P4 with 4.266 GB/s bandwidth?

[Raystonn]

"This is not actually true in normal operation."

Ah but it is. The latency killer in SDRAM is the dead wait cycles during changing from reading to writing or vice versa. This happens all the time during standard system operation. This even outweighs RDRAM's penalty on totally random accesses.

-Raystonn

= The views stated herein are my personal views, and not necessarily the views of my employer. =

 

[Sojourn]

How do the laws of physics change when you go from 2 channels to 4 channels? The physical implementation of 8 DIMM slots will increase the latency of the entire RDRAM bus compared to 4 DIMM slot configurations. That's just the way RDRAM works, as you've pointed out many times yourself. Its not the increase in channels that increases latency, its the increase it trace length between the bus and modules. I don't see this as being practicle for PC system memory, though I believe quad channel RDRAM implementations already exist in other applications.

-= This is our wading pool.
Stop pissing in it. =-

 

[Raystonn]

"The physical implementation of 8 DIMM slots will increase the latency of the entire RDRAM bus compared to 4 DIMM slot configurations"

We would be moving from 2 channels, each having 2 RIMM slots, to 4 channels, each still only having 2 RIMM slots. There would be no increase in latency per channel.

-Raystonn

= The views stated herein are my personal views, and not necessarily the views of my employer. =

 

[FUGGER]

Side note reguarding the PC800 running over 100Mhz FSB

There is a option to go from quad pumped 4x to tripple pumped 3x on the P4T motherboard for overclocking into the 130Mhz + range.

This will reduce bandwidth on the bus, but allows for overclocking into the 2Ghz area without overclocking your ram.

 

[leonov]

You said

"The latency killer in SDRAM is the dead wait cycles during changing from reading to writing or vice versa. This happens all the time during standard system operation. This even outweighs RDRAM's penalty on totally random accesses."

This article explains (very well) why you are incorrect. It is talking about RDRAM as compared to DDR SDRAM.

(from http://www.ntsi.com/DDRRam_Explained.htm)

"In other words, everything being optimized, the main difference in the two hardware configurations can be considered the memory subsystem itself. This, of course, raises the question why the solution blessed with a substantially higher peak bandwidth shows the lower performance in real world applications. To understand this, one must, however, take into account that, in real life situations, a high percentage of memory accesses occurs randomly. That is the data are not necessarily stored in memory in a sequential order but distributed to wherever there is enough space within the memory array, hence the term Random Access Memory.

Random accesses, however, do not simply draw from an already open page. As the name indicates, both row and column addresses need to be specified, decoded and accessed, introducing additional latencies. Exactly this fact is the Achilles heel of Rambus memory, since the initial latencies are substantially higher than in standard DRAM. Consequently, even if the maximum bandwidth achievable (peak bandwidth) by Rambus PC800 memory is 50% higher than the peak bandwidth of PC133 SDRAM, the average bandwidth under real life conditions is substantially less since random access-related latency becomes the main limiting factor. There has been enough coverage of Rambus DRAM vs. SDRAM, thus, there is no need to further go into details. The point, however, has been made that latency is a crucial component of memory performance."

It goes on to show how very significant improvements can be made to increase the performance at NO extra cost.

I know that HardwareCentral have put out an article about how RDRAM has lower latencies etc but let's face it, it CANNOT be true. If it were RDRAM, with its alleged better bandwidth AND latency would easily beat DDR SDRAM, and let's face it, it doesn't, except in some synthetic benches which don't behave like any real applications do.

L

 

[amebapk]

ok, soon the P4's bandwidth will be up to 4.266GB/s, but so? suck still sucks, even let P4 up to 10GB/s, but still can't go over AMD's thunderbird.

 

[Guest]

I wonder if QDR-SDRAM will be faster then RDRAM. Why is PC2100 (266mhz) sdram only using a 64-bit data path. The Geforce2 and radeon use 128bit DDR-SDRAM. Heck the Geforce3 later on will use 256-biy DDR-SDRAM.

I only steal the princess cause she is hot ;*)

 

[AmdMELTDOWN]

"AMD/VIA...you are the weakest link, good bye!"

 

[ksoth]

Actually, GeForce and Radeons use DDR SGRAM, not SDRAM The reason for SGRAM having 128-bit wide data paths and SDRAM having 64-bit wide data paths is the wider the data path, the more traces there needs to be on the PCB. That's the main reason why we aren't quite seeing dual-channel DDR SDRAM. I read somewhere that if someone were to implement it, it would require like an 8 level PCB, which is just too expensive and complicated to make right now. Because RDRAM works off of a 16-bit data path, implementing a dual channel solution for it doesn't require nearly as many traces or layers, even less than that of a single SDRAM channel I believe. Because a graphics card has such few parts to it compared to a motherboard, it can afford the extra real estate taken up by wide data paths. Something that backs this up is mobile graphics solutions, like the GeForce 2 GO, which I think only uses a 32-bit wide memory path because with mobile computers, space is everything, and more traces probably just won't fit. Plus, the GeForce 2 MX uses a 64-bit wide path to keep costs down as well. We will probably see QDR SDRAM before dual-channel solutions for this reason, considering QDR SDRAM will offer double the bandwith of DDR SDRAM without any more traces being required.

"Trying is the first step towards failure."

 

[Raystonn]

That whole article, and hence your whole post, is based off this one line they state, but do not back up: "Finally, the Rambus DRAM did not perform any better than PC133 SDRAM in actual systems." That may have been true in a Pentium 3 system, where the bandwidth was bottlenecked by the processor, but certainly not in a Pentium 4 system.

The article completely avoids whole dead wait state latency issue of SDRAM when switching between reading and writing, and uses Pentium 3 benchmarks to show RDRAM as slow. Thus, it's completely invalid. RDRAM really does have less latency in normal situations of medium and high memory use applications than SDRAM.

"I know that HardwareCentral have put out an article about how RDRAM has lower latencies etc but let's face it, it CANNOT be true."

So only articles that you agree with can be true?

"If it were RDRAM, with its alleged better bandwidth AND latency would easily beat DDR SDRAM and let's face it, it doesn't, except in some synthetic benches which don't behave like any real applications do."

The synthetic latency benchmarks are where SDRAM shines. This is because they do not mix reading and writing in the same test. Whether by design or by accident I cannot say, but it's a flawed test. RDRAM does much better in real world situations of mixed reading/writing.

-Raystonn

= The views stated herein are my personal views, and not necessarily the views of my employer. =

 

[Guest]

"The synthetic latency benchmarks are where SDRAM shines. This is because they do not mix reading and writing in the same test. Whether by design or by accident I cannot say, but it's a flawed test. RDRAM does much better in real world situations of mixed reading/writing."

Strangest thing.. real world test is where P4 doesnt exactly shine right now, which is why you always like to refer to SpecInt kind of synthetic benchmarks in every Athlon vs P4 thread..

 

[Raystonn]

"Strangest thing.. real world test is where P4 doesnt exactly shine right now, which is why you always like to refer to SpecInt kind of synthetic benchmarks in every Athlon vs P4 thread.."

Remember not to confuse the P4 vs. Athlon debate with the SDRAM vs. RDRAM debate. Those are two entirely different subjects.

RDRAM does well on synthetic _bandwidth_ benchmarks compared to SDRAM. SDRAM does well on synthetic _latency_ benchmarks compared to RDRAM. RDRAM continues to do well on real world bandwidth applications, such as Quake3A, when compared to SDRAM. SDRAM does not continue to do well in latency intensive applications when memory is being used normally (normal mixture of reads and writes) with a medium to high load on memory use, when compared to RDRAM. To see this, either run a few latency intensive applications at the same time, or run a synthetic benchmark that mixes reading and writing back to back.

-Raystonn

= The views stated herein are my personal views, and not necessarily the views of my employer. =

 

[Guest]

"Remember not to confuse the P4 vs. Athlon debate with the SDRAM vs. RDRAM debate. Those are two entirely different subjects."

Yeah you're right.. Im an idiot.. of course there are already RDRAM Atlhon boards and SDR/DDR P4 boards available, that allow us to really compare the different memory and cpu architectures and their impact on benchmarks and real world apps..

AS it is now, its very hard to seperate the P4 vs Atlhlon and RDRAM vs the world discussions. Sure, I know SDR/DDR/ and RDRAM are available for the P3.. but I dont think there are too many conclusions to be drawn from that platform. There is basically no difference to speak off between any of them.

Besides, I just wanted to point to the fact that you like to refer to real world application performance when discussing RDRAM, and synthetic benchmarks when it comes to cpu's.. Whichever suits your arguments best.

 

[Raystonn]

"Besides, I just wanted to point to the fact that you like to refer to real world application performance when discussing RDRAM, and synthetic benchmarks when it comes to cpu's.. Whichever suits your arguments best."

I usually discuss both synthetic and real world benchmarks when in a discussion about a certain technology as a whole. Certain threads are specific to certain benchmarks though, and hence are focused in one direction or the other.

-Raystonn

= The views stated herein are my personal views, and not necessarily the views of my employer. =