RAMBUS still king!

[speedy4500]

Check out the memory bandwidth comparison between the Asus P4T533 (32bit RDRAM) and the new Granite Bay DCDDR here . People who think DCDDR is the be-all end-all are going to be let down. As an owner of the original DCDDR chipset (nForce), I can attest to this. The fact is, dual channel memory architecture has proven to be difficult to implement effectively. Keep in mind Rambus's whole product line revolves around dual channel memory architecture, and they've been working on it for maybe 4 years now. With RDRAM's falling latencies which are now comparable to DDR, I would definitely go with RDRAM if I had the cash. The Granite Bay board is actually surpassed by single channel DDR in a few benchies on that site, if I remember correctly.

 

[theflyingrat]

Well, it looks as if the biggest lead RDRAM ever had in any of these tests was in the SANDRA memory bandwidth benchmarks.
Know what that means?
Approximately zero. Especially since the difference there, even, is only less than 1%.

Sure, RDRAM does have the lead in most of the real-world tests, but by a very, very, very narrow margin.
Quake III demo - RDRAM wins by .8 fps. That's eight tenths of a frame per second. Push.
3DMark 2001 - DCDDR wins by 5 points. Push.
Comanche - RDRAM wins by .04 (four hundredths) fps. Push.
CodeCreatures - RDRAM wins by .3fps. Push.

The comparisson to nForce has no validity here. They are chipsets for ENTIRELY different CPU architectures. nForce would have been a much bigger success if the Athlon XP was as bandwidth-hungry as the Pentium 4.

The beautiful fact about Granite Bay is its ability (like in the above tests) to run cheap PC2100. 512MB of PC2100? That's two 256MB sticks. $150 for two sticks of Crucial (very good memory by reputation; I also know this firsthand.)
Single stick of 32-bit, 512MB RDRAM? Cheapest I could track down on Pricwatch is $285 for Samsung. That's a $135 difference. Now, even if the GB boards cost $80-100 more than 850E boards, you still end up saving some cash, while basically getting the same performance, and in all likelihood, much better overclocking results.

Conclusion? Granite Bay, in all likelihood, will be a great success, even though it is marketed towards the workstation market, unless there are some rather great compatibility or bug issues with it. And RDRAM will die because of this. RDRAM had its time in the sun, where it was leaps and bounds ahead of everything else, and its huge cost penalty could almost be justified. If board manufacturers make some cost-effective GB motherboards, it's basically all over for Rambus, at least in the consumer PC market.

 

[masitti]

That's why I've never gone for the P4/RD combo - I went with the XP and DDR instead. It was a helluva lot cheaper.

Like said, RAMBUS is so expensive, and for the little performance gain you get, it just isn't worth it. Likewise for the P4 - it may have a little better performance, but the price to performance ratio can't be beat.

 

[larva]

some good points, but the real champ is 845

Well, it looks as if the biggest lead RDRAM ever had in any of these tests was in the SANDRA memory bandwidth benchmarks.
Know what that means?
Approximately zero. Especially since the difference there, even, is only less than 1%.

I agree 100%. Use synthetic low level benchmarks very carefully, and don't try to read too much into the result.

Originally posted by theflyingrat
Sure, RDRAM does have the lead in most of the real-world tests, but by a very, very, very narrow margin.
Quake III demo - RDRAM wins by .8 fps. That's eight tenths of a frame per second. Push.
3DMark 2001 - DCDDR wins by 5 points. Push.
Comanche - RDRAM wins by .04 (four hundredths) fps. Push.
CodeCreatures - RDRAM wins by .3fps. Push.

Another good point, as the application tests show a near parity between RDRAM and the DDR alternatives. But the really important observation is how well the 845ge board fared. It actually beat both RDRAM and dual channel DDR in Quake3, and bested the dual DDR setup in 1 other test. Of the four remaining, the 845ge machine finished with essentially identical (although fractionally lower) score as compared to dual DDR, only being bested by a substantial margin on one of the six tests. And since the only reason I bother to keep my fast PC fast is that I play Quake3, you can see why I just bought a board some would decry as obsolete, a BD7-II. Aided by PC3500 memory performance (achieved with the cheapest PC2700 ram I can buy, which is no more expensive than a pair of half-sized PC2100 modules), this is at least the equal of the RDRAM and dual DDR alternatives, and may actually be unmatched in some areas while at no point being seriously outclassed.

Originally posted by theflyingrat
The comparisson to nForce has no validity here. They are chipsets for ENTIRELY different CPU architectures. nForce would have been a much bigger success if the Athlon XP was as bandwidth-hungry as the Pentium 4.

I disagree here, there is a valid lesson to be learned from nForce and nForce2 boards. The lesson is the performance of either is nearly identical between the single channel and dual channel implementations. The basic virtue of the chipset is the biggest factor, not how big the numbers are. Granite bay is here to remind us that this is not a characterstic unique to the nForce. It also has a desperately tough time outperforming the single channel 845 series, even though the numbers are twice as large. And the only reason a P4 is "bandwidth hungry" is that it is faster, with a substantial edge in working frequency. The faster a chip processes data the more difficult keeping it stocked with data to process becomes, and the greater the gains possible through higher memory througput. But as this is problem wrought by the extreme speeds these cpus's attain, there are worse problems one could have.

Numbers don't really matter... I see many folks actually asking which dual channel DDR or AGP 8X products they should replace their (usually perfectly fine) existing motherboards with. The truth is if your current rig is optimized for value presently, none of this new stuff changes the selection. A properly built high fsb 845e, pe, or ge system with PC3500+ memory speed will at worst equal the newer alternatives, and in some cases outperform them. Just like the tiny performance edge PC1066 RDRAM shows doesn't make it the platform of choice for most, neither does the edge possesed by dual channel DDR. In reality getting a single channel DDR system running the ram at 400MHz and up is just as valid a solution to the (only comparitively) limited bandwidth as going to the hideously expensive dual channel RDRAM or new (and therefore costly) granite bay dual channel DDR solution.

If I needed a board already I might consider granite bay, but only after it has matured a bit, the price has relaxed, and the overclocking ability has been proven to surpass what the 845 series allows. As we've seen at equal clock rate granite bay's edge varies between unimportant and non-existant, so it better be able to keep up in the MHz race lest it be embarrassed by the lowly old 845e. To rip out a perfectly optimized 845 board at present to pursue the "2" implied by dual channel DDR and the "8" in 8X AGP is entirely a giant step sideways, and one that drains the old NV30 fund at the same time.

Originally posted by theflyingrat
The beautiful fact about Granite Bay is its ability (like in the above tests) to run cheap PC2100. 512MB of PC2100? That's two 256MB sticks. $150 for two sticks of Crucial (very good memory by reputation; I also know this firsthand.)
Single stick of 32-bit, 512MB RDRAM? Cheapest I could track down on Pricwatch is $285 for Samsung. That's a $135 difference. Now, even if the GB boards cost $80-100 more than 850E boards, you still end up saving some cash, while basically getting the same performance, and in all likelihood, much better overclocking results.

I agree the economics do not favor RDRAM, but they don't really favor granite bay either. You can buy a good 845 board for 75-100 bucks, a 512MB Kingston Value Ram PC2700 stick for 150 (or less, 144 at googlegear last week), and have equivalent performance to granite bay (which is essentially identical to RDRAM) for even less. And the overclocking abilties of the 845 series chipsets are well known, with 160MHz + FSB and 430MHz + memory speeds obtainable very easily.

Originally posted by theflyingrat
....it's basically all over for Rambus, at least in the consumer PC market.

I agree there, RDRAM tested the limits of reason when it had a substantial performance edge, and of course the DDR competition has improved to a point where there isn't much of an edge to point to. But the important fact is that single channel DDR solutions are just as competitive with RDRAM as the dual channel ones are. Two may be bigger than 1, but PC performance is a subject that requires entirely more complex evaluation to quantify accurately.

The real death of RDRAM stems from Intel's decision to stop pursuing this particular architecture. In the P3 days RDRAM was slower, but hideously expensive. Intel forced its use on the market in an attempt to obviate the AMD question. Just like the initial adoption of AGP, the move to RDRAM was a marketing ploy designed to make the non-intel platforms look second-rate as they could not offer these features. Of course the question of whether these features were of any value had an answer so obvious that Intel gave AMD more market share than it ever dreamed as AMD was able to adopt AGP by the time it actually started to matter, and took the substantially more elegant DDR solution to the memory throughput question. Without the 300lb gorilla of Intel's will to to pressure the adoption of RDRAM, no one will be using it. The fact that Intel is moving to a dual channel DDR is not responsible, just the fact that they are moving.

nForce2 is a slightly better chipset than the single channel KT400 Via on the AMD platform, but this is not because it is dual channel. Single channel operation of nForce2 reveals that it is a more refined chipset of higher performance than is KT400, with the dual channel calling card just a sideshow. When you compare to the 845 series DDR chipsets, this lack of refinement does not exist. As such dual channel DDR alternatives have a considerably tougher time impressing me. Sure more memory bandwidth is a worthwhile goal, but improvements in single channel memory technology such as the 650MHz DDR SDRAM graphics cards already use or DDRII are just as valid of ways to pursue this goal. And the current single channel DDR setups don't particularily suffer from their situation at present afforded by system ram than runs at "only" 400-450MHz.

 

[Kunaak]

if rambus was cheaper and got more support it would be a cool idea, but the simple fact is, what I pay for 1 256 rambus I could get the best DDR 512 meg around.
in the end, it's simply not worth it.

plus RD is an aging format, with no new boards in a hell of a long time and it's simply gonna fase away unless there are a few new boards to support it.
and I just don't see any...

 

[FIZZ3]

1) NForce's implementation does *not* say anything about the P4 DC DDR implementation because of the simple fact that DDR266 and now DDR333 are an exact match for the AMD XP's FSB, whereas it is not for the P4- it lags behind considerably. DC DDR266 remedies that situation, taking only the penalty of a less efficient memory controller (compared to rambus). Thus, the increased memory bandwidth is beneficial in a general. How much particular programs benefit from this is an empirical question.

2) DDR memory is possibly more dead than rambus is. DDR is at it's absolute limit at DDR400, and the last official spec ever is DDR333. Rambus has room to grow to PC1333 and beyond. The fact that Intel has no chipsets planned may sound like all is lost for rambus, but the truth is that it's much of the same story for DDR, save this final bout of Granite Bay and it's immediate DDR333 DC successor. These chipsets will hold the water for less than a year though- then DDR II strikes, rendering all DDR memory we have now obsolete.

3) The argument that the memory bandwidth rambus used to have means nothing is flat out wrong. Rambus chipsets have consistently outperformed SC DDR chipsets in various real-world programs by a margin that approaches a cpu speed grade. What Granite Bay offers is almost the same performance, but this comes at the cost of an expensive motherboard and single-sided ram only issues, detrimenting the 'value argument'.

4) The performance delta that exists between rambus and GB may be small at this point, but it's similar to buying high quality DDR ram so one can run better timings... everyone recommends doing that, knowing the gains are about as small as those shown here. It's a matter of personal choice in the end.

Also...

Asus P4T533, w/Audio/RAID/ATA133/2-Rimm slots(Supports 32bit-RDRAM), Retail Box - $162

RDRAM 256MB 1066MHz Original Rambus 32Bit Memory Manufacturer Part# MD16R1628AF0-CN9DF - $104

Total for Asus board+ 512Mb PC1066 32ns Samsung: $162 + $208... certainly a huge amount of money compared to the quoted $150 for Crucial + the unknown but likely high priced GB board...

 

[larva]

1) NForce's implementation does *not* say anything about the P4 DC DDR implementation...

I'm not saying it does, but neither of these dual channel solutions provides a convincing performance edge in applications.

2) DDR memory is possibly more dead than rambus is. DDR is at it's absolute limit at DDR400....[/B]

Ever see see a GF4-Ti4600? The DDR on one runs at 650MHz (and up). The technical challenge certainly rises, but it is very much feasible, and in production a year ago.

3) The argument that the memory bandwidth rambus used to have means nothing is flat out wrong. Rambus chipsets have consistently outperformed SC DDR chipsets in various real-world programs by a margin that approaches a cpu speed grade...[/B]

Hmm, where is that cpu grade difference now? At one time RDRAM did posses a more substantial edge but refinement of the single channel DDR products has reduced this vastly.

4) The performance delta that exists between rambus and GB may be small at this point, but it's similar to buying high quality DDR ram so one can run better timings... everyone recommends doing that, knowing the gains are about as small as those shown here. It's a matter of personal choice in the end.[/B]

The real point is not only is the diffence between RDRAM and Granite Bay tiny, but it is virtually identical to the difference between RDRAM and single channel DDR. Nothing to get excited about in either case. I certainly choose the best quality of DDR available, as it costs hardly more than the cheapest. My 150 dollar 512MB PC2700 runs fine at PC3500, at which point the tiny edge either setup shows over single channel evaporates.

Also...

Asus P4T533, w/Audio/RAID/ATA133/2-Rimm slots(Supports 32bit-RDRAM), Retail Box - $162

RDRAM 256MB 1066MHz Original Rambus 32Bit Memory Manufacturer Part# MD16R1628AF0-CN9DF - $104

Total for Asus board+ 512Mb PC1066 32ns Samsung: $162 + $208... certainly a huge amount of money compared to the quoted $150 for Crucial + the unknown but likely high priced GB board... [/B]

Well congratulations, you've just proven a RDRAM system competes cost wise with a dual channel DDR system. Kind of the point. Both of these alternatives cost $150 more than a good 845 board and PC2700 DDR (that runs at PC3500 speeds). And both produce a performance edge over 845 that varies between insignificant and non-existant. And with one trip through the bios the 845 system eliminates those tiny disparities.

 

[Placid]

The cpu is limiting. Oc that thing
Lets see rdram=4200 dc ddr at 2100= 4200 seems like a tie.
Now how high can you oc the 4200 at x4 and how high can you oc the ddr? Expect to see asynch dc-ddr boards early next year and maybe even GB chipset asych from some of the more creative mb mfg's.

 

[FIZZ3]

Originally posted by larva
I'm not saying it does, but neither of these dual channel solutions provides a convincing performance edge in applications.

You did by mentioning this in an earlier post:

I disagree here, there is a valid lesson to be learned from nForce and nForce2 boards. The lesson is the performance of either is nearly identical between the single channel and dual channel implementations.

Which is not a valid comparison with the single channel vs dual channel chipsets on the P4 platform, as I said.

Ever see see a GF4-Ti4600? The DDR on one runs at 650MHz (and up). The technical challenge certainly rises, but it is very much feasible, and in production a year ago.

The memory used in videocards is an entirely different matter and the above statement is largely irrelevant to the question at hand. The state of the industry in this area does not bear a direct relation to the system ram field. DDR333 is the final DDR-1 specification to be approved by JEDEC and thus the last one to be properly certified. So called DDR400 memory is out there, but support and specifications (timing!) are fuzzy. There will be no chipsets supporting any kind of faster DDR-1 memory.

The real point is not only is the diffence between RDRAM and Granite Bay tiny, but it is virtually identical to the difference between RDRAM and single channel DDR. Nothing to get excited about in either case. I certainly choose the best quality of DDR available, as it costs hardly more than the cheapest. My 150 dollar 512MB PC2700 runs fine at PC3500, at which point the tiny edge either setup shows over single channel evaporates.

That is a matter of personal opinion. One can choose to go for value and stick to single channel DDR and swallow the small performance penalty. Others like to spend more and get the highest performing solution. This carries over to the next point:

Well congratulations, you've just proven a RDRAM system competes cost wise with a dual channel DDR system. Kind of the point.

Yes. My point. I did not address your previous post in particular, as you may verify now. There was at least one other member posting how GB solutions would still have the cost advantage. I have shown that it has not.
My conclusion is that you can get either Rambus or GB for about the same price, and Rambus is still on top. I understand that you would choose neither, for cost reasons and that is your good right. Do note that this opinion is just your personal liking though.

 

[larva]

Originally posted by FIZZ3

You did by mentioning this in an earlier post:

What I said was that nForce2 does not say anything about the DDR implementation on granite bay, but there is still a lesson to be learned from comparing its results in single channel mode vs dual channel mode. But you know this, as you read my post so carefully

Originally posted by FIZZ3

The memory used in videocards is an entirely different matter and the above statement is largely irrelevant to the question at hand. The state of the industry in this area does not bear a direct relation to the system ram field. DDR333 is the final DDR-1 specification to be approved by JEDEC and thus the last one to be properly certified. So called DDR400 memory is out there, but support and specifications (timing!) are fuzzy. There will be no chipsets supporting any kind of faster DDR-1 memory.

You could easily have said the same bout DDR itself during the SDR sdram days. The point is DDR speeds far in excess of 333MHz are feasible, and do already exist. Intel and JEDEC may have decided against this, but they make decisions for their concerns, not ours. And to call DDR dead because single channel DDR chipsets are on the way out is ludicrous. By the same token RDRAM is already dead then because the modern implementations are dual channel. Single channel RDRAM left a long time ago, and the dual channel implementations are going going gone in the next product cycle. But there is no reason to differentiate between single and dual channel when it comes to evaluating the usefullness of a particular module, if it's dual channel you use two. This does nothing to obsolete the memory type being used, whether it be DDR or RDRAM.

Besides, what part in your system or mine conforms to any ratified standard? This is the first thing abondoned in the search for optimal performance, so why pay homage to it now?

Originally posted by FIZZ3
That is a matter of personal opinion. One can choose to go for value and stick to single channel DDR and swallow the small performance penalty. Others like to spend more and get the highest performing solution. This carries over to the next point:

It's not a matter of opinion, look at the highest scored on 3DMark and SETI benchmarks and you will see they are occupied by good 'ole single channel DDR IT7 Max (usually) 845e boards as often as they are RDRAM rigs. Whatever performance penalty that you associate with DDR disappeared already. And since single channel systems are producing equal performance to RDRAM, and all dual channel DDR does is more closely resemble RDRAM, there is pressing little reason to use either over single channel DDR. Those that already have spent more will defend their choice to the last, but the truth is neither RDRAM nor granite bay are higher performance solutions than optimally equiped and configured 845 based platforms. I've been building PC's long enough to have paid as much as $2650 dollars for a mb, cpu, and ram, but I cannot recommend spending the money on either RDRAM or granite bay, especially to an overclocker.

Originally posted by FIZZ3
Yes. My point. I did not address your previous post in particular, as you may verify now. There was at least one other member posting how GB solutions would still have the cost advantage. I have shown that it has not.
My conclusion is that you can get either Rambus or GB for about the same price, and Rambus is still on top. I understand that you would choose neither, for cost reasons and that is your good right. Do note that this opinion is just your personal liking though.

Well your conclusion is purely opinion as well. And the top 3DMark and SETI performance held by IT7 Max's tend to back my opinion. The whole point made was neither RDRAM nor granite bay produce worthwhile improvements over 845. Continued develpement of single channel DDR as well as 845 would continue to define value and overclockablility in P4 chipsets, and sacrifice no performance in doing so. Intel has it's own agenda to follow, and it has nothing to do with our priorities. If you feel compelled to march to the beat of their drum so be it, but "note that this opinion is just your personal liking though".

 

[FIZZ3]

You could easily have said the same bout DDR itself during the SDR sdram days. The point is DDR speeds far in excess of 333MHz are feasible, and do already exist. Intel and JEDEC may have decided against this, but they make decisions for their concerns, not ours. And to call DDR dead because single channel DDR chipsets are on the way out is ludicrous. By the same token RDRAM is already dead then because the modern implementations are dual channel. Single channel RDRAM left a long time ago, and the dual channel implementations are going going gone in the next product cycle. But there is no reason to differentiate between single and dual channel when it comes to evaluating the usefullness of a particular module, if it's dual channel you use two. This does nothing to obsolete the memory type being used, whether it be DDR or RDRAM.

With DDR-I the current type of DDR ram is meant, not the amount of channels. Here is a small news story related to that & DDR-II:

http://www.siliconstrategies.com/story/OEG20020325S0065

I agree that all this may be a matter of business politics and not much else, but the fact remains that this is the reality we have to deal with.

It's not a matter of opinion, look at the highest scored on 3DMark and SETI benchmarks and you will see they are occupied by good 'ole single channel DDR IT7 Max (usually) 845e boards as often as they are RDRAM rigs. Whatever performance penalty that you associate with DDR disappeared already. And since single channel systems are producing equal performance to RDRAM, and all dual channel DDR does is more closely resemble RDRAM, there is pressing little reason to use either over single channel DDR. Those that already have spent more will defend their choice to the last, but the truth is neither RDRAM nor granite bay are higher performance solutions than optimally equiped and configured 845 based platforms. I've been building PC's long enough to have paid as much as $2650 dollars for a mb, cpu, and ram, but I cannot recommend spending the money on either RDRAM or granite bay, especially to an overclocker.

...

Well your conclusion is purely opinion as well. And the top 3DMark and SETI performance held by IT7 Max's tend to back my opinion. The whole point made was neither RDRAM nor granite bay produce worthwhile improvements over 845. Continued develpement of single channel DDR as well as 845 would continue to define value and overclockablility in P4 chipsets, and sacrifice no performance in doing so. Intel has it's own agenda to follow, and it has nothing to do with our priorities. If you feel compelled to march to the beat of their drum so be it, but "note that this opinion is just your personal liking though".

Well 3DMark is but one benchmark discipline, and one that is most related to high FSB and videocard performance, not so much ram throughput. Altogether, it's not very clear why you should take this one to test chipset differences. Maybe if you live and die by 3DMark, like some do for distributed computing efforts...
I like to use programs that I use myself. Archiving, statistical analysis software like SPSS, office apps, games etc etc. I think it's also reasonable to maintain that such an array of programs is a better indicator of performance and also more relevant than a particular synthetic bench. When I chose to buy RDRAM, the lead it had over sc DDR was there already. Only now GB comes close, and at price parity too. I've enjoyed and still am enjoying my speedy chipset. =)

 

[Hamburger]

Rambus owns all if your not overclocking, when you wanna overclock, dig a hole in the backyard and throw it in it, and go get you some ddr.

DDR owns all if your not video editing or playing around with some Cad, when you wanna do that, dig a hole in the back and throw it in it, and go get you some Rambus.

 

[larva]

Originally posted by FIZZ3

2) DDR memory is possibly more dead than rambus is.

I know exactly what DDR-I and DDR-2 are. I was referring to the statement made by yourself in a the previous post quoted above. I forgot to point out the falsehoold present in my original reply to that post, but it is such an obviously incorrect statement I felt compelled to mention it.

Originally posted by FIZZ3

Well 3DMark is but one benchmark discipline, and one that is most related to high FSB and videocard performance, not so much ram throughput. Altogether, it's not very clear why you should take this one to test chipset differences. Maybe if you live and die by 3DMark, like some do for distributed computing efforts...
I like to use programs that I use myself. Archiving, statistical analysis software like SPSS, office apps, games etc etc. I think it's also reasonable to maintain that such an array of programs is a better indicator of performance and also more relevant than a particular synthetic bench. When I chose to buy RDRAM, the lead it had over sc DDR was there already. Only now GB comes close, and at price parity too. I've enjoyed and still am enjoying my speedy chipset. =)

I agree that 3DMark is far from the end-all be-all of machine performance, but it no RDRAM supporter worries about this when RDRAM does demonstrate an advantage here. I agree it is infuence heavily by video card and FSB, but the top RDRAM rigs don't suffer in these areas compared to the top DDR rigs. Granted you may have to mod it till you are blue in the face, but RDRAM rigs have proven capable of 200MHz + FSB in determined hands. It certainly is easier to get extreme FSB out of an IT-7, but P4T533's are doing it too, and still not demonstrating a clear edge.

SETI is one of the best torture tests for the performance of a memory subsystem. It is obvious from the top 10 performers that DDR is no longer a poor second-cousing to RDRAM. Also, in the granite bay test linked by the thread starter, an 845 rig beat both RDRAM and granite bay in Q3, very much real world app. In the other real world apps the difference was extremely close as well, and this is at the stock clock rates that give RDRAM and granite bay the best chance to demonstrate a convining superiority, should it exist.

Perhaps at the time you bought it RDRAM really had an edge, but it has dissapeared due to stagnation on the RDRAM front, and continual evolution of the DDR alternatives. At this point granite bay does provide equilivalent performance, but I agree the price is too high. In six months or so granite bay will not be priced as unfailry as it is now, and will make more sense. But as a good 845 rig can compete with it (at stock clock) and even surpass it (in overclocked trim) I certainly won't be paying a premium for it just like I wouldn't invest in RDRAM at the present. Fortunately granite bay is not the worst thing to happen, but it will be some time before it attains the value 845 provides now. The best we can hope for is that it proves as overlockable as 845 has, or it will won't even be returning what we already had when the price is more reasonable.

 

[FIZZ3]

Rambus owns all if your not overclocking, when you wanna overclock, dig a hole in the backyard and throw it in it, and go get you some ddr.

DDR owns all if your not video editing or playing around with some Cad, when you wanna do that, dig a hole in the back and throw it in it, and go get you some Rambus.

Rambus is just as good if you consider the bandwidth you get at 155Mhz 4x already. No need to dig any holes.

A while back a tip on the front page showed how Rambus at 3x can be just as effective as ddr on high fsb too. Throw the shovel on the ground and keep your hands where we can see em!

 

[FIZZ3]

I know exactly what DDR-I and DDR-2 are. I was referring to the statement made by yourself in a the previous post quoted above. I forgot to point out the falsehoold present in my original reply to that post, but it is such an obviously incorrect statement I felt compelled to mention it.

That quote you gave is not a falsehood in any way. I explained how DDR has only about a year to go until DDR-II renders current DDR-I single AND dual channel chipsets obsolete. Until that time, the performance you get with Rambus is comparable and sufficient. The lifespan of both DDR-I and Rambus ends at the end of that year.

Perhaps at the time you bought it RDRAM really had an edge, but it has dissapeared due to stagnation on the RDRAM front, and continual evolution of the DDR alternatives. At this point granite bay does provide equilivalent performance, but I agree the price is too high. In six months or so granite bay will not be priced as unfailry as it is now, and will make more sense. But as a good 845 rig can compete with it (at stock clock) and even surpass it (in overclocked trim) I certainly won't be paying a premium for it just like I wouldn't invest in RDRAM at the present. Fortunately granite bay is not the worst thing to happen, but it will be some time before it attains the value 845 provides now. The best we can hope for is that it proves as overlockable as 845 has, or it will won't even be returning what we already had when the price is more reasonable.

From your performance comparisons I get the feeling you are comparing overclocked SC DDR chipsets to others by means of 'hotlists'. While I don't object to doing so, I do think that Rambus overclocking is underestimated and that not many people are doing it accordingly. The few that do, score the highest scores. So looking at Top100 scores means that you're ignoring the distribution of systems... 99% of users have DDR rigs and thus the percentage that reaches higher numbers by extreme overclocking floods the ranks.

Let's compare some bandwidth numbers to show what the difference really is.

DDR - 333: 166.000.000 cycles * 64 bits * 2 double data rate = 2656Mbyte per second.

DC DDR - 266: 133.000.000 cycles * 64 bits * 2 DDR * 2 channels = 4256Mbyte per second.

Rambus - PC1066: 133.000.000 * 8 mp * 16 bits * 2 channels = 4256Mbyte per second.

Let's overclock to get to the heart of this discussion. I am taking hypothesized FSB differences into account:

DDR system @ 166Mhz 3:4 = 221M cycles * 64b * 2DDR = 3536Mb/s.

DDR system @ 183Mhz 3:4 = 244M * 64b * 2DDR = 3904Mb/s.

Rambus @ 150Mhz 4x = 150M * 8 * 16b * 2ch = 4800Mb/s.

Rambus @ 157Mhz 4x = 156M * 8 * 16b * 2ch = 5024Mb/s.

Conclusions:

1) DDR can't even match PC1066 stock speed at 183Mhz 3:4 which makes the ram run at DDR 488 already.

2) Rambus at top speed of 157Mhz 4x reaches 5Gb of bandwidth in sync with the FSB. This is simply superior performance.

These numbers are more relevant than biased/schewed TopX ranks in my view. I do agree with you that getting this extra bandwidth may be decided by value considerations related to what one does with the pc.

 

[larva]

Those numbers are a typical spec sheet suedo-knowledge. There are many factors other than raw bandwidth that effect the realized performance a memory subsystem achieves. Looking at the raw bandwidth numbers and declaring RDRAM superior ignores the reality of the situation. Even the latest PC1066 is composed of 32ns devices. PC2700 DDR is 6ns typically. This indicative of the enormous latency advantage sdram enjoys that until rather recently made it outperform RDRAM. And at persent it allows DDR systems to provide equivalent performane despite possessing a (very small at this point) disadvantage in raw bandwidth. Since I already covered this point in a post I made yesterday I shall quote it now.

"Yes, latency is RDRAM's dirty little secret. The latency of SDRAM, whether it be SDR or DDR is vastly better than RDRAM. This is why DDR systems with lower sandra scores can indeed compete with RDRAM systems posting higher bandwidth numbers. Once you get the DDR system to the same bandwidth level, it is a faster system due to the latency advantage. Back in the single channel RDRAM days PC800 got it's lunch eaten by PC133 SDRAM. Intel had to go to the dual channel RDRAM implementations to give RDRAM systems a large bandwidth edge over SDRAM systems before it finally showed any promise, much less an advantage. As the raw bandwidth of DDR sdram has improved to near parity with RDRAM the advantage that RDRAM posses varies between insignificant and non-existant."

Yes it is ineed hard to get single channel DDR systems to equal the raw bandwidth exhibited by RDRAM systems. But the difference is small (although some are indeed running 187 at 3:4) it never has to equal RDRAM's bandwidth to make a system of equal performance on the whole. Different tasks respond to different characteristics, and those that need raw bandwidth do still prefer the RDRAM characteristics. Most however, including the vast majority of home user PC tasks respond much more favorably to the latency characteristics of DDR memory subsystems.

And as far as RDRAM outliving DDR, dream on. Intel has already anounced that 850e will be the last RDRAM platform. RDRAM accounts for a tiny percentage of the ram produced at present and will dwindle to non-existant very shortly without Intel forcing its adoption on the PC market. As you know DDR-1 chipsets are still being designed and introduced, whether they be single or dual channel. And the huge installed base of DDR machines means the production will continue long after the appearance of DDR-2, just like PC133 is still produced today.

 

[runsalone]

Why would anyone buy an RDRAM board with the dual channel DDR boards coming out??

With some good corsair DDR modules and a dual channel DDR board, you'd be able to get probably 4000 mb/s or more in sandra memory benchmark with a 3:4 divider and a bump in FSB speed.

 

[Placid]

 

[FIZZ3]

Those numbers are a typical spec sheet suedo-knowledge.

System efficiency is similar and only a small factor. The numbers I gave do hold relevance.

Even the latest PC1066 is composed of 32ns devices. PC2700 DDR is 6ns typically. This indicative of the enormous latency advantage sdram enjoys that until rather recently made it outperform RDRAM. And at persent it allows DDR systems to provide equivalent performane despite possessing a (very small at this point) disadvantage in raw bandwidth. Since I already covered this point in a post I made yesterday I shall quote it now.

The above quote truly is pseudo-knowledge. The 6ns vs 32ns numbers have no indicative relation to real latency. Judging from this text of yours, you apparently think Rambus latency is 4 times that of DDR...

Hardly:

http://www.tech-report.com/reviews/2002q4/i845pe-ge/index.x?pg=5

A 23% difference is more like it. Compare this to the DDR333 2.7Gb vs 4.2Gb RDRAM bandwidth (a 56% difference). That is a favorable trade-off by all accounts.
*Dual channel is not taken into account here because it's latency and efficiency haven't been properly measured yet.

And as far as RDRAM outliving DDR, dream on. Intel has already anounced that 850e will be the last RDRAM platform. RDRAM accounts for a tiny percentage of the ram produced at present and will dwindle to non-existant very shortly without Intel forcing its adoption on the PC market. As you know DDR-1 chipsets are still being designed and introduced, whether they be single or dual channel. And the huge installed base of DDR machines means the production will continue long after the appearance of DDR-2, just like PC133 is still produced today.

I did not say outlive. I said both would become obsolete in about a year. Also, claiming that an installed base will somehow postpone obsolesence is irrelevant to the current discussion we have on the memory's performance.

 

[OC-Master]

Ram bus is dead.

This time next year, it wont even exist and everyone will be buying DDR and DDRII sticks in dual channel configurations.


OC-Master