<<
<< Oh blah blah... my RDram setup at 150Mhz FSB get near the 4000 mark on Sisoft. Does that matter? Sisoft is a synthetic benchmark that has no definitive correlation with real world results, it is based off optimal results. Sisoft also says my P4's FPU is stronger than an Athlon XP1800 or a McKinley 1Ghz, is that true? Heck no. >>
Agreed, same can be said about 3DMark 2001. Just look at R8500 vs. GF3/GF4!!
However, the real world results as I have stated more times than I can count, are there for PC1066 vs. DDR400 and DDR333.
<< Well said. >>
Agreed
<< BTW, Has DDR come to an end - interms of speed and scalability?. I know RDRAM is not. >>
Agreed. DDR400 will prolly never hit the market.
<< I just noticed the title and the conclusion to Tom's article... and dang... thats really funny. If I wrote an English paper like that, I'd fail the class. Looking at the prices of memory + motherboards, it looks like if you buy 512MB of ram (high quality RDram and high quality PC2700 DDR SDram) with a motherboard (Asus P4T-E and Asus P4b-266C for comparison's sake) they end up at roughly the same cost. This is pertenant to the United States, it may be different in Europe, but then again Europe has really high protective tariffs anyways. >>
Dang. That's what I keep on trying to tell everyone, now I know someone will opint out that there is no PC1066, but high quality RDRAM (which I consider Samsung) with the right board easily hits PC1066 speeds as long as you've got either a 533fsb CPU or one that can be overclocked to it.
<< I'm a EE student at U of Washington Seattle, most of the time, serial is *inferior* to parallel for practical purposes. I have no idea why Rambus on serial is more superior than SDram on parallel, but im not a tech wiz. I'd have to say a lot of design is revolved around parallel, not serial (easy example is to take 2 60W light bulbs and put them on serial with a fixed voltage and then the same 2 bulbs and put them in parallel). >>
I'm no expert either alright, but I disagree. Maybe Fkloster can confirm this, but to me, its like the P4 and it's long pipeline compared to the P3 and its shorter pipeline.
The P4 on a Long Pipeline at first seems to perform terribly alright, but once the benefits of the longer pipeline vs. the P3 are put to its advantage, you end up with a faster product. Just take a look at the final 2GHz .18 micron Williamette, and compare its performance to the last .18 Coppermine P3 (which I believe was 1.1GHz on the final stepping). The Williamettte kicks butt!!!! I think pararell vs Serial is very similar to that. Now, there is one other key factor however added to the equation besides Serial getting more MHz, the other factor is because of its lower pin count (16 bit width), it is fairly easy and fairly cost effective to implement a Dual Channel RDRAM Memory COntroller, where with DDR, I don't care what the real prices are on nForce chips and mobo's, nVidia and the mobo makers have got to be just swallowing up profit losses to sell nForce 420-D and 415-D boards where they are. So, anyway. That's all I have to say. Does this make any sense?
edit:
<< I guess the only way to get individuals like yourself to consider any testimony is to provide links to other sites confirming what I'm saying. There is nothing about any of the review sites that makes their reviews more credible than any independent benchmarks you can peform yourself. They offer break downs of how they went about performing the benchmark, which are time consuming to prepare, and there's no way I could take the time to do that. >>
Alright then, you've made your position clear ok, how about you tell if not Fkloster, then me what your benchmarks have said and how they have been run? I am curious to see where your conclusions have come from >>
Athlon4All, the reason why serial architectures do better is because you're only dealing with one chip at a time. Essentially with RDRAM only one chip is transfering information down the pathway at any one time. Thus, since it has no other "Brothers" to slow it down (other chips, like in DDR where information is drawn in paralell) data access can occur at obscene speeds. PC800 runs at 1600MB/s
Couple it with a narrow fast bus, (Which means less traces running around the motherboard, and less problems with EMI causing a clockspeed barrier) and you ultimatley have a more scalable architecture. The problems with latency are related with waking up the chip and waiting for data transmission to start. The reason why RDRAM exhibits less latency as frequency increases is because, like I said, it doesn't have any brothers to slow it down and thus doesn't have to wait on all the other chips to start data transmission like DDR does. Thus, with RDRAM the latency problem is with the individual. With DDR, it's with the collective. Where DDR has to do a bank refresh and column access strobe stuff with all of the chips, RDRAM just engages one chip and starts the data flying off it's merry way.
It's a bit like this.
DDR is like controlling a squad of weaker soldiers that can work together
RDRAM is like controlling a single very powerful and fast soldier
With DDR, the thing limiting the soldiers performance is how well you can coordinate them to work together
With RDRAM the limit is with the soldier itself.
Another advantage of RDRAM is it's relativley lower datapath. 16 bits on RDRAM versus 64 on DDR means it has more headroom. As quickly as you solve the EMI problems, you can march up the datapath. Thus, at the same width RDRAM will utterly smash DDR into oblivion. And since as I said with RDRAM's serial architecture the limit is with the chip itself, it's easier to ramp up the speeds since you have no weakest link in the chain, since the chain is one link.
I know i'm sounding really uncoherrant, just trying to provide a sliver of useful information.
Hope that helped, Athlon4All. Too bad the Athlon can't benefit from tremondous memory bandwidth RDRAM provides. Perhaps one day when RDRAM's latency is lower than DDR II's, the hammer could work with RDRAM. What a day that will be...