Slashdot Mirror
Intel Tests Show PC133 SDRAM Bests RDRAM
SteveM wrote citing an Semiconductor Business News article which begins: "SANTA CLARA, Calif. -- Here's a surprise. Benchmark test results from Intel Corp. show its new 815E chip set with PC133 SDRAMs beating the performance of its 820 chip set with Direct Rambus memories. Moreover, Intel has posted those unexpected test results on its Web site, not intending to show PC133 SDRAMs beating the Direct Rambus memory format, which is favored by the Santa Clara chip giant." The results actually show some fairly unspectacular differences, but those differences lean overwhelmingly in favor of the SDRAM. Surely someone will come up with a benchmark that always makes RDRAM look better.
No, I fully understand your point and agree with what you are saying. All I'm trying to point out is that benchmarks that deal exclusively with performance and do not mention cost are necessary. In this case there is a clear winner and a clear loser, but that isn't always the situation, so as a general approach it doesn't work....
2 1337 4 u!
Perhaps they're just slowly removing their ties from RAMBUS. Coming straight out and saying "we didnt know what the f*ck we were doing, we just wanted the stock options" doesn't make them look very good.
The fact that intel decided to use SDRAM and/or DDR SDRAM in its next generation of chipsets instead of RDRAM outright shows that intel knows a bit better than to push technology that is at best marginally better at 5x the cost. The conspiracy really isn't much, I quote from Tom's hardware:
When Intel 'decided' to go for Rambus technology some three years ago, it wasn't out of pure believe into technology and certainly not just 'for the good of its customers', but simply because they got an offer they couldn't refuse. Back then Rambus authorized a contingency warrant for 1 million shares of its stock to Intel, exercisable at only $10 a share, in case Chipzilla ships at least 20% of its chipsets with RDRAM-support in back-to-back quarters. As of today Intel could make some nifty 158 million Dollars once it fulfills the goal.
20% os the market is quite a bit, but intel doesn't have to be a RAMBUS zealot to pull this off. If RAMBUS really does work better, for, say, the server market, this is acheivable without incredible loads of propeganda that we've seen from them last year and much of this year.
The fact that intel itself would come out and say DDR SDRAM is better than RDRAM pretty much ends the conspiracy theory. But that doesn't mean they're still not biased twords it.
Adding a second RIMM channel also reduces the likelihood you'll take a "bank hit" in the RDRAM, and it allows the chipset to prefetch on the second channel if it thinks there's going to be a subsequent access over there when it sees an access on the first channel.
Of course, CPU and chipset designers have never been all that good at ESP. And, as on-chip caches grow larger, the traffic at the CPU boundary looks increasingly random because all of the redundant and predictable traffic has been absorbed/filtered by the cache, making ESP all the more important. (And yes, I mean Extra Sensory Perception, as in the chipset needs to psychically know where the CPU's going next.)
The other comments about making the channel wider rather than deeper to reduce latency also apply.
--Joe--
Program Intellivision!
> I agree with your entire position on Intel, but logically you cannot exempt
2 57&cid=170 ], I argued that x86 is the "open-source ISA" since anyone can use it, while Intel and HP will demand steep royalties for anyone wanting to do IA64 processors. As long as you don't have to code in assembler for it--and few code in assembler these days, anyway--there's nothing wrong with x86 since modern x86 CPUs are really a RISC core with an x86 decoder tacked on, which according to Ars Technica only adds about 1% penalty to the processor's speed. My point was that I find it contradictory that so many people hate x86, but love Intel. People just hate x86 because it's old and ugly as an ISA, but these days it's not such a real-world problem since few people code in hand assembler. ISA is really less important than how efficient the actual RISCy core of a modern CPU is; a 1% speed penalty is really insignificant in exchange for compatibility with the last 20 years worth of x86 apps, and despite people claiming for the last 5 years that x86 is going to hit a performance ceiling "soon", it still hasn't and probably won't for some time.
> AMD from your ire. While they are surely less evil than Intel, they are
> still evil for contributing to the continued existence of x86.
Actually, I never said that I personally think x86 is bad, evil, or otherwise undesirable. I used the phrase "since everyone here hates the x86 architecture so much"--and generally they do, but I'm an exception. In a recent post [ http://slashdot.org/comments.pl?sid=00/06/29/2227
So, I never said Intel was evil for pushing x86 for so long, I said that it's dumb for people to hate x86 but not fault Intel for creating a better ISA long ago. That leaves AMD in the clear as far as I'm concerned, since I'm glad they're going to extend x86 to 64bits and maintain backwards compatibility and maintain an open, freely usable ISA--putting the next big ISA into Intel's licensing control is a very, very, very dangerous idea--I'll keep incurring that 1% penalty in exchange for keeping an open chip platform, thank you. The reasons Intel is evil include its sloth, especially in keeping the P6 core so long, and its predatory M$ like nature. I congratulate AMD for starting out as having really crappy inferior processors, but making honest and huge leaps with almost every generation, almost every year, while Intel sat on its hands with the P6 core *for 5+ years*. AMD processors are now at least equal to their Intel brethren, and most benchmarks put them at a slight edge now that cache is all on-die, and in price/performance they whomp Intel completely and mercilessly.
> Quality, high-performance workstations from Sun, SGI, and Decompaq can
> be had for less than USD 5000
Yes, I agree that the PC architecture is lacking woefully, but the oppenness of that platform is what allowed the Internet boom and Information Age to happen. Cheap commodity hardware that even people who live in trailer parks can afford, but which scales up to performance powerhouses which equal the horsepower (for most applications, but obviously not all) of a RISC unix workstation for a fraction of the price. The sheer brute force and clockspeed of a commodity x86 processor, even on the hobbled buses of the PC platform, make Alphas and Ultrasparcs unnecessory for all but the highest-ed uses. It may take an 800MHz Athlon to get the FP performance of a 400MHz Alpha, but when the Athlon and its mobo are so inexpensive, there's no contest as to which is most useful. Why in God's name would I pay $5000 for a DEC or Sun box which won't run most things any faster than a $2500 x86 box I could build myself? For the elegance? Fuck elegance, give me just as fast for half the price and I'll take x86 ugliness any day. Depending on which processor the DEC or Sparc has, either an Athlon Tbird or SMP P!!!s could get equal performance for between $1600 and $2500 total, not near the $5000 for a non-x86 workstation or server. If you need those big caches, the 500MHz Xeon with 2MB cache goes for between $700 and $900, though for most applications regular P!!!s at higher clockspeed/smaller cache would be better, or a regular 1 GHz Athlon Tbird. Jeezus, one could build a Quad Xeon for less than the price of a typical DEC workstation: mobo $2500, processors P!!! Xeon 733MHz $500 each, add a hard disk and video card to taste. Unfortunately, AMD is still behind with its multiprocessor solutions...
Most PC platform problems could be cured by moving to faster and wider buses, and a Unified Memory Architecture like SGI used on its short-lived line of Wintel workstaions. And, most existing operating systems and the software which run on them would work fine with just a minor OS patch, like the one SGI used to get NT 4.0 to run on its UMA Visual Workstations.
"The more corrupt the state, the more numerous the laws."--Tacitus, *The Annals*
Uh, it's because the i815 is based on the i810 design, but with newer and more bells and whistles, and separate AGP support (i810 just used the embedded graphics controller, i815 lets you disable it). i810 was intended as a low-end "bridge" chipset for use in entry-level systems while the i820 and i840 became established.
I really think Intel wasn't backing Rambus out of any sinister conspiracy scheme - I think they really thought that PC100/PC133 wasn't going to hold up long-term in their roadmap and they needed something better. They had the Rambus investment, and didn't forsee DDR SDRAM. That's why they got caught flatfooted with the i810 being their only non-Rambus chipset and what opened the door to both Via and AMD.
I bet if they could do it all over again they would have started with the i815 as the low-end chipset, which would have both closed the window of opportunity that Via and AMD used to get business, and it would have eliminated the demand for SDRAM support on the i820 (and we all know how that worked out...), since there would have been an equivalent performing SDRAM chipset.
Even Intel screws up sometimes, though.
- -Josh Turiel
-- Josh Turiel
"2. Do not eat iPod Shuffle."
Ahemmmm, can you enlighten us on that one a bit? I'm sure it'd be an interesting topic.
And something offtopic, but it crossed my mind when I read this comment:
OTOH, I think AMD is shooting itself in the foot by making claims that it'll stick to the x86 architecture for many years to come.
I'm no longer fed up with MS Windows: I go rid of them
Look, I really hate to use buzzwords (is it ok if they have fallen out of use?) but you need to think about total cost of ownership. If you have to pay someone $50 an hour plus benefits and taxes to fix things when they break that $3500 peecee suddenly looks pretty expensive. Real workstations are much more reliable, and when they do break it's just a matter of pulling out the broken piece and popping in the new one. If you've ever worked in real hardware you know what I mean. Any repair job is 2 minutes, and there are no bloody hands and extra screws to deal with. If we're talking about individual use systems, then the TCO depends on how much you value your time. I consider playing around in cramped, cable-rat's-nest-ified, sharp-edged, poorly labeled peecee cases to be a complete waste of my time. It's well worth the extra money to have a machine that always works; and even if it doesn't, it's trivial to fix it. If you've never owned a real workstation, you can't really argue with me. Try it; you'll never go back.
AMD Athlon pipeline is only 10 or 12 stages.
Intel P4 (aka Willamette) has 20 stage pipeline, and it remains to be seen whether the high clock rates this enables makes up for the hits it'll take due to latency and branch mispredict penalty.
As if we couldn't just take Intel's word for it!
Drag n' Drop DVD Recommendations
Perhaps Intel is trying to scare Rambus through some public relations... Rambus may be demonstrating attitudes that Intel doesn't like, or thinks will threaten Intel's investment in them. Subjective qualities like becoming complacent, cocky or too aggressive in their dealings is not something Intel wants to see.
Perhaps Intel is just doing this to keep Rambus on their toes, make sure that they are always using notch 11 on the 10 notch amp, for that little bit of extra energy.
Come on people, need more conspiracy material!
OK, just for the hell of it, I'll bite. According to Tom's Hardware, Intel stands to make about $158 million off of Rambus. That's pocket change for Intel -- they probably spend more money than that on offices cleaning supplies. But by buying into RDRAM, Intel gets to confuse AMD and forces it to spend money licensing the technology that could be better spent on research. In the meantime, motherboard manufacturers scramble to license RDRAM and incorporate it into their products. Only a small number of mavericks try to stick with SDRAM after mighty Intel has spoken.
Then suddenly Intel does some benchmarks and plays innocent -- "those Rambus bastards lied to us!" So Intel does an about face and bring back SDRAM. Maybe it even buys out a couple of those mavericks (who are probably hurting for cash) and stick Intel labels on their mobos to get them out the door quickly.
Where does this leave AMD and competing mobo makers? Up a creek that's where. The big PC makers want to follow Intel's lead and go with SDRAM mobo manufacturers can't afford to switch back to SDRAM quickly enough -- Intel wipes out a bunch of competitors and solidifies its grip on the mobo market in one fell swoop. AMD is pushed away from the PC mainstream and relegated to the extreme low end and hobbyist markets -- again. And Intel thaws out Elvis in time for the launch of Itanium.
Does this
From what I recall SynchLink was 800Mb/s per pin (that is,a small 'b' as in Megabits per second). So, you'd need a 16-pin interface to reach the same bandwidth as RAMBUS. (Hey wow, that's the same number of pins as RAMBUS uses. Think that's a random coincidence? Think again.) I remember hearing about SyncLink before they'd added the 'h' to become SynchLink, and when their bandwidth per pin was still 400Mbit/s. From what I recall, they upped it to be competitive with RDRAM.
--Joe--
Program Intellivision!
RDRAM lost EVERY test except for FPU thoughput with L2 cache disabled. HOWEVER, when the cache was enabled, it got trounced! The theory of RDRAM operation and performance is ok (not great) but the reality is a different story.
There are plenty of options out there; many are surprisingly inexpensive. Quality, high-performance workstations from Sun, SGI, and Decompaq can be had for less than USD 5000, often less than half of that, which do not use x86 nor the peecee architecture. You'd better hurry, though, before everyone drops their quality architectures for IA64 and gives Intel the market chokehold it has been lusting after for years.
Fight the power; insist on quality; boycott the peecee!
Comparing similar boxes based on i815 and i820, I can get an i815 based box with 256 megs RAM for cheaper than a 128 meg i820 box, and if I even wanted to go to 256 on an i820, it'd cost me an extra $500 or so. And -- you have to be really careful. Dell has apparently been shipping PC600 or PC700 with many of their units, to keep costs down. And PC600 RDRAM should *really* be called PC534 but it's been rounded up. If it doesn't say PC800 in the "configurator," be suspicious.
Bottom line, screw minor benchmark differences, when it comes down to it, RDRAM cost is prohibitive and if you compare boxes of the same cost with the SDRAM based box loaded up with extra RAM, you'll be better off with SDRAM.
> are a few percentage points difference that big a deal?
As others have pointed out, it's a big difference in price for almost the same performance.
Worse (or perhaps "better"), we're expecting DDR SDRAM to hit the market in the fall, adding a big performance boost on the SDRAM side of the equation, with very little increase in price.
RDRAM is dead. Or at least it would be if Rambus and Intel weren't doing everything in their power to cripple the competition.
--
Sheesh, evil *and* a jerk. -- Jade
One sentence:
Intel -- Just short of intelligent.
Free Software: Like love, it grows best when given away.
Now this HAS to tell people something? Does Intel think that its end users are going to remain dumb forever. I think Intel needs to come clean as to why exactly it's still pushing Rambus memory so hard. I've seen so many benchmarks that show PC-133 on top that why a company that claims to be the processor leader would choose an inferior product.
Some people take their .sig way too seriously
Tom's Hardware did a moderately detailed benchmark of SDRAM vs. RDRAM a while back.
Which is better? It depends on both the montherboard configuration and on what you're doing.
Intel's high-end RDRAM motherboard beat the hell out of SDRAM systems. It had two interleaved RIMM slots, doubling effective bandwidth.
Intel's more recent SDRAM offerings have generally been pretty bad. Via chipsets put out a good effort, but were still beaten out by the high-end RDRAM systems and the BX board.
The best SDRAM offering was a 440 BX board overclocked to 133 FSB. Tom swears it's stable. YMMV.
As far as load is concerned, RDRAM is optimized for throughput, SDRAM is optimized for latency. Something that hits many cache rows in more or less random order taking only a little data from each will work well with SDRAM. Something that processes large amounts of data in more or less linear order will work well with RDRAM. It depends on what you're doing.
My personal opinion? RDRAM is a bad implementation of a good idea. In five years we might see something better. For now, by DDR SDRAM. YMMV.
It doesn't look like the PC133 results were massivly better, but almost all of these tests showed minor performance increases. The article states that the tests were done in the same lab, I'm kind of surprised that no one realized they were getting nearly the same results. If the same group of people did the measurements, I would think that someone would have gone "hey, those numbers look familiar." Hmmm.
---
That's sorta true - however, the 8086, which was BEFORE the 8088, had a 16 bit bus. The 8 bit bus was actually from the 8080 (or the competing Z80), which was an 8 bit processor.
The 8086 was short lived for cost reasons, so most people (i.e. you) associate the 16 bit bus with the 286.
--- "So THAT's what an invisible barrier looks like!" - Time Bandits
>That's sorta true - however, the 8086, which was >BEFORE the 8088, had a 16 bit bus.
:) The 8086 was not *before* the 8088; the 8085, 8086, and 8088 were introduced simultaneously.
:) However, the busses used varied widely. To the best of my knowledge, the IBM PC bus did not relate to any 8 bit busses . . .
That's almost true
>The 8 bit bus was actually from the 8080 (or the >competing Z80), which was an 8 bit processor.
I remember those quite well. I even have an 8080
>The 8086 was short lived for cost reasons,
short-lived? It was in wide use by almost everyone except IBM until the 286 became commonplace. At that point, it fell largely out of use, and the 8088 was used in budget machines.
>so most people (i.e. you) associate the 16 bit >bus with the 286.
Uhh, no. Aside from that I remember all of this from when it was happening, I most certainly do not make any such association.
However, the *particular* 16 bit bus that was being discussed is the IBM PC/AT bus, which was introduced with the attached 286, and extended the 8 bit bus of the IBM PC which used an 8088. There were several other 16 bit busses at the time, including Olivetti's and Vector's, which extended the 8 bit pc bus, and an extension to the S-100 favored by companies such as Compupro.
hawk
Well, consider the fact that SDRAM, and even DDR SDRAM, is considerably cheaper than RDRAM. Thus, for (nearly) the same performance, you can have a cheaper solution that holds up to the more expensive one. Now which do you choose?
And you're welcome to spend $1000+ for a Creator 3D card that's probably no faster than a $200 PC video card.
I paid $80 for mine. FFB2+. Very nice.
And then you can get screwed when you need a patch for Solaris that's only available to contract customers.
So don't use Solaris. It isn't very good anyway. Linux runs exceptionally well on Sun hardware, much faster and more reliable than on peecee hardware.
You think everyone would rather spend 3 times as much because you're too lazy to work inside a computer for a few minutes longer?
Laziness has nothing to do with it. I was discussing cost. If something takes longer, it costs more. In an environment where you're paid to do so, the costs are immediate and direct. In other environments you must evaluate the worth of your time. Personally, I'd rather just use my computers to do the work I want to do and not spend lots of time screwing around trying to get broken, misdesigned hardware to function. YMMV of course.
It's not easy to do that. Everyone has different weightings that they put on those categories. A honda gives fantastic price/performance, but if you want to win the Indie500, it is definitely not the right choice. Some people will pay a lot more for a small gain in performance because they need all the performance they can get. Others will take a significantly inferior product for even a small price drop because they just don't have the extra $100, period.
2 1337 4 u!
While many end users are actually more interested in price/performance than they are in performance per se, the idea of listing price per performance is still a bad one. There are two main reasons for this:
Both of these factors suggest that rating by price/performance is a bad idea, and that rating just by performance is much better.
There's no point in questioning authority if you aren't going to listen to the answers.
Try putting in 8 slots of fully interleaved RDRAM vs SDRAM and you will find that the RDRAM has one hell of a better bandwidth.
In some servers youll be dealing with large chunks of data, not quake 3. If you can move one bit from a to b really fast, but if its a half gig of data your fucked with current SDRAM setups.
RDRAM has its place. Mostly its 'cause most RDRAM implementations are fucked up.
-- dieman - Scott Dier
Intel's philosophy is no different from Microsoft's: Embrace, extend, extinguish. I'm just amazed that your typical Microsoft-bashing /.ers aren't Intel bashers, too, because Intel deserves a big ol' can of whoopass opened right by their corporate asses. Let's examine a little...
/.ers such big Intel fans? They're the companywhich kept pushing x86 for decades instead of developing something new and improved and more RISCy, so why so many Intel apologists and AMD naysayers? After all, as good and serviceable as the P6 core was, it didn't deserve to stay in service for 5+ years. AMD may have been a dog back then, but at least it made radical improvements with almost every product cycle; Intel just wasn't trying at all. And look at the disaster which is the new Celeron/Culeron: it may be overclockable to 900MHz easily, but because of the set associativity lost by savagely destroying half the cache like Huns sacking Rome, it barely rivals a P!!! 700MHz and gets blown away by a lower-clocked Duron too--and the Duron is also very OCable. Intel is being just as evil as M4.
First off, Intel has been in the process of developing standards for the PC architecture for some time, as well it should. However, they've doing it the same way Microsoft has been "contributing" to Internet standards. For example, they developed AGP up to 4x, which has proven to be very useful; however, rumours are churning out from reputable sources discussing an Intel project to create a successor to AGP 4x, and this successor is to be limited to Intel chipsets and chipsets made by select Intel partners--i.e., anyone who annoys Intel will get left behind. Intel developed PC-100 memory standards--a great service, but...then it refused to develop PC-133 standard or DDR-SDRAM specifications, because of its own interest in RDRAM as a wholesale replacement for all SDRAM.
Many have questioned that Intel has much to gain from Rambus becoming the new standard instead of DDR-SDRAM; after all, contrary to popular belief Intel doesn't completely own Rambus, and their deal with Rambus would only give them compensation in the tens of millions, which isn't much for a company whose revenues are in the billions each year. But what Intel has to gain isn't direct monetary compensation by Rambus, it's *control* over the standards for memory and memory controllers--and the rights to manufacture and license those memory controller technologies. This is exactly what MS did with IE--it didn't directly make a profit by developing a new web browser and bundling it with Windows; it gained market control and the ability to manipulate the Internet protocols so that all its products, from IIS to Frontpage to NT Server and the rest, had an advantage of guaranteed interoperability and increased functionality over competing products.
Intel wants to do the same with RDRAM and its new IA64 architecture, and its new forays into the emerging appliance market. Intel will make royalties on all chipsets which support RDRAM. Intel will make direct profits on its IA64 processors and has probably been hoping to licence the ISA to competitors once x86 plateaus. Intel has purchased the StronARM and other embedded/appliance hardware companies, hoping to leverage its market dominance to push it into every area. And, let's not forget that they tried and tried and tried to force their way into the graphics market, but failed there due to too-short product cycles and competitors with much more graphics experience.
It's clear that Intel wants to be the Microsoft of the hardware world. If they leverage enough tech patents on all fronts, they can force use of their products in the same unfair ways Microsoft leveraged itself into every crevice: big OEMs unable to get the best prices on Intel desktop processors unless they agree to use StrongARM in their embedded/appliance products instead of Transmeta or MIPS, or unable to get hold of ahort-supplied IA64 for workstations/servers unless they use P4 in their desktops, VIA unable to make the most advanced RDRAM chipsets unless they cut back on DDR or agree not to pursue QDR, etc. Don't think it won't happen, even with M$ as an example: there are many sneaky, below-the-board ways to hint at such matters without bluntly making demands.
And, since everyone here hates the x86 architecture so much, why the Hell are so many
"The more corrupt the state, the more numerous the laws."--Tacitus, *The Annals*
You are a stupid fuck.
Ouch.
In realworld applications, Rambus does perform better.
Where is your fucking PROOF asshole? Is Bryce 4 not a 'real world application'? How about CorelDraw 9? Naturally Speaking? Quake III? Netscape Communicator? Paradox 9.0? Photoshop 5.5? Powerpoint 2000? Windows Media encoder 4.0? Word 2000?
I think all of these are 'real world applications' and guess what, the 440bx at 133 smacks the i820 all over the fucking place. The only real world app where I saw Rambus with an advantage was excel 2000.
--Shoeboy
Maybe I'm just not enough of a hardware junkie, but are a few percentage points difference that big a deal?
I think the big deal is the fact that RDRAM is suppose to be so much better in terms of performance than SDRAM. The very fact that SDRAM matches or beats or loses by so little causes one to wonder why spend the extra $$$ for RDRAM. So, no... in terms of performance only a few percentage points don't matter. But if you look at the overall picture: price, availability, compatbility, APPLCATION.... which technology do you really need?
---
RAMBus is over eight years old. This is something like the fourth major revision (in '92 it was a 400Mhz 8bit interface). This is not a repeat of the Celeron story.
If you fully interleve the SDRAM it has pretty impressave bandwidth numbers too. Of corse that takes (about) four times as many pins. In fact 8-way interleved PC100 SDRAM excedes the bandwidth the Intel and AMD can get off their CPUs, so the only thing that will matter is latency, which'll make the SDRAM a better choice...
If you can come up with all those pins. If.
Low pin count is one of RDRAMs few remaining advantages (RDRAM systems with no CPU L2 cache run about as well as systems with small L2 cache and a normal memory system -- but that's not a good deal with L2 caches so large these days...I can list other obsolete advantages if you like). You can four way interleve RDRAM with (about) the same number of pins you need to interface to stright (not interleved) SDRAM. So if IBMs high density packaging catches on, RDRAM loses that (as more pins will be cheep). If DDR SDRAM really uses a 16 bit interface RDRAM loses it's advantage.
Of corse I don't see many chipsets using this advantage. Where are the motherbord chipsets with four RDRAM controlers?
Test them on price/performance instead of performance; for general-purpose memory, I see no compelling reason to use RDRAM except to say that you're using it. (As in, "Wow, RDRAM, that's new, isn't it? I bet that set you back quite a bit...")
:)
Now, for some special-purpose applications, RDRAM might be an excellent choice, just like in some circumstances, a P-III might work out better than an Athlon, or an 8086 might be the better choice than a G4, or a hammer might work better than a screwdriver. But for general purpose, plain old RAM, RDRAM is underwhelming.
...now watch the price of RAMBUS drop. I can hear the screams from here.
---
pb Reply or e-mail; don't vaguely moderate.
pb Reply or e-mail; don't vaguely moderate.
I wouldn't be suprised if some tech-head came and showed me RDRAM spanking the benchmarks for big server apps, but why do I care?
Really? It would shock the hell out of me. RDRAM latency degenerates rapidly as you add chips. Get a Quad Proc system with 4GB of RDRAM and you'll see some truly abysimal benchmarks. That's why intel was trying to position RDRAM as a desktop/workstation tech for Williamette (the P4) while pushing SDRAM for Foster (the P4 Xeon).
--Shoeboy
Rambus handily outperforms PC133 DIMMs, and is worth the extra expense
I think that the benchmarks make you step back and think. Do you really need to spend the money on Rambus? Think of it this way, if you were about to invest in a Rambus system just because you thought it was faster than PC133... you might be surprised to find out that whatever your application is, SDRAM performs just as good.
So, think of it in that respect, it all depends on the application and if the application warrents the cost. If your specific application won't gain anything out of it, why spend the money? On the otherhand, you might be able to rest assured that the money is well spent.......(which I know most people here won't think that way, they'll just look at the numbers, but hey that's life).
---
Open one up sometime. Look at the chips. You will see that they use standard peecee components like ATI graphics, IDE, and Goldstar (yes, Goldstar) CD drives. In my book, that makes it a peecee.
I wonder if you use Matlab, which doesn't seem you do.
It's not my primary application, no. I use my systems for development. Since I know for a fact that matlab does not run on sparc-sun-linux (I do admin matlab, I just don't normally use it), I would strongly suggest that your disappointment with your Suns is the fault of your choice of operating systems, not hardware. Solaris has a reputation, backed up by benchmarks for whatever they're worth, for offering poor performance, especially on fewer than 16 processors.
I once worked on a project to translate a matlab program into C. I do not know whether the original program played to matlab's strengths or weaknesses, but I can say that my portable ISO C program averaged 23 times the performance of the matlab version. The point? I don't think matlab is a very good benchmark. Obviously, it's your application so it's the only benchmark you care about, but I suspect that in the grand scheme of things matlab doesn't necessarily mean much. It also has no way whatever to test things like disk I/O and internal bandwidth which are nearly irrelevant to matlab but of critical importance for virtually every other application, areas in which peecees lose to any real workstation, often by a factor of 5 or more.
My point is if I have an Alpha or Sun with a 500 Mhz processor and they cost 5000 and I can buy an Athlon 900Mhz for 2500 I pefer the Athlon and I KNOW it has very good chances of outperforming the others.
Good for you. I'm glad you've found systems that work well for your application. I'm sure you'll enjoy repairing them numerous times in the six months before they stop working completely. *shrug* It's your maintenance nightmare, not mine.
The other part of your argument ... AMD is evil because they make wintel-class chips? I think not. AMD would be out of business if they made some little off-brand CPU architecture. With more than 90% of the intalled base of desktops and workstations running under the PC architecture, you'd be a fool not to consider making hardware for it! Even SGI has been moving their software to the PC platform because there's just more of it out there and they know they can't keep up when it comes to price vs. performance.
I don't know what planet you're from if you consider US$5k for a workstation (even a high-end workstation) "surprisingly inexpensive" either. I can build a pretty damned sweet workstation by any standard for US$3.5k and that's including a monitor better than my current 21" and some very nice (if expensive) input devices. You said it yourself, the PC architecture wasn't planned beyond build something that "works"(?) as cheaply as possible. Until other architectures can deliver as much or more performance at a comparible or lower cost down in the mid- to low-end workstation range as well as the high-end and our respective mothers can still play Solitaire and Minesweeper... The resurgance of unix and unix-like platforms, especially those which are developed portably and openly with such a focus on ease-of-use may as they mature make it easier to throw away the tired PC architecture. That time just ain't here yet. Until then, AMD looks like a mighty promising choice the next time I build a box.
The RDRAM is serial unlike most RAM. This means that the 800Mhz serial speed of RDRAM has the same transfer rate as 100Mhz DDR SDRAM. RDRAM can go faster but do to the serialness of it there has been problems. Here are some general peak transfer rate stats for RAM.
PC133 SDRAM: 1GB/s
DDR SDRAM 100Mhz: 1.6GB/s
DDR SDRAM 133MHz: 2GB/s
RDRAM 800Mhz: 1.6GB/s
Synchlink DRAM: 800MB/s PER PIN***
*** Reminder, this is per pin so in theory if there was 2 pins there would be a 1.6GB/s transfer rate. This RAM today is expensive and is meant for workstations/servers.
That was an extension to the 8-bit bus for the 8088, and the reason that the connector is two pieces--the big section is the old 8088/8bit section, with the smaller section added on the AT.
hawk
If you bought your machines directly from Sun, that's your own fault. You can buy them elsewhere for a fraction of the cost. The Ultra 10 is a peecee, so I don't see how it fits into the comparison. Of course it sucks, it's a peecee. Go buy a used dual CPU ultra 2. It's elegant, fast, and inexpensive. The performance of peecees compared with real computers depends, as always, on application. If you want to spin your cpu in a tight loop of integer instructions, then a real computer is not necessary. If you actually want to get anything done, you need some I/O bandwidth and floating point performance, things that aren't available in anything with a BIOS in it.
Rambus handily outperforms PC133 DIMMs, and is worth the extra expense (which means little to companies who want the extra bandwidth). Like was mentioned before, it all depends on the application. Some applications will run slower on RDRAM, it is just that simple. The problem is that Intel doesn't make any distinction about the differences.
They present it as if it is a straight upgrade path, the same as upgrading from a 486 to a pentium. They "forget" to mention that the technology is completely different, and will perform differently (sometimes radically) under different cirsumstances.
That being said, I also think that RDRAM may not be dead. Look at the celeron. The first celerons were crap. Now it is just about the most common low end processor out there. There may be a little more lag time since RDRAM isn't being developed directly by Intel, but I think that Rambus will do whatever Intel tells it to. (At least they better!)
As far as the SDRAM patent issue, I don't think Rambus has a chance in hell, and they are wasting their time and resources trying. The way US patent law is written is that something that has been in common use for over a year cannot have a patent put on it retroactively. I don't know all the details of the case, but the question rambus is going to have to answer is "Why didn't you deal with this earlier?"
-----------------------------
I've had enough abrasive sigs. Kittens are cute and fuzzy.