y?
I just can't wait to bolt some waveguide to the system clock and use it to cook my food - coming soon to a pc near you - the latest and greatest in pc cases, with a built in microwave - just open the convienient door, slip in that slice of cold pizza, and by the time your code is compiled, your food is hot...and just think, no more burnt popcorn - you can nuke it and have it right in front of you so you don't forget about it, just click a button on your desktop and it stops cooking (not recommended during a q3 fragfest)
of course, I guess that means we will need a bigger power supply....
The first thing we do, let's kill all the lawyers. Shakespeare, Henry VI, Part 2, Act 4, Scene 2
if you run the well-known BX chipset, and both cpu's are run to near saturation for longish periods of time (more than a few weeks, usually), the system WILL lock up.
I found this while stress testing my dual cpu BX boards (asus, tyan and abit). I ran 2 instances of seti@home on each cpu (yes, 4 processes on a dual system). ALL my BX systems hung eventually. even the ones with extra fans on the bx chip and heat compound under the sink.
my guess is that when the BX was released, intel felt it was "good enough" and that no one would totally saturate their CPUs. certainly NT wouldn't - and I bet linux 2.0 wouldn't either. not until linux 2.2 would both cpu's be used to this level.
so I can't believe intel anymore when they lie about their own products. dual 1ghz - gimme a break; they can't even make ONE ghz work under normal consumer conditions!
--
--
"It is now safe to switch off your computer."
There isn's nearly one program that most people use that would actually make getting a dual PIII Ghz machine worth it. I can't speak for all devleopers, but for myself, cutting down times in compiling code, and being able to multi-task and do somthing else while the code is compiling, without speed-downs, is absolutly crucial for productivity.
I Have a Dual-500mhz machine, and i max out both cpu's all the time. Having Twice the power would almost double my productivity. As for most users - Anything that is multi-threaded and cpu-intensive will take full advantage of these cpu's. I believe that apps like 3d max are multi-threaded and can take advantage of the SMP. My artist friend has to take whole afternoons off while his animations are rendering because he can't use his system for other things while it is rendering. he has a single cpu. If he had a dual-cpu system, im sure he could be doing other things(photoshop,email,etc) while his images rendered. Just my 2 Cents (3 in canada)
Geek: I can download 5 billion pictures in 7 seconds.
Homer: But I want them now.
Discussion Never Hurt Anyone.
Libertarians
Actually, the poster child for (good) motion blur isn't FPSes at all. It's racing games and flightsims. Having the nearfield flickering semirandomly loses the sense of great speed. Even a small amount of general purpose motion blur would help this quite a bit. I concede that's Not Your Problem :)
[WARNING: The contents of this post are slightly of topic, until you consider A) This is a subject regarding excessive CPU power that nonetheless consistently gets overrun and B) This site has been overrun and I can't comment on the actual contents of it.]
:-)
:-)
:-)
[WARNING 2: This is one of the more bitter posts I've made to Slashdot. You've been warned.]
And so it was so ordered, after legions upon legions of sites fell to That Which Was The Slashdot Horde:
If the content of the web page is not dependent on the identity of the user, then the content of the web page MUST not be generated specifically for that user.
Yes, that's an IETF must, damnit
This isn't a complicated concept, folks. If each user gets a very different page(think search engine), then you dynamically generate the new content live. If each user only gets a slightly different page...well, gee, dynamically generate that slight difference, but leave static everything else.
If you're dependent on the user, change the page for each user. If you're dependent on some local index of news, then change the page each time the local index of news changes. If you're dependent on an angel coming down and teaching you to code the goddamn meaning of life in Perl, *THEN CHANGE THE PAGE WHEN SOME GLOWING HALOED CREATURE WALTZES IN YOUR STUDIO*, but for *CRYING OUT LOUD* don't regenerate your page every time I try to read some godforsaken article!
It's simple stuff like this that make me feel like I have a moral obligation to be a Comp Sci major. Grrrr.
One other point...ya wanna talk overcommitment? The Linux kernel lists are going nuts about the reasonably rare situations that can arise when the OS allows processes to overcommit memory, on the probabalistic assumption that not all processes will actually use the memory they request. What to do when the memory actually commited actually becomes used? Should the OS die, so that the processes may live? What processes does the OS kill to keep itself alive? There's alot of argument about how to deal with overcommitment on the OS level, and I'll leave that fight to the experts.
But lemme tell ya, just view the Slashdot Victim of the Day to find web pages that deal with overcommitment. Since these sites aren't too likely to change their entire codebases all that soon, may I suggest that expressing Database Errors *might* not be the most graceful method of expressing degradation of resources?
In other words, faced with the choice of fewer ad impressions and less readers vs. temporarily switching to a cached copy of the page which is 99.9% accurate, might it not be nice to have built as a core element of Apache's modperl something along the lines of, "Run this script to generate this page UNLESS we're getting hammered; in that case, use mod_rewrite to change the URL to a static equivalent of our now thoroughly overloaded page"?
Ahhhh. I might actually be able to view pages about Gigahertz SMP
Yours Truly,
Dan Kaminsky
DoxPara Research
http://www.doxpara.com
P.S. Irony #1235235: It's taking me forever to finally get this comment posted
secondly, it has always been true that you need to properly spec your system based on what you plan to use it for. If it's going to be serving static webpages, you want lots of memory and a PCI architecture that can support multiple concurrent bus masters (to handle the high bandwidth). CPU speed is less important than cache size as well in this configuration.
If you're wanting to run Quake, you want a smaller amount of RAM but it needs to be low latency (say, 6ns) and have a high bandwidth (PC133). You'll want a fast processor and a faster video card. HDD is unimportant for quake, as is the bussing architecture (PCI, ISA - not the FSB).
But don't tell me there's a single metric for measuring system performance.. that's a lie, as much as saying that average access time for harddrives is "the" metric. I'd disagree, it's the track-to-track latency and aureal density that I happen to have used to spec my system. Not that it's "the" metric, but it's the two I used (as well as internal xfer rate, of course) which is directly proportional to the RPM the HDD is rated for!
To recap, the article has 6 pages on modifying the Iwill Slocket IIs (albiet with graphics).
I like the item mentioned in the conclusion better: "a new revision of the Slocket II is currently in the works that will support FCPGA SMP out of the box, making the configuration of a dual CPU system a matter of plugging the CPUs in the slockets, no soldering required."
> Granted that it is a royal pain in the butt
Yet still less trouble than actually rounding up two 1GHz processors.
--
Sheesh, evil *and* a jerk. -- Jade
DB Error: Unable to get author information!
:(
Cant read the article
-- iCEBaLM
Nearly 30 replies and nobody has made the obligatory Beowulf Cluster comment?
Information wants to be free -- but informants want to be paid.
They may or may not have been right, but in any case they certainly did not run a pair of factory 1GHz CPUs in an SMP configuration. I'll grant that the core is almost certainly the same; but even if they are correct in their contention that the stepping is the same for the two parts, it is a trivial matter for Intel to change the packaging to render the 1GHz version incapable of supporting SMP -- don't bond the SMP pin to the die, and they've done it!
The big question from my point of view is this: why would Intel say their flagship processor won't support SMP? This isn't like the case with the Celeron, where they clearly wanted people to buy the more-expensive processor instead of the cheapie... so why don't they want people to buy their most expensive product in pairs? The Celerons were too cheap, and the early FC-PGA Coppermines (the 500E and 550E) were just too overclockable; it makes sense that Intel would want to disable SMP for them, and so they played hide-the-SMP-pin. It also appears they've gone further with the CeleronII (AKA Coppermine128), and simply not bonded the SMP pin to the die -- again, a pricing issue... But why the 1GHz part? It's hard to buy Sassen's argument that it's just heat -- that applies to single-CPU systems as well as to duals.
So I don't think the controversy is over; it has just gotten more complicated, is all.
---
---
Politics is about making compromises. Religion isn't. --Michael Horton
I'm not sure about the 1GHz machines, but AMD was having real problems with thier caches, at least when the 800 MHz Athlons came out. They ended up setting the cache multipliers at 3/8. If you're doing hard core rendering and simulation, you'll want an Athlon, but you'll also want to be running an SMP machine. Unfortunately, there aren't a lot of SMP Athlon machines out there.
If, on the other hand, you're looking at gaming. I beleive the video card is the current bottleneck in a high performace gaming system. The Athlon/PII decision becomes a mattter of personal preference.
Database manipulation and general OS tasks would seem to be where the PII would shine, given that it's cache multiplier is 1. That whole cache multiplier problem is going to be a real big problem for AMD if they don't get it figured out soon.
Karl
I'm a slacker? You're the one who waited until now to just sit arround.
Copyright Violation:"theft, piracy"::Anti-Trust Violation:"thermonuclear price terrorism"<-Overly dramatic language.
Two 1-GHz processors on one motherboard seems like an incredible source of bus contention. How fast is the bus these CPUs are connected to? The CPUs are at 10 times the bus speed! (afaik)
Intel on the other hand has a shared front side bus, so all the CPUs can see each other's traffic and results in a substantially less complex chipset, with a lot fewer pins. Probably lower performance (depending on several factors), but certainly a cheeper solution for dual and 4-way SMP machines
- Mike
Molog
So Linus, what are we doing tonight?
So Linus, what are we going to do tonight?
The same thing we do every night Tux. Try to take over the world!
Do dual Athlons motherboard exits?, and at what cost?
I do not believe there is one out there yet, but here, near the bottom it says that tyan is making a board to be released Q4-2000 , Codename Dolphin, that is supposidly Dual-CPU capable.
I'm tired of all you people saying that faster procs are useless because of other bottlenecks. I got news for you, in most non-server environments the proc is still the biggest bottle neck. To tell the truth, I enjoyed much more the 50% boost from 200 to 300 MHz than I did the 50% boost from a 66 to a 100MHz bus. For the most part apps are still computer bound, EXCEPT in server space. Thats why the Xeon is still chugging along at 550MHz. Examples of apps that are compute bound.
1) 3D, games, rendering,modling, you name it.
2) Any kind of realtime graphics.
3) Photoshop type apps depending on wether you use filters more often or just edits to large files.
4) Compiling.
5) Audio editing
6) Real-time video editing. (What, I have to wait 2 minutes to render the changes!)
Things not compute bound
1) Serving webpages, files, etc.
2) Working with large photoshop files.
3) Some types of scientific computing where data crunching is high volume, low workload.
Yes I've forgotten a few tihngs, but the market for more CPU power is clearly more important than the market for higher bandwidth.
A deep unwavering belief is a sure sign you're missing something...
A GeForce should be able to run Q3 at 200 fps at 400x300 (r_mode 1) or possibly even 512x384 resolution if the cpu was fast enough. A dual willamette at the end of this year will probably do it.
We currently see 100+ fps timedemos at 640x480 with either a 1ghz processor or dual 800's, and that isn't completely fill rate limited. DDR GeForce cards are really, really fast.
Yes, it is almost completely pointless.
The only reasonable argument for super high framerates is to do multi frame composited motion blur, but it turns out that it isn't all that impressive.
I did a set of offline renderings of running Q3 at 1000 fps and blending down to 60 fps for display. Looked at individually, the screenshots were AWESOME, with characters blurring through their animations and gibs streaking off the screen, but when they were played at 60hz, nobody could tell the difference even side by side.
Motion blur is more important at 24hz movie speeds, but at higher monitor retrace rates it really doesn't matter much.
There are some poster-child cases for it, like a spinning wagon wheel, but for most aspects of a FPS, realistic motion blur isn't noticable.
Exagerated motion blur (light sabers, etc) is a separate issue, and doesn't require ultra-high framerates.
There are still plenty of things we can usefully burn faster cpu's on...
John Carmack