why are Sandra results so bogus? they claim to be running the Stream benchmark, but the numbers are wrong. for instance, pretty much any FSB100 BX-based celeron/pii/piii/etc will deliver around 400-450 MB/s running the real Stream code. and Athlon or i840 systems deliver around 550.
Dell, known for its particularly close ties to Intel and hence
Rambus, has also performed some tests. Their results show that under many conditions
even PC100 SDRAM beats DRDRAM. http://www.inqst.com/Dell-Rambus.htm
I don't particularly find fault with the general thrust of the article, which seems to be "here's what's good about Rambus and what's bad about SDRAM". There has certainly been alot of investigations into the reverse. But there are some points that stick out in my mind.
Sander, again you insist that "most of us" are accusing you of being "bought" by Intel or Rambus. Again I find this untrue. While a couple of messages have made that accusation, most have addressed the points in your article. Implying that you're being persecuted makes it easy to dismiss the points being made that are contrary to the ones in your article.
I find it ironic that you basically used the benchmarks that many other sites have used. Toms Hardware also does full disclosure of hardware, software, and drivers that it uses for its tests.
I also find it ironic that, for all of the advantages you found for RDRAM (and disadvantages for SDRAM), it could only muster a slight lead in BAPco SYSMark, and was edged out in the Quake III benchmarks. This despite the fact that you used the slower CAS3 PC133 SDRAM against the fastest current variant of RDRAM.
You ask that we "step away from the endless price/performance discussion and open our eyes to the potential Rambus Direct RDRAM has to offer". This seems to be saying 'be willing to overlook the criminally high prices, because one day RDRAM will be really fast!'. That is hardly convincing. The fact is that RDRAM is prohibitively expensive, even after recent price drops. And that must be factored into any buying decision, when you realize that the performance difference is basically nil (except in a synthetic benchmark, yet another bit of irony).
The comparisons to SDRAM are also forgetting something. When SDRAM was introduced, Intel did not force motherboard makers to drop SIMM support overnight. Many motherboards had both connectors, and it was still possible to get TX and VX motherboard with SIMM sockets, and not have to suffer a performance penalty for its use. With the i820, Intel wanted to force a much costlier DRAM technology on consumers and leave them with no choice. When they finally relented to allowing SDRAM on i820, they did it in a manner which cripples performance. Intel's message to consumers was clear- pay too much, or get too little (or both, to be honest).
And considering that you are evaluating RDRAM's potential as a future technology leader, you neglect to mention, even once, its likely adversary, DDR/QDR SDRAM.
I think that your article is, at best, incomplete. At worst, it is a slanted attempt to promote a technology that is being almost universally panned with good reason.
Sander apparently has leant nothing from
the previous discussion. The article
is really a piece of job! I would not
call it a fantasy or lies, but just
simply a job which apparently was
well paid. Othervise no one in sober
mind is capable to assemble so
much of misinterpretation and desinformation. Just reading the
article across the diagonal, I found at least 6 instances of stretching things
beyond any comprehension.
Let's start with "uniform-impedance transmission line"
Where the heck he got this idea? First, it
is theoretically impossible to provide any impedance
match between simple connection of two wires (bus!) to any single
input pin on the Rambus chip. People who have even
little experience with coaxial TV cables
know this, and use inpedance matching
splitters which attenuate signals BTW.
Rambus people also started to realize
this in their "third generation Rambus
technology" and arrived with NON-UNIFORM-
Impedance transmission lines, where some
areas of connectors are made narrower, so
two wires connected to the chip "load" add
up to the same 28 Ohm. The so-called "loaded"
portion of the tracing requires careful simulation
and analysis to minimize reflections that
ineviably present in the inherently mismatched
design. When you have 32 of those stumbling
blocks in the chain and multiplied by manufacturing
deviations, the signal quality
become far from theoretically desirable.
So much for the technical accuracy of the
Next: low pin count and "cheap" trace routing.
Pin count may be low, but traces are
not cheap by any measure. Because of UHF
signalling frequency that exceeds our
broadcast UHF radiostation's frequencies 8
times, the traces are REQUIRED (BY RAMBUS and
INTEL) to have ground guards of certain size
along the whole and each trace. So the number
of wires effectively doubles ot least, and no
freedom is allowed in routing, every turn
counts, every via.
Next: "2-cycle addressing" - again, where he
got this I wonder? The Intel document he links
contains NO INDICATION ABOUT ANY 2-CYCLE
ADRRESSING PROBLEM. None. Zilch. As I tried
to explain to him earlier, a SDRAM
controller MAY CHOOSE to use FREE CLOCKS to
drive addresses in advance since those two
cycles ARE REQUIRED ANYWAY to maintain the
standard timing. This leads to very relaxing
requirements, and this is why almost
any PC100 has a room to work up to 160-170MHz,
Next: Sony Playstation2. The reference to
this Rambus application is at least silly
on the PC site. As he states himself, the
PS2 uses two channels WITH SINGLE RDRAM CHIP
IN EACH CHANNEL! For single chip there is none
of the problem that plagued RIMMs! The depth
of this misunderstanding is just shocking!
Next: Granularity. What the nonsense it is!
Who would ever need a system with 96MB of memory
instead of 128MB? Or 256? What is the difference?
Even more, it was recently reported that Intel
is having problems on i840 boards when using
4 and 6-chip RIMMs, it just does not work, and
the problem has not been solved for half a year!
See "errata 28" in
"Errata" 28 description:
"In system configurations with a single 4 or 6
device PC800 RIMM installed in each channel,
the system may experience memory errors...
...observed behavior can range from being
completely transparent to the user to causing
Workaround: Configure systems with PC800 RIMMs
with 8 or more devices... or use PC600 RIMMs"
Next: "max power of 4.6W is smallest among
I don't know how Rambus marketing arrived
with those numbers, but experience shows that is
you do not blow a fan over the RIMMs, you can
burn your fingers off the thermal spreaders!
Nothing close have been experienced with any
SDRAM, they run slightly worm, that's it. For
the RIMMs, every chip incorporates a singal
that flips if the Rambus chip experiences
excessive heat, so the controller can slow
down the rate of data activity and prevent RIMM
from delamination and desoldering. There is no
need for such sensor in any SDRAM yet. The fact
of "smoke detector" in RAMBUS itself speaks pretty
loudly about which memory consumes more. It is also
worth to mention that during patented initialization
process, the RIMM may consume up to 7.5Amperes of
current, so Intel recommends to take a special care
to RDRAM local power supply.
Again, so much for low-power design.
I guess this is enough to say. Only a highly
compensated person can write such a misleading
and erroneous article as Mr. Sander did. His
continous reminders that he has some EE education
become really annoying too. He apparently "tackled"
not too much about the real technical issues that
really hammers the lookin-good-theoretically
Now that dual channel 800mhz rambus I840 look
Its weird that most, if not all the review I seen show rambus clearly slower.
In this review rambus is only barely faster then PC100...
Humm... I guess its worth paying 560$ for 128meg instead of 88$ for PC133.
Intel should look at a 440BX running at 150mhz... Humm, wonder why they detail the
i815 so much... could it be that it will win all the benchmarks? unless they are now looking at crippling PC133 so rambus dont look bad.
I'm very pissed at the rambus situation. Intel waisting the industry time!
Great we had a 64bit FSB and a 64bit memory bus in 1994 on the pentium...
Its Mid 2000.. where are we at? 64bit FSB,
In 1994, video card had a slow 32bit bus.
Today, 166+mhz 128bit DDR, some with multiple
bus & advance bufering.
Many people have pointed out serious flaws with the limited scope offered by this article. Showing BX @ 100Mhz FSB is not an appropriate comparison. A far more appropriate comparison would be to compare i820 with BX @ 133Mhz FSB and challenge Intel for ever releasing a memory scheme the failed to deliver the goods. My BX based board runs @ 133Mhz just fine (PIII-550E OC to 733, it will run faster except I have to lower ram timings to 3-2-2 from 2-2-2 @ 133Mhz) and my AGP card also works fine (Matrox G400 MAX). The Quake frames are great and the Sandra memory tests yeild 392/437 which is directly comparable with the i820 equipped with PC800 RDRAM @ triple the cost.
A challenge should have been issued to Via to improve their memory and AGP performance to the levels offered by the 2+ year old BX. Another challenge should have been issued to Intel to deliver an RDRAM platform capable of outperforming their old BX clocked @ 133Mhz by more than 20%. If Intel could not deliver such an RDRAM solution (which they clearly have not done), then Intel should deliver an update on the BX using state-of-the-art processes (to get the potential clock speed up to between 133-160Mhz and get the heat down), integrate ATA-100, AGP 4X (if not AGP Pro), and offer an AGP divider. Clearly RDRAM is NOT a performance leader for consumer level computing and offers very little to even workstation level computing. Via will clean up if Intel continues to push RDRAM with poor performance and high cost in its consumer products.
Clearly there was an agenda in the choices of comparison. While I am not a Tom fan, his analysis of the situation is much more tuned to reality and is unaffected by marketing schemes.
The addition of the high-speed interface results in a small increase in die size, but also allows higher performance to be obtained.
Higher than what…PC100? OK, certainly, but not even equal to DDR at 133Mhz.
In contrast to current SDRAMs, memory systems that use RDRAMs transmit addresses and data in a wave-pipelined manner on both edges of the clock.
They mean like DDR.
The high per-pin and per-device bandwidths of RDRAMs have several system-level advantages. Providing high bandwidth from a small number of pins means fewer bus traces to route. In addition, the fact that the Rambus channel is routed in parallel means simpler routing and reduced complexity of motherboard designs. Often, simplified routing can result in fewer motherboard layers, reducing manufacturing costs as well.
In reality, the higher bus speeds along with the protocol nature of RAMBUS cause the motherboard design to be MORE limited often requiring additional layers to be added to the board to allow a lower signal to noise ratio. While I personally like to see all my bus lines straight from point A to point B, it is nice to be able to move one around a bit to fit things in if needs be. This luxury is not an option with RAMBUS. It is very sensitive to impedance differences from one line to the next. The net result is that the system is more expensive.
The high price is often attributed to factors including low yields, large die sizes, the cost of new equipment, and royalties charged by Rambus. However, several of these explanations have recently been called into question. An interesting point is that the price premium for RIMMs in the retail channel is higher than the premium for RIMMs purchased from the major OEMs. The Dell website, for example, allows prospective buyers to configure systems with varying amounts of RDRAM. The cost of upgrading a system to include an additional 128 MB RIMM module is much lower than for purchases made through retail channels, and has even been dropping recently. This calls into question the exact cause(s) of RDRAM price premiums. Central to this controversy is the difference between price (what a consumer pays for RIMMs) and cost (what it costs a manufacturer to produce RIMMs).
LoL. I love this stuff. On the DELL web site it shows a $480.00 128mb upgrade. This is to compete with 128mb of PC133 at $89.00 on price watch (granted, this is cheep memory, but so is the memory in the DELL. I have heard it is only PC700 at best...at worst PC600. By my math that is a 539% increase in price……hardly a bargain in my book.
Any new technology has lower yields than the incumbent technology, and has development costs that must be recovered, so one would expect some premium to be paid initially for RDRAMs. But this article suggests that the cost of manufacturing RDRAMs is not the reason for the RDRAM price premiums in the retail market.
Although the first sentence is true, RAMBUS has inherent yield issues. The higher clock frequency needed (currently 4X the clock as SDRAM) makes the design much more expensive…..but everyone knows this. A 4x faster processor will certainly cost more and have lower yields than the slower one. It will additionally have more sensitive traces, more expensive mb design, higher heat dissipation, and larger power requirements.
As an aside, it has been rumored that RDRAMs are first tested when they are placed on RIMMs, and that low device yields mean that the entire RIMM must be thrown away if one bad device is found. This is not the case. RDRAMs are tested in much the same manner as SDRAMs, with functionality being verified before devices are assembled onto RIMMs. The high operating speed of RDRAMs, which exceeds the operating speeds of previous DRAMs, means that new high-speed testers are needed to test some functionality. While there is an investment for this new capital equipment (which must be recovered), it has the advantage of being able to test RDRAMs at full speed, meaning that test time per RDRAM is reduced and test throughput is increased. Agilent describes this advantage in a recent press release.
Yea, yea, sure, sure. You can test each chip prior to assembly; however, you still have to validate the entire package once assembled. RAMBUS is so sensitive to timing, that manufacturing differences in one PCB to the next can affect the max speed rating of a RIMM. Besides, this is an invalid arguement since the entire assembly must be tested anyway right?
Analysts estimate the royalty Rambus receives to be about 2% for each RDRAM sold. This is very small compared to price premiums seen in the retail channels today, and is thus not a major contributor to these premiums.
I have heard 5%. We must be talking to different analyst. As I understand it, the royality percent is on a sliding scale. You get a better rate if you sell more. Intel is on the hook for quite a few of these jewels. With their projected volumes, 2% may be an accurate number, but even so, this is still more than the 0% of SDRAM or DDR.
Dean Kent at realworldtech (the links from the above posts) has always been a pretty non-biased source of technical information. I don't consider his analysis to be discounted easily.
RAMBUS IMHO is simply not a viable solution when compared to DDR. It is not really a good solution compared to PC133. Give the incredible amount of technical information against RAMBUS memory, it is understandable how some have questioned your motivation. I agree with UncaMilty on this one (nice post by the way)
I think that your article is, at best, incomplete. At worst, it is a slanted attempt to promote a technology that is being almost universally panned with good reason.
Any defense or replies to my arguments would certainly be welcomed.
[This message has been edited by OneEng (edited 05-10-2000).]
With greater knowledge comes greater understanding!
You seem generally offended! I appologise for my comments, here is a sucker and your balloon back. When you offer a section for feedback and get emotional when people disagree with you and point out your flaws and mistakes, you are not a balanced person!
First of all let me confess that in this case I am definitely guilty of being a bad reader. I went to the index of the article and jumped around a bit to review the conclusion, then look at some of the supporting benchmark charts, and config lists. I wasn't really interested in the technology overviews, I've read several on this topic.
My reason for coming to this discussion thread was to chide Sander a bit because I didn't find the information I was looking for in a consistant, easy to read, format. Some illustrations (particularly the screenshots) were not readable, even when I linked to display them individually. I also didn't see the kind of head-to-head charts I was expecting; the benchmarks I saw were shown individually. I might as well mention that, overall, I thought there were too many words, and that I didn't like some of the colors in the pictures either.
When I got here to post regarding some minor qualms concerning form, I found that a war was in progress concerning content, and opinion.
As a quick aside here I'd like to mention that I have never seen a thread with more first-time and "single digit", posters in it. Clearly this article has drawn a new audience, or inspired many who have never posted before. From that point of view: well done Sander.
While most of the articles here can be categorized as one of the following: previews, reviews, or buying guides, this one doesn't really fit any of those classifications. Obviously in hop-scotching through the article I missed the part that read: "Buy this stuff, buy it now, pay any price. All Hail Intel, and all of its minions!"
I'd suggest to any of you who have problems with this article, that you explore this site for earlier reports on the same topic. I pretty clearly remember one which left me thinking that RDRAM was not only way too expensive, but that any benefits provided by the underlying technology were dubious, at best. Perhaps then you would see that this piece is simply adjusting back towards center, not leaping off the far right edge.
Well, you see RoadWarrior, that is the main difference between us. You jumped around, looked at the pretty pictures, and missed the part that says "all hail the Holy Rambus." By contrast, I did wade through all the hype. And through carefully selected facts which, although very few of them were blatant mis-information, also suspiciously omitted all the bad facts about Rambus. Like impedance and EMI problems on the lines, the penalty of waking up a chip, and so on and so forth.
Basically it's like me telling you car A is great because of this and that. Look what a neat engine it has, and look how fast it can go, etc, etc, etc. But carefully omitting any reference that it burns more fuel than a Boeing 747, that its acceleration is horribly low, and it's unstable when taking a turn.
Basically, no, it doesn't spell out "go buy RDRAM, damnit!" But for the last few articles on the subject, they all sound like a paid advertisment for Rambus. It's even got all the hype and buzz-words of an ad. In many cases: word for word.
Again: I'm not saying they're actually paid ads. That I cannot know, of course. But if you actually read it (as opposed to skipping to the cute pictures), it looks and feels like reading an ad.
Dunno, maybe those companies Sander asked for info sent him ad brochures instead. I'm sure there's some reasonable explanation there
Moraelin -- the proud member of the Idiots' Guild
-RoadWarrior, the objections being raised by many have nothing to do with whether the article in question is simply a "rah-rah" puff-piece on RDRAM. There are serious objections being raised about the accuracy of the article and the technical points that were raised. These are valid concerns, and not just character attacks (which I consider to be in the minority, so far).
-There are some who feel that it's unfair to characterize HWC as "pro-Rambus/Intel", due to the fact that other articles and reports have taken a decidedly different tack on the subject. That's all well and good. But it doesn't have anything to do with whether or not the articles in question are fair or not, or whether they're based on fact, or whether the technical explanations are inaccurate (or just wrong).
For example, if a site that is primarily pro-AMD decides to post an article that takes a more positive view of Intel, hurrah. But if that article contains false information and misleading comments and data, it isn't excused just because it went against the grain. I think that some folks have raised some interesting counter-points to the article, and I think that those questions are valid. It doesn't matter that HWC has shown balanced reporting on the issue. A bad article is a bad article.
This article is well written, but poorly thought out. As mentioned by many commentators, comparing Rambus to SDRAM is a sham argument. Double Data Rate DRAM is the technology you need to compare Rambus to. It is not even mentioned. SDRAM is the old technology, and you are comparing it to Rambus' new technology. Why not do a comparison of new vs. new? Answer? Because Rambus would come out looking terrible. It would loose in the areas of cost, interoperability, power consumption, and heat dissipation, and yet the bandwidth is about the same. This huge article was a waste of time, because it compares apples to oranges. The author will no doubt argue that Double Data Rate DRAM is not available for main memory systems yet, but for the vast majority of consumers, Rambus is not available or affordable yet either. Double Data Rate DRAM-containing AMD "Thunderbird" systems running on the AMD 770 chipset will be available in 5 or 6 months.
Intel's recall of 1 million i820 motherboards yesterday attests to the fact that trying to get Rambus and SDRAM to work on the same motherboards is fraught with pitfalls. Most of Intel's current problems can be traced directly to their pigheaded adoption of Rambus. Articles like this that compare Rambus to older technologies in an attempt to bolster Intel's untenable position just help make the case that Rambus is the wrong new technology. An honest article would have compared Rambus DRAM with Double Data Rate DRAM. In any case, consumers will vote with their hard earned dollars, and eventually DDR will win out. In the process, Intel will loose more market share to AMD and VIA. Consumers want more power for the same, or less money, and don't want to pay more money without a significant performance boost, which Rambus does not offer in it's current form.