The beginning of the end for DDR memory?

[KenAF]

The main advantages of DDR over RDRAM thus far have been: 1) price; 2) latency.

With the substantial increase in DDR memory chip cost, there is no longer a massive difference between DDR and RDRAM pricing.

There is no doubt that PC2100 (266Mhz effective) DDR offers lower latency than PC800 (400MHz) RDRAM. Latency has a significant impact on the performance of office and general productivity applications. However, unlike DDR, where latency either increases or stays the same as memory speed increases, the latency of RDRAM decreases with increases in frequency. PC2100 (266Mhz effective) DDR memory has 200 cycle latency with the latest chipsets (Nforce, KT266A, etc), whereas PC800 RDRAM has a 270-300 cycle latency. Obviously, the lower the latency (latency = delay), the better the performance. Put another way, PC2100 DDR kicks the butt of PC800 RDRAM when it comes to latency.

With the move from PC800 (400MHz) RDRAM to PC1066 (533MHz) RDRAM, latency drops from 270-300 cycles to 200-210 cycles, or essentially the same latency as PC2100 (266Mhz effective) DDR memory. Moreover, the PC1200 (600MHz) RDRAM demonstrated this past week offers latency of 160-170 cycles, or 20% lower (=faster) than PC2100 DDR memory. In contrast, most PC2700 (333Mhz effective) DDR memory chips increase in latency from the 200 cycle time of PC2100 to 220-230 cycles; that puts PC2700 (333Mhz effective) DDR with a latency that is 30+% faster than current PC800 RDRAM, 10% slower than upcoming PC1066 RDRAM, and 30+% slower than PC1200 RDRAM.

As Tom's latest tests demonstrate (and he never tested higher than 472MHz RDRAM on the P4-2600, compared to the 533MHz RDRAM Intel will be using after March 28), the P4 performs better at 2600MHz with 472MHz RDRAM than it does at 3000MHz with 137x2=274MHz DDR memory. One can only assume that the performance delta will widen as Intel launches 533MHz RDRAM (again, Tom tested at a max of 472) on March 28. Moreover, many enthusiasts will undoubtedly try to overclock their PC1066 (533Mhz) RDRAM to PC1200 (600MHz), for a significant improvement in both latency and bandwidth.

Comments?

 

[AndyHui]

A lucid and balanced account....very good. I agree with most of your points.

I put forward that there are 2 points that you have missed in regards to RDRAM....latency increases with the number of devices present on the bus, but chipsets like the i850 use aggressive pre-fetching and buffering to hide as much latency as possible.

DDR itself only needs a small modification according to Kentron for QBM...Given their existing designs and FSBs, the Pentium 4 would take more advantage of QBM than the Athlon. Since QBM is yet to come out, this is all speculation.

 

[Texmaster]

I have had 3 rambus boards and I have defended it well.



<< As Tom's latest tests demonstrate (and he never tested higher than 472MHz RDRAM on the P4-2600, compared to the 533MHz RDRAM Intel will be using after March 28), the P4 performs better at 2600MHz with 472MHz RDRAM than it does at 3000MHz with 137x2=274MHz DDR memory. >>



FOR THE LAST TIME.

Tom's hardware did NOT overclock the DDR Memory as the Rambus was in his test.

In actuality, the scores are VERY close once you overclock DDR memory.

 

[KenAF]

<< FOR THE LAST TIME.

Tom's hardware did NOT overclock the DDR Memory as the Rambus was in his test. >>

Huh? Yes, he absolutely did. He made this clear in the article right here, and reiterated this fact in his posts to the community board on his web site. He clocked the P4 2.2 at 3.0GHz using the GIGABYTE GA-8IRXP motherboard configurated at 22x137MHz with 256Mb of Micron DDR333. The default DDR memory speed on P4 boards is 100MHz (200MHz effective), so he overclocked it to 137MHz to jump from 2.2GHz to 3.0GHz.

If you actually go to the page I link above, you will see his statement, "Our RDRAM system (Intel 850 chipset) was able to run at a maximum clock speed of 2600 MHz." He goes on to say, "Here, we can deduce that a DDR SDRAM module clocked at 137 MHz cannot deliver the necessary bandwidth in combination with the Pentium 4, thereby accounting for performance that is somewhat lower than expected. By comparison, our Pentium 4 system with the Intel 850 chipset shows better results in a few of the benchmarks, despite a lower clock rate of 2600 (Pentium 4/2600)." As Tom was not able to get his RDRAM platform up to the necessary FSB for 3.0GHz, he used the Gigabyte DDR motherboard instead. Compared to the actual PC1066 RDRAM that Intel intends to use, the RDRAM in the 2400MHz and 2600MHz tests was significantly underclocked.

 

[Texmaster]

<<

<< FOR THE LAST TIME.

Tom's hardware did NOT overclock the DDR Memory as the Rambus was in his test. >>

Huh? Yes, he absolutely did. He made this clear in the article right here, if you actually read it, and reiterated this fact in his posts to the community board on his web site. He clocked the P4 2.2 at 3.0GHz using the GIGABYTE GA-8IRXP motherboard configurated at 22x137MHz with 256Mb of Micron DDR333. The default DDR memory speed on P4 boards is 100MHz (200MHz effective), so he overclocked it to 137MHz to jump from 2.2GHz to 3.0GHz. >>



Obviosuly its you who do not read enough.

Many i845 boards have a 3:4 raitio and turbo features to overclock the ram bringing the ram overclock to 150+fsb.

Tom conviently used one of the few i845 boards without this feature.

But since you obviosuly need an example I will give you one. Sigh.

My best score on 3dmark 2001 is 9533 t500 265/590 i845 P4@2700

Amused one's best score on 3dmark2001 is 9498 t500 265/602 i850 P4@2400

And bofore you get any ideas it was amused who showed me how to overclock.

And if you knew anything about 3dmark you would know it likes two things. Bandwith and Video Card speed.


Amused and I both have t500s, his with a slightly greater overclock. CPU speed is not nearly as big of a factor.


So you can see now that you were wrong in your assessment because you took Tom's word. Try not to make that mistake again.

 

[christoph83]

All I can say KenAF is with the right board you can get good bandwith with DDR. The Asus P4B266 has an 3:4 function which allows the cpu to run at 100FSB and memory at 133FSB. Right now im running 20x123. So really my memory is running at 156FSB. I get 2400/2400 scores in sandra 2002 which is equal to PC800 RDRAM. If tom used this board with he could have given the P4 much more bandwith than the 2000/2000 he benched it at. This feature really makes this board a step up from all the other i845D boards in my opinion.

 

[flexy]

<< All I can say KenAF is with the right board you can get good bandwith with DDR. The Asus P4B266 has an 3:4 function which allows the cpu to run at 100FSB and memory at 133FSB. Right now im running 20x123. So really my memory is running at 156FSB. I get 2400/2400 scores in sandra 2002 which is equal to PC800 RDRAM.. >>



according to sandra 2k2 i also only get slightly lower scores (like 2250,2100) using a very moderate overclock 145 FSB (via kt266a), and i use only DDR2100 Ram specced at Cas2.5 (run at CAS2 of course).....i think this is NOT bad for DDR2100 ram and it does NOT make me feel like i have to go out and get a rdram board...i'd rather get DDR2400/2700 and experiement with that then... and i am sure i can reach/beat rdram800 speed...no big deal for me...

 

[Pabster]

RAMBUSt memory bandwidth means diddly squat. The fact is, today's speediest official DDR modules (PC2700) not only provide significantly lower latency, but greater bandwidth. Either we've got some big RAMBUSt fans here, or too many people not taking advantage of DDR.

I still maintain that RAMBUSt is a dead-end technology. PC1066/PC1200 be damned. DDR is moving forward, despite any claims to the contrary. QBM is just around the corner, and DDR-II is shaping up nicely. Why stick an investment in to a proprietary technology like RAMBUSt? Not to mention the company itself.

 

[Athlon4all]

I agree. DDR is not going to get any faster IMO. QBM dos have potential, but right now, I am not too comfident in it because VIA recently, they had planned to include QBM in P4X333, well recently, they took QBM out of P4X333 and put in it P4X600, I am not too comforted by this move, and I'm not confident that it will ever see light of day. And plus, sure DDR and RDRAM are pretty close right now but when we see 533fsb+pc1066, I feel that there will be a big enough increase for RDRAM to be worth it, and with current prices, why not? We will find out soon about how 533fsb+845-D compares to 533fsb+PC1066 because I had a discussion with Johan from Ace's hardware and he says that he is writing an article that will compare 533fsb on 845-D, SiS 645, P4X266, and 850 with 850 using PC1066 and all the DDR platforms (except for SiS 645) using PC2100. But though, the one thing I will note is that I don't expect DDR to die however. I doubt that AMD CPU Memory Bandwidth requirements are going to increase any time soon (that includes Hammer) and DDR will stay for the Athlon, but for the P4, I think that RDRAM is in for a comeback this year. We'll see, I agree though KenAF with most of what you said.

<< Tom's hardware did NOT overclock the DDR Memory as the Rambus was in his test. >>

Texmaster, I am not going to dispute you on this, but I am not certain what your point is. The reality is, that a P4 3.0GHz 533fsb CPU on 845-D will not run at PC2700 speeds but at PC2100, BUT 850e (533fsb) will run at PC1066 speeds and Tom wasn't even running at PC1066 but more inbetween PC800 and PC1066. I will not argue that you are correct, but I fail to see how that changes things because if anything, that is more resembant of the picture that we will see at the end of the year.

 

[christoph83]

Texmaster, I am not going to dispute you on this, but I am not certain what your point is. The reality is, that a P4 3.0GHz 533fsb CPU on 845-D will not run at PC2700 speeds but at PC2100, BUT 850e (533fsb) will run at PC1066 speeds and Tom wasn't even running at PC1066 but more inbetween PC800 and PC1066. I will not argue that you are correct, but I fail to see how that changes things because if anything, that is more resembant of the picture that we will see at the end of the year.

The only thing we are saying is that the asus p4b266 board has an option that theoretically you can reach pc2700 speeds. I'm running 156FSB memory and 123FSB CPU on this board. This equates to 2400/2400 in sandra 2002. Toms 3ghz setup had 2050/2050 memory benchmarks. We all know that more memory bandwith is better for the P4, and that extra bandwith from this option would have really helped out. Technically the memory was not really overclocked in toms benchmark because it was running around 133FSB= PC2100.

I dont know how you can think RDRAM is dead end technology after the toms review. The P4 Needs RDRAM or a DDR board like the Asus P4B266 than can supply extra bandwith. How is PC1066 and PC1200 not moving forward? Theres nothing wrong with the company Rambus. They had patents in DRAM related areas that they thought were being used by other companies and not getting payed for it. So tell me....if you were rambus...you invented some DRAM related patents,then other companies stole it and never payed you,would you fight to get some money out of it? If calling Rambus names makes you feel better about owning DDR,thats fine.

Edit: I'll post this here as well as I wanted to see how much of a different the extra overclock on the P4B266 makes. I just did the first two quake benches from toms review. I put the clock at 2.4ghz and the DDR is running at 153FSB. Tom's geforce 3 was running 250/400 and I was only able to put my ti200 at 220/400. The core wouldnt go any higher. So I have a slightly slower video card.

The 640x480 quake3 Test I got 295FPS . 10 slower than the 2.4ghz equipped with overclocked RDRAM, 13 slower than the 3ghz with slower DDR and 15 faster than the 2.2ghz with RDRAM.

The 1024x768 test I got 214FPS. Only 3 FPS slower than the 3ghz with DDR. 5 slower with the 2.4ghz with RDRAM and 2 faster than the 2.2ghz with RDRAM. And whats suprising is this was done with a slower clocked video card. From these tests with the 3:4 option on the Asus board,RDRAM and DDR is almost identical and possibly faster in the 1024x768 test.

Did another test with the same setup. 3dmark2000 with my video card at slower speeds I got an 10848. Very strange score as it beat all of the setups on tom's review. All I can say is this. If you go DDR get an Asus P4B266! Maybe I should start another thread with these scores.

 

[JellyBaby]

<< I still maintain that RAMBUSt is a dead-end technology. PC1066/PC1200 be damned. >>

Good sentiment. I'll overclock to that.

 

[Athlon4all]

<< The only thing we are saying is that the asus p4b266 board has an option that theoretically you can reach pc2700 speeds. I'm running 156FSB memory and 123FSB CPU on this board. This equates to 2400/2400 in sandra 2002. Toms 3ghz setup had 2050/2050 memory benchmarks. We all know that more memory bandwith is better for the P4, and that extra bandwith from this option would have really helped out. Technically the memory was not really overclocked in toms benchmark because it was running around 133FSB= PC2100. >>

Ok. I see what you're saying. But how does that pertain to this discussion? The bottom line is that it is very unlikely that 533fsb will yield any performance boost on 845-D/P4X266A/SiS 645, while Ace's Hardware has proven that there is a good 5-10% increase by going from 400fsb+PC800 to 533fsb+PC1066.

<< Yeah that's true! When Intel released the i845D I'm all DDR... Rambus is suck! It's overexpensive. Some people like Rambus - oh why? >>

What do you mean overexpensive? From Crucial and Newegg:

256MB Crucial PC2100 DDR $75
2 sticks of Kingston 128MB PC800 RDRAM 2x$43=$86 a monstrous $11 more!

Need I say more?

 

[christoph83]

Ok. I see what you're saying. But how does that pertain to this discussion? The bottom line is that it is very unlikely that 533fsb will yield any performance boost on 845-D/P4X266A/SiS 645, while Ace's Hardware has proven that there is a good 5-10% increase by going from 400fsb+PC800 to 533fsb+PC1066.

You are correct that as of now DDR will not be helped much on the 533FSB. xbitlabs has a review on DDR at 533FSB and it gives about a 2-5% boost. Although once Dual channel DDR chipsets come out this should help some in that 400FSB to 533FSB jump. They will provide 4.2gb/s bandwith vs 2.1gb/s right now. That will be very interesting to see since PC1066 RDRAM will provide around that same amount of bandwith.

 

[LucJoe]

You are comparing a higher quality DDR ram to a 2 sticks of Kingston Rambus ram. That seems like a bad price comparison to me.

Best way I can think of to compare bottom-line prices is pricewatch. From Pricewatch:

256mb pc2100 DDR: $50
256mb RDRAM: $70

That makes rambus ram 40% more expensive than DDR

 

[AGodspeed]

DDR-II and/or QBM will arrive soon enough. DDR will live on for years to come, no doubt about it to me.

Rambus will be niche market product for a long time until they come out with something besides their 4i RIMMS and Yellowstone hopes.

 

[splice]

Heh, as long as JEDEC is around, DDR will be too.

 

[oldfart]

256 Meg Samsung PC800 $76

Price is no longer a factor.

 

[dullard]

I'll make one point. Yes the PC1200 revealed this weekend has a lower latency than current RDRAM. And yes DDR has a lower latency than DDR. However as Crucial says, lower latency is most likely unnoticible to the end user. Suppose you request a chunk of memory be sent to the processor. The first data that reaches the processor is delayed by the latency a few nanoseconds. From then on, the rest of the data comes without a latency penalty. Even with 0 latency memory, that same data chunk would only be a few nanoseconds faster. No one can tell the diffence in speed between a process that takes 1 second and a process that takes 1 second + a few nanoseconds.

Latency has an effect when want to overclock - usually lower latency memory overclocks better. But now you are talking a very small fraction of the market.

RDRAM is here to stay and so is DDR. I think PC1200 RDRAM will be great (especially since you won't have to buy in pairs anymore and the price isn't supposed to be any more than PC800).

 

[Athlon4all]

<< You are comparing a higher quality DDR ram to a 2 sticks of Kingston Rambus ram. That seems like a bad price comparison to me. >>

Well I disagree, Kingston is high quality. Sure they aren't the best overclockers but they are definately in the same league as Crucial.

<< DDR-II and/or QBM will arrive soon enough. DDR will live on for years to come, no doubt about it to me. >>

I agree in terms of Hammer/Athlon XP. DDR is here to stay but for the P4 I feel that RDRAM is here to stay as well. QBM, well I think we've discussed this previously AGod and you know where I stand

<< RDRAM is here to stay and so is DDR. >>

Very true, couldn't agree more.

<< I think PC1200 RDRAM will be great (especially since you won't have to buy in pairs anymore and the price isn't supposed to be any more than PC800). >>

Where does it say that PC1200 won't need to be installed in pairs? The reason current RDRAM requires to be installed in pairs is because of 850 is a Dual Channel RDRAM chipset, and actually, originally PC1066 was supposed to be the flaship RAM on Tulloch (Single Channel 850 but gets same bandwidth with 32-bit RIMM's) and that wouldn't need it to be installed in pairs.

 

[AGodspeed]

DDR is here to stay but for the P4 I feel that RDRAM is here to stay as well.

I agree Athlon4all. However, it seems to me that RDRAM will be a niche product (in P4 land) until they develop a low-cost high-performance product (like DDR is).

Rambus is/will be niche kinda like Transmeta's ultra low power processors or Matrox's Razor sharp 2D Gxxx video cards.

 

[Athlon4all]

<< However, it seems to me that RDRAM will be a niche product (in P4 land) until they develop a low-cost high-performance product (like DDR is). >>

Well, it all comes down to weather RDRAM prices stay where they are. If they do, then I wouldn't call it a niche product, but i guess, I also feel as I have stated that with 533fsb+PC1066 I feel that the performance difference will be enough for it to be worth taking over DDR.

 

[Agent004]

KenAF, that's a nice account.

Rambus is quite a suitable solution for the further processor. The history repeats itself, first we started with a serial path memory, then a parrallel path memory. Rambus (if it can survive that long) is converting also from serial to parrallel.

 

[Remnant2]

Dullard, that's untrue. In actuality, memory latency can have quite a large effect on memory performance.. it's not all about bandwidth. The prime example of this?

Pointer chasing. This is a very common occurance, as you traverse a tree or skip down a linked list. At each node, you have to retreieve the memory contents of the next node before the processor can continue. Unless you're clever or lucky, those nodes will not be in the cache, causing a main-memory access. The latency of the memory directly influences how long it takes before processing can continue.

So lets say you save 60 cycles as seen by the CPU from one memory type to the other. Big deal, 60 cycles is nothing. But if your tree has 10,000 elements, you're looking at 600,000 cycles of idled CPU time. Much worse but not too bad. Now say that this tree is the scene graph of your 3D game scene -- you're drawing it 60 times a second. Suddenly you're wasting 36,000,000 cycles of CPU time doing nothing but waiting for the nodes of the scene graph to be brought in from main memory.

On an unrelated note, is there an article or reference somewhere that states that the latency of DDR does NOT decrease with frequency? I may be mistaken, but most usually the memory latency will decrease with frequency, as a natural side effect of increasing the frequency. IE if you're driving memory at 150mhz vs 100mhz, the first word of data is retrieved in a shorter period of time than 100mhz memory -> lower latency.

 

[AluminumStudios]

Aside from all of the specs, benchmarks, and technical arguments there is another side to the decision to go with DDR vs RD. Flexibility. OEMs can equip budget systems with a single DDR DIMM whereas RD has to be used in pairs. Also, most RD systems only have 2 RIMM sockets so upgrades require removing current memory and buying larger amounts for the upgrade, many DDR systems have 3. With DDR you can slap in another DIMM. This is a factor for OEMs and for non-leading edge performance oriented customers.

It's sad to say that non-leading edge performance oreinted customers (laymen) have significant influence because their numbers are great!

 

[BFG10K]

All you people proclaiming the death of DDR ask yourselves one question: is there RDRAM available that provides 8.8 GB/sec memory bandwith (10.5 GB/sec expected on the GF4 Ti4600) that DDR RAM is able to provide? Also RDRAM is totally unsuitable for video cards because it's too hot and uses too much power.

Also its latency increases as you add more RIMMs and you pretty much have to use them in pairs.