In this article you’ll find an in-depth analysis of Rambus’ RDRAM, what makes it tick, what the benefits are and above all we’ll look into its technological advantages and widespread misconceptions. We’ll round up this article with a broad range of real world benchmarks, which are aimed at finding the technological advantages and disadvantages, thereby checking the theory with practice. And naturally, the conclusion will summarize the results of our findings as well as give our verdict on the SDRAM vs. RDRAM issue.
Well, it is a whole lot better than the previous articles. At least it did show some benchmarks. And I should definitely start by saying: thanks.
However, it's not problem-free. For starters:
1) It is well known that the 4.20 drivers from VIA have serious performance issues. In fact, the AGP driver in those produces horribly low 3D performance, compared to the 4.17 ones. That's why VIA went back to offering the 4.17 ones, too. Why not try with those, too?
2) It's also well known that the 3.68 detonator drivers do not yet enable fast writes on VIA boards.
3) Also... why not use 2-2-2 SDRAM for the VIA board? Check it out, they ARE available. E.g., from Mushkin. (Yes, they are slightly more expensive than 3-2-2 DIMM's, but still a helluva lot cheaper than RDRAM.)
Briefly: what I see here is crippling the VIA board, to make the i820 and RDRAM show on top. At all cost.
4) Also... the fact that RIMM's are placed one after the other on the bus is good and fine. But it also means that latency increases when you add RIMM's on the bus. Yes, adding more memory on an i820 lowers your scores. Care to also show us what happens to those scores when you add more memory?
------------------
Moraelin -- the proud member of the Idiots' Guild
[This message has been edited by Moraelin (edited 05-08-2000).]
BTW, how about a comparison with DDR? I mean, ok, comparing RDRAM to SDR is good and fine, and interesting. But, well, all major manufacturers have announced DDR support already, and a lot of us have been holding off upgrades until DDR is here. It would be most useful to know what can we expect from it. Is it worth waiting for?
------------------
Moraelin -- the proud member of the Idiots' Guild
You say you don't, but your actions are contrary to that!
The issue here is not that sdram is better than rdram. No one has ever said that. The price issue is part of all of this but you missed the point altogether, which this site seems to do a lot. Of course that happens when you have deadlines and advertisers to keep happy. Lets not argue trivial issues that we know that your right in this matter. RDRAM(PC800) is better than SDRAM(PC133), just 750% more expensive.
The issue we had before is your supporting of RDRAM as the next wave of ram. There are two new technologies on the horizon that have already implemented(in graphics cards), and those are:
DDR SDRAM and the much waited for QDR SDRAM
-You have heard of these right?
Compare your RDRAM to those and let us see what technology is cheaper and faster.
I gave your benchmark results the quick once over and the BX look like its pretty close to the I820. Given the price differential I'd stick with the BX for the time being.
Jeff
I think this article was much better than lies, but I would like to see DDR done up with this type of in depth look. We are only a month or two away from DDR in MB's, ALL sorts of manufacturers are making chipsets for DDR; VIA, AMD, SiS, Ali, and even Micron is going to sell it's DDR chipset. I think this is a VERY strong group of manufacturers that can stand up to RAMBUS and Intel.
Actually, penclboy, it's a lot worse than lies, because it presents some flawed data. What we have here is a competition where one of the competitors has been crippled.
E.g., it is well known that the 4.20 VIA drivers lower Quake frame rates by as much as 25-30%. It's been known for some months now. Yet guess which drivers are used in this benchmark? Yup, these slow ones.
With the proper drivers, the VIA board scores a bit better than either the BX at 100 MHz, and the i820 in most tests. And an 133 MHz BX spanks both the VIA and the i820 with Rambus. Despite running on that SDRAM which we're told is slow. Go figure. But then, showing that in a benchmark would put Rambus in a bad light, and we wouldn't want that
Basically, I'm not necessarily saying Sander is bought by Rambus. Maybe he isn't. Let's give him the benefit of the doubt. Maybe he's just totally clueless, and totally out of touch with drivers and performance issues
------------------
Moraelin -- the proud member of the Idiots' Guild
"Splitting the control bus in this manner enhances transaction
pipelining by allowing a RAS operation for one transaction to be specified at the same time as a CAS
operation for a different transaction. Traditional technologies like SDRAMs require that Row and Column
addresses be transmitted on the same set of address lines, resulting in a resource conflict when the
memory system is placed under heavy load."
What? How can you address a row and column for different transactions at the same time? sounds like magic to me. Unless Gandalf designed the chips a RAS and CAS are needed for each unique memory address. To be able to assert them at the same time cuts out that particular latency problem. Of course if a DUAL channel rdram solution has its own full set of wires for each channel that would Almost make sense. Of course there goes the simplicity of design/low pin count for the mobo.
"The reason that this increase in bank count is
so important is that large increases in latency and decreases in bandwidth are caused by bank conflicts.
The larger bank count of RDRAM-based memory systems means that the probability of encountering a
bank conflict is smaller when using RDRAMs versus SDRAMs. "
What? Dont bank conflicts come into play when a series of memory address requests cross over from one bank to another? dont dont a small number of LARGE size banks of reduce the occurence rate this happening?
Doesn't a LARGE number of small banks Increase the occurence rate of this happening?
"Any new technology has lower yields than the incumbent
technology,"
Often correct but not universal. I.E. AMD's change over to .18 micron process and .18 copper process seems to have gone Flawlessly . When trying to push a new technology up against its limits right From the start is where many problems arise.
Woops, I posted my reply as a newtopic by accident. Here it is in the proper place
-----
"a doubt that overall it does perform better than any other chipset currently on the market. "
Excuse me, it came last in 3 out of 4 Quake3 benchmarks. More importantly, how it performs at 640x480 is almost irrelevant, as this is not a resolution that will be used by people purchasing such a system.
Secondly VIA has long been known to have a worse implementation of AGP and memory than intel. What EVERYONE out in consumer land wants to know, is why doesnt intel just release a BX based chipset with ATA-66 and appropriate multipliers for the AGP/PCI bus when a FSB of 133 is used.
So if your such an authority Sandra, please explain why a BX chipset at 133, without ATA-66, Hub Architecture or AGP 2.0 can outperform the top of the line i840 chipset?
Van Smith of Tom's Hardware gave full disclosure concerning any possible holdings he may have relating to Rambus or DDR-SDRAM. (BTW, he didn't have any.) I believe you should do the same if you wish to maintain credibility.
Speaking of full disclosure, I find it highly suspect that you chose BAPCO's SysMark 2000 as the standard benchmark, considering that Intel partially owns BAPCO. Proof resides here. As such, I'm sure Intel highly optimized this benchmark for Intel platforms, most notably the one they're currently trying to foist onto the buying public.
Considering that rather mature DDR-SDRAM platforms exist, with reference boards available from Micron, it also casts doubt on the fairness of your Rambus assessment. Hey, look, this Camaro beats a model T, so it definitely is tomorrow's vehicle! Of course the comparison suspiciously leaves out the Corvette as one of the contenders.
If you want to play with the big boys in this arena, you better be more thorough and fair-handed than you have with this poor showing. Shame, really.
1. The 820 supports a maximum of 8 open pages across all DRDRAMs in the system. If the desired Rimm is not one of those eight (and there are 128 in an 8 device Rimm), there is an additional 40-50ns wait for the chip to become active. Any program requesting data spread across multiple Rimms will encounter this delay. In some applications this is insignificant, in others, it can allow even PC100 SDRAM to outperform DRDRAM.
2. Mentioned in both aticles is the fact that adjacent pages cannot be simultaneously open. At first this seems not important considering only eight pages can be open on the i820; but, first, it would matter in future motherboard designs that could have more open pages (much greater power consuption, however). Second, reading from a small group of adjacent Rimms will require frequent switching of active Rimms because of the this limitation. For example, if data is being randomly accessed from adjacent Rimms, there would be a 50% chance that a data access would require retiring/waking up Rimms.
Though I still prefer DDRRAM, it was interesting to see an article pointing out the advantages of DRDRAM. The memory battle in the next few years will be interesting.
Well, on the theory side he did keep all the omissions and hype from the previous article. What can I say, at least he's consistent
E.g., interesting to note how when talking about the latencies of RDRAM he fails to mention the effects of taking a nap for thermal reasons. What makes this omission even more interesting is that he mentions only the active chip as dissipating 4W, which makes me think the others on the stick are taking a nap. (Either for inactivity or for thermal reasons.) And he does mention the need to spread that heat around, via that flat heatsink. Yet I don't see anywhere a mention of the time penalty for waking up such a chip to active state.
E.g., he fails to notice that while RDRAM does work with fewer traces on the motherboard, those traces operate at a much higher frequency. Hence, trace length and interference problems are much bigger for the manufacturer. Hence, that motherboard won't really be any less expensive. Add the more complicated chipset, and the mobo is actually more expensive.
Etc, etc, etc. But then, as I've said before, I'm more interested in benchmarks than in that hype anyway. Assuming he can ever bring himself to do them right
------------------
Moraelin -- the proud member of the Idiots' Guild
You are very quick to point out deficiencies in the SDRAM design but fail to do the same for RDRAM. I guess RDRAM's deficiencies have been thoroughly aired by other sites though, so fair's fair.
One thought though. In a modern general purpose computer design, many of the deficiencies of the underlying system memory are hidden by the processor cache are they not? So whereas RDRAM may be the ideal solution for the Sony Playstation for a variety of reasons, system designs based on the Pentium III (Celeron incl.) and Athlon may be less fussed.
Reply With Quote
Bookmarks