Slashdot Mirror

More info on Williamette
( I think Williamette is named after a river in Oregon close to Beaverton.)

Link for you:
Apples power failure.

I believe that the new intel mobile chips with speedstep actually consume fewer watts than the G4.

What's definite is that x86 slaps the g4 all over the place on both bandwidth and integer performance. And the ghz t-bird beats the 500mhz g4 by a very wide margin in floating point.

As far as your assertion that x86 instructions require a lot of transistors, that's only if they're implemented in hardware. The crusoe uses software translation. (not the same as software emulation)

--Shoeboy
(former microserf)

As for SDRAM and RDRAM, I read Toms Hardware, Anandtech and Sharky Extreme. Yes, there have been benchmarks that showed SDRAM in the lead, but I have ALSO seen benchmarks by them that say the opposite. I really don't feel like digging it up, you can either take my word for it, or go research it yourself. I'm sure their articles are well marked. I'd be curious to learn of your findings.

The fact is that even if the performance of RDRAM is better than SDRAM (and it hasn't generally lived up to expectations) there are still plenty of things to not like about RDRAM such as the heat and cost issues. Infact, just about the only thing RDRAM has going for it (in a desktop machine) is the high bandwidth and that advantage is more than nullified by DDR SDRAM.

As for SDRAM and RDRAM, I read Toms Hardware, Anandtech and Sharky Extreme. Yes, there have been benchmarks that showed SDRAM in the lead, but I have ALSO seen benchmarks by them that say the opposite. I really don't feel like digging it up, you can either take my word for it, or go research it yourself. I'm sure their articles are well marked. I'd be curious to learn of your findings.

The fact is that even if the performance of RDRAM is better than SDRAM (and it hasn't generally lived up to expectations) there are still plenty of things to not like about RDRAM such as the heat and cost issues. Infact, just about the only thing RDRAM has going for it (in a desktop machine) is the high bandwidth and that advantage is more than nullified by DDR SDRAM.

I've been hoping that someone would bring this up so that I could rehash the discussion we had on RDRAM back when this whole latency story broke. Below you will find a number of links to other places. Suspiciously, out of the holy trinity of hardware review sites (Tom's Hardware, Anandtech, and Sharky Extreme), the ONLY one that speaks up in favor of RDRAM and doesn't talk about its latency problems is Anandtech. Hmm...

From Sharky Extreme on this page:

The memory bus we are all used to operates at 100MHz and is 64-bits wide. Rambus' offering runs at 400MHz (transferring on the rising and falling edges of the clock) and is 16-bits wide. What this essentially translates into is a faster Rambus interface (in terms of frequency) with added latency because of the smaller "width" of the bus.

From Tom's Hardware: This page tells what the theoretical bus bandwidth is for SDRAM, DDR SDRAM, and RDRAM. I quote from the following page:

Continuously managing multiple latencies would be a nightmare for the memory controller. In order to work around this, when a system is booted the RDRAM subsystem performs an involved initialization process to determine what the greatest latency is for the entire RDRAM system and adjusts all RDRAM devices to have the same latency as the slowest RDRAM device on the system. And remember that in a real world system each RIMM will have many RDRAM devices so this latency balancing is quite complex.

(Emphasis is mine.) The next paragraph reads:

An RDRAM chip typically has a normal 20 ns page read access latency. To balance latencies, these chips have a TPARM control register that can be programmed with a 2.5, 5.0, 7.5 or 10.0 ns of artificial compensating latency. This means that the normal chip latency can be as much as 50% higher than the minimal 20 ns often quoted as RDRAM's page read latency. Compare this with the fastest PC100 SDRAM with a latency of only 20 ns, but again remember that RDRAM has even other issues that bring its total latency much higher still.

Finally, An article from Real World Tech explains just what the timings are like, why they occur, and why they mean that DDR SDRAM is going to be faster for the forseeable future. A very instructive paragraph on the general problems with RDRAM follows:

RIMMs also generally require a metallic heat spreader enclosure to avoid an excessive localized heating of any single memory device. Finally, the computer system motherboard into which RIMMs plug must have tightly controlled electrical characteristics that match RIMM circuit cards to avoid unwanted impedance mismatches and signal reflections. This can require extra signal layers and power planes, which along with the tighter manufacturing tolerances, results in a more expensive computer motherboard.

So let's see, RAMBUS memory has higher latency, less bandwidth, consumes more power and therefore dissipates more heat, and it's more expensive. It basically sucks compared to DDR SDRAM in every way... Where's the plus side?

I've been hoping that someone would bring this up so that I could rehash the discussion we had on RDRAM back when this whole latency story broke. Below you will find a number of links to other places. Suspiciously, out of the holy trinity of hardware review sites (Tom's Hardware, Anandtech, and Sharky Extreme), the ONLY one that speaks up in favor of RDRAM and doesn't talk about its latency problems is Anandtech. Hmm...

From Sharky Extreme on this page:

The memory bus we are all used to operates at 100MHz and is 64-bits wide. Rambus' offering runs at 400MHz (transferring on the rising and falling edges of the clock) and is 16-bits wide. What this essentially translates into is a faster Rambus interface (in terms of frequency) with added latency because of the smaller "width" of the bus.

From Tom's Hardware: This page tells what the theoretical bus bandwidth is for SDRAM, DDR SDRAM, and RDRAM. I quote from the following page:

Continuously managing multiple latencies would be a nightmare for the memory controller. In order to work around this, when a system is booted the RDRAM subsystem performs an involved initialization process to determine what the greatest latency is for the entire RDRAM system and adjusts all RDRAM devices to have the same latency as the slowest RDRAM device on the system. And remember that in a real world system each RIMM will have many RDRAM devices so this latency balancing is quite complex.

(Emphasis is mine.) The next paragraph reads:

An RDRAM chip typically has a normal 20 ns page read access latency. To balance latencies, these chips have a TPARM control register that can be programmed with a 2.5, 5.0, 7.5 or 10.0 ns of artificial compensating latency. This means that the normal chip latency can be as much as 50% higher than the minimal 20 ns often quoted as RDRAM's page read latency. Compare this with the fastest PC100 SDRAM with a latency of only 20 ns, but again remember that RDRAM has even other issues that bring its total latency much higher still.

Finally, An article from Real World Tech explains just what the timings are like, why they occur, and why they mean that DDR SDRAM is going to be faster for the forseeable future. A very instructive paragraph on the general problems with RDRAM follows:

RIMMs also generally require a metallic heat spreader enclosure to avoid an excessive localized heating of any single memory device. Finally, the computer system motherboard into which RIMMs plug must have tightly controlled electrical characteristics that match RIMM circuit cards to avoid unwanted impedance mismatches and signal reflections. This can require extra signal layers and power planes, which along with the tighter manufacturing tolerances, results in a more expensive computer motherboard.

So let's see, RAMBUS memory has higher latency, less bandwidth, consumes more power and therefore dissipates more heat, and it's more expensive. It basically sucks compared to DDR SDRAM in every way... Where's the plus side?

I've been hoping that someone would bring this up so that I could rehash the discussion we had on RDRAM back when this whole latency story broke. Below you will find a number of links to other places. Suspiciously, out of the holy trinity of hardware review sites (Tom's Hardware, Anandtech, and Sharky Extreme), the ONLY one that speaks up in favor of RDRAM and doesn't talk about its latency problems is Anandtech. Hmm...

From Sharky Extreme on this page:

The memory bus we are all used to operates at 100MHz and is 64-bits wide. Rambus' offering runs at 400MHz (transferring on the rising and falling edges of the clock) and is 16-bits wide. What this essentially translates into is a faster Rambus interface (in terms of frequency) with added latency because of the smaller "width" of the bus.

From Tom's Hardware: This page tells what the theoretical bus bandwidth is for SDRAM, DDR SDRAM, and RDRAM. I quote from the following page:

Continuously managing multiple latencies would be a nightmare for the memory controller. In order to work around this, when a system is booted the RDRAM subsystem performs an involved initialization process to determine what the greatest latency is for the entire RDRAM system and adjusts all RDRAM devices to have the same latency as the slowest RDRAM device on the system. And remember that in a real world system each RIMM will have many RDRAM devices so this latency balancing is quite complex.

(Emphasis is mine.) The next paragraph reads:

An RDRAM chip typically has a normal 20 ns page read access latency. To balance latencies, these chips have a TPARM control register that can be programmed with a 2.5, 5.0, 7.5 or 10.0 ns of artificial compensating latency. This means that the normal chip latency can be as much as 50% higher than the minimal 20 ns often quoted as RDRAM's page read latency. Compare this with the fastest PC100 SDRAM with a latency of only 20 ns, but again remember that RDRAM has even other issues that bring its total latency much higher still.

Finally, An article from Real World Tech explains just what the timings are like, why they occur, and why they mean that DDR SDRAM is going to be faster for the forseeable future. A very instructive paragraph on the general problems with RDRAM follows:

RIMMs also generally require a metallic heat spreader enclosure to avoid an excessive localized heating of any single memory device. Finally, the computer system motherboard into which RIMMs plug must have tightly controlled electrical characteristics that match RIMM circuit cards to avoid unwanted impedance mismatches and signal reflections. This can require extra signal layers and power planes, which along with the tighter manufacturing tolerances, results in a more expensive computer motherboard.

So let's see, RAMBUS memory has higher latency, less bandwidth, consumes more power and therefore dissipates more heat, and it's more expensive. It basically sucks compared to DDR SDRAM in every way... Where's the plus side?

For a technical article to not report where he's gotten his benchmark data, his historical facts etc. is not just suspect, it's ludicrous.

The other thing that strikes a weird chord is the chorus of clock speed not being the same as but being strongly corelated to performance. This is a fallacy in chip design, propigated by the truth that clock speed is strongly corelated to chip performance within a chip family but outside the family, clock frequency is meaningless.

In the end, DeMone's article is no less Intel propaganda than Apple.com is Apple propaganda. I suspect many of his claims are erroneous, since they contradict statements from third party journals, but I can't contest them seriously without sources.

So there you are.

Ushers will eat latecomers.

I used to use macs exclusively, for design / layout / everything. First the quality of the supporting media (e.g. Smackworld ecause of their obvious habit of getting high before attempting any review - and its competitor Crackuser likewise) sucked, then the sheer droning of designers in general justifying themselves with an identity linked to the presumed - and for a while confirmed - superiority of their machines, this last bit made me want to suffer Windoze more than stick around with crappy overpriced hardware.

This article which i linked to already is a indictment of Apple's complacency. It also suppports why you are right, that building an intel box may be better, but does not confirm your thought that IRIX on MIPS is equally worth eschewing. SGI is all about custom graphics hardware, backplanes bandwidth and the like. Most of their stuff, even from a few yrs ago, will just cream a G4, no matter how pretty the latter might be.

What this screams out to me is that Apple is yet again delusioned in thinking its core markets are designers and graphic pros. If RealWorld Tech is right in their analysis, it is precisely the same as APPLs desired target market WHO SHOULD IN FACT CONSIDER ANOTHER COMPUTER. Can no one whip Color Sync and get good font handling. Will RealDesigners one day finally use DisplayTeX?

You are also right in finally saying that unless you are "image conscious" - i.e. want an iMac (stupid name for anything that) you may be out of the running for APPL these days. If the high end of their desired market will hit on Intel or another arch altogether and they ream ppl on price, they HAVE to offer something really exceptional in their hardware. Underlying tech aside, there is no OS I should rather use. (especially if they reverted to ver 6.1 :) Regarding your comment about networking performance, I should only hope that BSD internal actually does show some performance. But I have no evidence to support any other thinking re that.

If APPL should do anything its what they should have done a long while ago - focus on making smart low form factor expadables (pizza boxen like the PowerMac 6100) for business users, and big ugly beasts for designers w/ like 12 PCI slots, preferable 64bit. I miss the build quality of the older tower Macs. They felt so good to have deskside. I want this feeling back - not shiny G4 shells - and I am prepared to pay for it

Do read that artic le its a sad and timely remider that maybe,just maybe APPL is still a difficult case we should be wary of.

I used to use macs exclusively, for design / layout / everything. First the quality of the supporting media (e.g. Smackworld ecause of their obvious habit of getting high before attempting any review - and its competitor Crackuser likewise) sucked, then the sheer droning of designers in general justifying themselves with an identity linked to the presumed - and for a while confirmed - superiority of their machines, this last bit made me want to suffer Windoze more than stick around with crappy overpriced hardware.

This article which i linked to already is a indictment of Apple's complacency. It also suppports why you are right, that building an intel box may be better, but does not confirm your thought that IRIX on MIPS is equally worth eschewing. SGI is all about custom graphics hardware, backplanes bandwidth and the like. Most of their stuff, even from a few yrs ago, will just cream a G4, no matter how pretty the latter might be.

What this screams out to me is that Apple is yet again delusioned in thinking its core markets are designers and graphic pros. If RealWorld Tech is right in their analysis, it is precisely the same as APPLs desired target market WHO SHOULD IN FACT CONSIDER ANOTHER COMPUTER. Can no one whip Color Sync and get good font handling. Will RealDesigners one day finally use DisplayTeX?

You are also right in finally saying that unless you are "image conscious" - i.e. want an iMac (stupid name for anything that) you may be out of the running for APPL these days. If the high end of their desired market will hit on Intel or another arch altogether and they ream ppl on price, they HAVE to offer something really exceptional in their hardware. Underlying tech aside, there is no OS I should rather use. (especially if they reverted to ver 6.1 :) Regarding your comment about networking performance, I should only hope that BSD internal actually does show some performance. But I have no evidence to support any other thinking re that.

If APPL should do anything its what they should have done a long while ago - focus on making smart low form factor expadables (pizza boxen like the PowerMac 6100) for business users, and big ugly beasts for designers w/ like 12 PCI slots, preferable 64bit. I miss the build quality of the older tower Macs. They felt so good to have deskside. I want this feeling back - not shiny G4 shells - and I am prepared to pay for it

Do read that artic le its a sad and timely remider that maybe,just maybe APPL is still a difficult case we should be wary of.

as it happens you may be wrong regards "network oriented tasks and graphics". I used to use macs exclusively, for design / layout / everything. First the quality of the supporting media (e.g. Smackworld ecause of their obvious habit of getting high before atempting any review - and its competitor Crackuser likewise) sucked, then the sheer droning of designers in general justifying themselves with an identity linked to the presumed - an dfor a while confirmed - superiority of their machines, this last bit made me want to suffer Windoze more than stick around with crappy overpriced hardware.

This article which i linked to already is a indictment of Apple's complacency. It also suppports why you ar eright, about building an intel box may be better, but does not confirm your thought that IRIX on MIPS is equally worth eschewing. SGI is all about custom graphics hardware, backplanes bandwidth and the like. Much of their stuff will just cream a G4, no matter how pretty the latter might be.

What this screams to me is that Apple is yet again delusioned in thinking its core markets are designers and graphic pros. If RealWorld Tech is right in their analysis, it is precisely the same as APPLs desired target market WHO SHOULD IN FACT CONSIDRE ANOTHER COMPUTER. Can no one whip Color Sync and get good font handling. Will RealDesigners one day finally use DisplayTeX?

You are also right in finally saying that unless you are "image conscious" - i.e. want an iMac (stupid name for anything that) you may be out of the running for APPL these days. If the high end of their desired market will hit on Intel or another arch altogether and they ream ppl on price, they HAVE to offer something really exceptional in their hardware. Underlying tech aside, there is no OS I should rather use. (especially if they reverted to ver 6.1 :) Regards your comment about networking performance, I should only hope that BSD internal actually does show some performance. But I have no evidence to support any other thinking re that.

Do read that artic le its a sad and timely remider that maybe,just maybe APPL is still a difficult case we should be wary of.

as it happens you may be wrong regards "network oriented tasks and graphics". I used to use macs exclusively, for design / layout / everything. First the quality of the supporting media (e.g. Smackworld ecause of their obvious habit of getting high before atempting any review - and its competitor Crackuser likewise) sucked, then the sheer droning of designers in general justifying themselves with an identity linked to the presumed - an dfor a while confirmed - superiority of their machines, this last bit made me want to suffer Windoze more than stick around with crappy overpriced hardware.

This article which i linked to already is a indictment of Apple's complacency. It also suppports why you ar eright, about building an intel box may be better, but does not confirm your thought that IRIX on MIPS is equally worth eschewing. SGI is all about custom graphics hardware, backplanes bandwidth and the like. Much of their stuff will just cream a G4, no matter how pretty the latter might be.

What this screams to me is that Apple is yet again delusioned in thinking its core markets are designers and graphic pros. If RealWorld Tech is right in their analysis, it is precisely the same as APPLs desired target market WHO SHOULD IN FACT CONSIDRE ANOTHER COMPUTER. Can no one whip Color Sync and get good font handling. Will RealDesigners one day finally use DisplayTeX?

You are also right in finally saying that unless you are "image conscious" - i.e. want an iMac (stupid name for anything that) you may be out of the running for APPL these days. If the high end of their desired market will hit on Intel or another arch altogether and they ream ppl on price, they HAVE to offer something really exceptional in their hardware. Underlying tech aside, there is no OS I should rather use. (especially if they reverted to ver 6.1 :) Regards your comment about networking performance, I should only hope that BSD internal actually does show some performance. But I have no evidence to support any other thinking re that.

Do read that artic le its a sad and timely remider that maybe,just maybe APPL is still a difficult case we should be wary of.

I have beeen real close recently to shellingout for a shiny new G4, thinking hard how much I want my hands on OSX.

Then yeasterday I read this analysis of G4 / mot performance and im starting to think all over again.

Im feeling pretty bummed out that Apple could be so misleading with their advertising.

Looks like they are over using a whole bunch of ppls goodwill to remember them as an idealistic company so not to question the assertions they make. I should welcome myself to reality or something.

If I were you, I'd have a mid-range Thunderbird at the top of my list. It will debut at 700-1000 Mhz, and probably hit 1.1 GHz soon thereafter, so we're probably talking somewhere in the 800-900 MHz range. The process will have a lot of headroom in it, so you'll probably be best off getting a 700 MHz T-bird or Duron and overclocking it, assuming you're comfortable with that.

The big question with this system is whether DDR is worth it. The actual cost of DDR should be barely more than the cost of PC133--it's just as easy to make, but it might cost a bit more because of limited supply. Instead the "cost" of going DDR is measured in time; DDR mobos don't look to be available until late this summer, a month or three after T-bird is released. If you can wait, don't buy until you see the first benchmarks of T-bird on DDR vs. T-bird with PC133. Don't make your decision on benchmarks of the PIII with DDR, which will probably be available sooner: the PIII's chip-to-Northbridge bandwidth is only 1.07 GHz/sec, so it is usually saturated by PC133. DDR ought to show a bigger performance increase on the T-bird's EV6 bus, which will run at 266 to match the DDR. On the other hand, don't plan on going DDR before you check out the benchmarks; no one really knows exactly how well it will perform.

As far as chipsets, you want to be looking at a VIA KZ133 if you're sticking with PC133 (NOT the KX133, which is apparently incompatible with the T-bird), and probably a KZ266 if you're going DDR, although there may be other DDR chipsets available for T-bird as well, most notably Micron's Samurai chipset.

The big unknown in the future x86 market is Intel's new Willamette core. It probably won't be out in volume until January or so, and thus might not be an option for you. On the other hand, they'll be looking to do a paper release in late summer or early fall to compete for press time with AMD, so we should have a good idea of how well it'll perform months before we can actually buy one. (And maybe if you're lucky and willing to pay a lot, you might even be able to buy one in 2000!)

The reason I mention it is that while the general consensus seems to be that Willy won't be such a huge deal, Paul De Mone, one of the most respected semiconductor analysts on the net, has had some very positive things to say about it. If you're interested after reading the two mondo articles there (and if you're any kind of hardware geek, I can't see how you won't be), you might want to check out what he's had to say about Willamette in Ace's Hardware's technical forum. Frankly, Paul really knows what he's talking about, and he seems to think Willamette will solidly give Intel back the performance crown. Whether Intel will use that as an excuse to price Willy out of the upper-mainstream market where you're looking to buy is another question entirely.

Any other variables I should be asking about?

Well, depending on what you're looking to do with this computer (and with what OS), the video card is probably the most important component these days--certainly more important than chipset and arguably more important than CPU. If you want good 3D performance, then at the moment that means running Windows. Period. This is changing relatively quickly, though, so it may be less of an issue when you get your computer. (It will NOT be a non-issue.) 3dfx has always had some of the better Linux drivers, so if you're going Linux one of the new V5 cards is probably your best bet. nVidia is well known for having terrible, and closed, Linux drivers, although they claim that that's changing. If you're going to be running Windows, a DDR nVidia GeForce 1 is probably going to be the best bet to match your computer (i.e. just-below-really-high-end).

If you're going to be running Windows games a lot, then this will be the most important part of your purchase, hands down. The two obvious choices will be the V5 and the GeForce 2; the GF2 is a bit faster with full-screen anti-aliasing off, while the V5 is faster with it on. The GF2 (and GF1) has T&L to speed up future high-poly games; the V5 has Glide which provides the fastest play in games like Ultima IX and anything using the Unreal engine (Unreal Tournament, plus many upcoming games like Duke 4). To really decide, however, you should go to some indepth benchmarks from a gaming site and look at the resolutions/settings you'll be playing at in the games you'll be playing and see which card performs better. Also ATI has a card aimed at the high-end 3D market due out this summer.

If you don't need good 3D, take a look at an ATI card for great DVD playback (depending on your ethical opinion thereof), TV-tuner, and general features, or take a look at a Matrox G400 for top-notch quality and the best dual-monitor support around.

Other than that, you should be fine performance-wise. (I'm sure I don't need to tell you that the monitor is the most important piece of equipment for overall computer satisfaction, and that a nice keyboard and mouse are close behind.) Of course, for "general purpose developer's desktop plus scientific number crunching in the background"--i.e. compiling stuff and running distributed.net--IMO any computer sold today is more than adequate, although of course extra compiling speed is always nice, as is a higher ranking. In this case, I'd say the most important factor is how much cache is typically consumed by compiling. (Anyone care to inform me?) If it's a low-cache amount low-memory access activity, go with a Celeron or Duron; if it's low-cache amount high-memory access, go with a Duron; if it's medium-cache amount, go with a T-bird or Coppermine; if it uses up a whole lot of cache you may be better off with an Athlon "Classic" or a Katmai PIII for the money.

Hope this helped!

That aside, I still can't find any specInt or specFP #s... which are a fairly good representation...

That's because SPEC doesn't run on MacOS. Thus the only PowerPC SPEC scores available are for high end IBM type systems, which have far superior i/o systems to Macs, and thus ought to perform much better on memory-intensive benchmarks like SPEC.

Unfortunately, the latest high-speed PIIIs kick the crap out of them, as do the actually-purchasable GHz Athlons. AltiVec is a great SIMD design, but even if there was enough software to take advantage of it, it'd be starved for data by the Mac's pitiful system bus in most real world conditions.

Basically, the only reasons to buy a Mac for your desktop is the OS, the color, or the fact that the new iMacs don't need fans. The only good points of the PowerPC architecture are low power consumption (which makes it a great laptop chip, and allows for the lack of a fan) and AltiVec (which, again, is starved by the poor i/o system).

For a much more detailed explanation of what I just said (including PPC vs. x86 SPEC comparisons), see Paul DeMone's excellent article.

As for the best SPEC/$ platform, right now it's far and away the GHz PIII. In the near future, it will be the Thunderbird on the low end and the new high-clocked Alphas (finally) for FP stuff. In the medium term, Willamette just might post some amazing SPEC results, although it's too early to tell. In any case, it's clear that the PPC architecture will lose badly for at least the next 12 months or so.

Besides, SPEC has a lot of things wrong with it as a benchmark, not least of which is the fact that even the SPEC "base" tests can be compiled based on optimizations from a previous test run (this is why Itanium will have good SPEC scores but terrible performance), and that it has no graphical component, and that it doesn't model typical code very well, unless you spend your time running scientific simulations all day.

On the other hand, if there were a better cross-platform benchmark, it would also show that the PPC sucks. As a CPU for a personal computer, that is. As well it should: it's designed for embedded/signal-processing applications.