8.4GB/s of memory bandwidth next year

Page 2 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
"Ray memory bandwidth is only one aspect of performance...just as FPU performance is only one aspect, etc...I could care less what a single part of a system's performance is as long as the overall system performs well."

That is fine. The sum of total performance is the performance of each component combined in some form. For this thread we are discussing the memory subsystem. This is one of the main components that compose your system. It is the main bottleneck in your system when dealing with large quantities of data. The CPU is the main bottleneck in your system when dealing with relatively small chunks of data but performing a great deal of work on that data. To achieve better performance, you need better and faster memory and CPUs.


"Athlon systems still significantly outperform the P4 as far as bang for buck."

That is a monetary issue only. The Pentium 4 delivers the best performance, but does so at a higher price. If price is an issue, buy what is less expensive. Price will always fluctuate, but hardware is what interests me. I leave the economics to the sales and marketing force and embrace the best technology. Noone can say what any piece of hardware is really worth, as that is relative to the individual. We can however look at specifications for peak performance and run benchmarks to obtain actual average performance. We can look at these performance figures and analyze what is the best performer and what has the best potential for the future. I find this fun, and I am sure I am not alone.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
 
If I had an Athlon system, I would be lobbying AMD for a DRDRAM chipset. Since I do not, I am not very much affected.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
 
Never thought of AMD and DRDRAM together, interesting thought, what’s the possibility of that ever happening though?
If they could make it viable it could give them another avenue for product expansion maybe.
I have been waiting for AMD to make the next BIG leap with something soon, but how long until we see an improvement on existing technology?



<font color=orange>Beam</font color=orange><font color=red> me</font color=red><font color=green> up</font color=green><font color=blue> Scotty</font color=blue>. :wink:
 
It is inevitable that within 5 years they will be forced to create a DRDRAM chipset. Looking at the roadmaps for DRDRAM and SDRAM (including DDR, QDR, DDR-II, etc.), all forms of SDRAM will have been far surpassed by then. I doubt they have any plans to die off like Cyrix.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
 
If price is of no concern why don't you go buy a super computer and stop spewing Intel is Supreme biased BS!!!
As was mentioned earlier the P4 has superior memory but look at how it benches as a whole compared to far cheaper and slower clocked machines and memory types (DDR).
I get the feeling that if Intel would re-release the Pentium 2 you (and others) would be telling us how great it is compared to Athlons.
The bottom line is Intel was the king until Athlon. Now it is not as clear. Yes, the P4 2 GHz is slightly faster than an AMD 1.4 but at more than double the price when counting everything CPU, MB, & memory. If money is no object why even argue about the nForce vs. i850 performance(especially before it exists)? Think about it who would have guessed a Amd 1.4 would outperform a P4 1.8 before the P4 was introduced. Nobody. Don't waste your breath. You, Fugger, Meltdown, and Juin will go and buy whatever the latest and greatest product the Intel Gods have to offer regardless of price or performance. Some will buy AMD no matter what and the rest of us will see how they compare in performance and price and go from there. Talk is cheap.
 
"why don't you go buy a super computer and stop spewing Intel is Supreme biased BS!!!"

Go elsewhere troll. You are not welcome in my thread. This thread is about various forms of memory and their roadmaps and specifications. Nowhere do I even attempt to compare various companies. So again, go elsewhere troll. You are not welcome here.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
 
We all know that DDR handles memory bandwidth more efficiently than RDRAM. And even if cpu bandwidth to memory is only 2.1 GB/s Total system bandwidth will be 4.2 GB/s!
Athlon in it's current form is more than capable of handling the Current P4. The only advantage I see for RDRAM is it's bus speed. I believe DDR333 if we ever get to see it on a Athlon system will out perform PC800 RDRAM!

As many of said memory bandwidth is only one aspect of the bottlenecks in todays system. So don't put all your chips in one basket raystonn.

These are just my views, I believe the only way to truly see an even benchmark is to have both Cpu's using the same memory types. Cause the design differences in both DDR and RDRAM wouldn't not make for an fair comparison!
 
"We all know that DDR handles memory bandwidth more efficiently than RDRAM."

Where did you hear that? This in untrue. DRDRAM has a much greater peak bandwidth than DDR SDRAM and makes use of it much more efficiently with a greater effective bandwidth. The only area in which SDRAM can claim superiority is latency, and this completely disappears with PC1066 DRDRAM. Once this is introduced, DRDRAM will have lower latency than all current forms of DDR SDRAM.


"And even if cpu bandwidth to memory is only 2.1 GB/s Total system bandwidth will be 4.2 GB/s!"

And most of that will remain unused unless you use the integrated graphics chipset. Sure, your hard drive and such will use a bit. But 100MB/s is hardly a drop in the bucket compared to the 2.1GB/s that is going wasted, sitting idle. The nForce's design was meant to boost the performance of their integrated chipset. Clearly nVidia hopes to move their performance video cards into their chipsets at some date, but that is not currently the case with the MX chipset they are now using.


"Athlon in it's current form is more than capable of handling the Current P4."

This is not a P4 versus Athlon thread. We are discussing different forms of memory, along with their specifications, benchmarks, and roadmaps. DRDRAM would do quite well when paired with future versions of AMD's processors, but I do not foresee this happening for another 5 years or so.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
 
"Where did you hear that? This in untrue. DRDRAM has a much greater peak bandwidth than DDR SDRAM and makes use of it much more efficiently with a greater effective bandwidth. The only area in which SDRAM can claim superiority is latency, and this completely disappears with PC1066 DRDRAM. Once this is introduced, DRDRAM will have lower latency than all current forms of DDR SDRAM."

In theory, but I have seen no benchmarks or even previews that suggest this Ray. It's Great if it's true, but right now it's so much rumor until released and benched.

What I think the poster you were responding to was talking about is the actual use of the ram and it's theoretical bandwidth. I haven't seen any benches out there showing more than around 55 or 60% of theoretical bandwidth of RDRam, but I've seen benches showing 80-90% usage of DDR Ram theoretical bandwidth (KT266A reviews)

I also have yet to see proof of your assertion that as the P4 CPUs become faster they make more efficient use of RDRam. All the benches I've seen of P4...right to P4 2.2 overclock seem to indicate at most a 60% bandwidth usage...and stilll serious latency.

Mark-

When all else fails, throw your computer out the window!!!
 
This is the most interesting, educational thread I've read in a long time. Raystonn, I've
always had respect for your opinion. FatBurger, I'm getting some for you as well.

Forums as they should be, Thank You.

RC

How do you eat a elephant? One bite at a time!
 
Here's my take. The current P4 design was developed with the RDRAM in mind, whereas the current Athlon was designed with DDR in mind. Would Athlon even benefit from RDRAM? Maybe, but I think that it's ability to take information from the RAM would peak out somewhere before the actual bandwidth was exhausted. Therefore, as things stand right now, I think that AMD would be unable to justify the use of RDRAM for their CPU. With the last release of an Athlon CPU being a couple (?) months ago, I think it may be safe to say that the current incarnation of the CPU is at it's end, and that AMD is putting the time and effort into the Barton and the 'Hammers to make them able to make use of those high theoretical bandwidths. I may be wrong, but call me optimistic. We will not know which version of RAM, be it RD, DDR, or QDR will triumph in the next year. With the various events in the technology world, it would be hard to imagine one winning over the other without both chip makers throwing support behind it.

And my question to you is this Ray: At what point does this massive increase in memory bandwidth become a matter of white papers and specs and no longer actually results in any real world boosts? I mean, would the P4, when it reaches 3.5GHz in Q2'03 (18 months from now) really be able to use that much bandwidth, or will it just sit by the wayside, lonley and unloved? Also, I know that eventually, things (i.e. software) will catch up, but if the RAM bandwidth increases faster than the needs for bandwidth (who really need 250 FPS in Quake 3?), why bother, especially if it requires much money for us end-users. I know, I'm kind of a wet blanket here, but I really don't see the need for these things in the mainstream market for at least 2 years. And by then, the next latest and greatest thing will be just over the horizion. This might be a great development for businesses running 64-bit apps in a year, but we, the real market (at least the one that accounts for a majority of CPU sales) really won't need it for quite a while. Personally, I think that for the computer enthusiast, there should be more of a push for tweakability and CPUs that don't require a jet turbine on top of a pure silver 'sink (this goes for both CPUs), and a push from the software makers to actually make something that isn't bloated to the gills with frills that we don't want and loose code that could have been tightened with a few extra man hours. Then we might not need mammoth amounts of bandwidth to access and store that code, since it will be nice and compact (until we move to 64-bit).

If that made no sense, I apologize, since I'm under the influence of cold meds, noisy roommates, and random coughing and sneezing fits.

-SammyBoy
 
"In theory, but I have seen no benchmarks or even previews that suggest this Ray. It's Great if it's true, but right now it's so much rumor until released and benched."

This is in specification of course. For actual benchmarks we will have to wait until it is released. The specifications indicate peak performance. The benchmarks indicate effective performance.


"I've seen benches showing 80-90% usage of DDR Ram theoretical bandwidth"

Can you paste some URLs for me to peruse?


"your assertion that as the P4 CPUs become faster they make more efficient use of RDRam"

I do not think I ever asserted that. Efficient use of DRDRAM stems from the speed of the DRDRAM as well as its electrical characteristics. It also has much to do with the software being run. Typical software performs approximately 70% reads and 30% writes, accessing memory mostly in areas close together, with the occasional memory access off in a remote memory location. Most synthetic benchmarks do not take this into account at all, doing pure reads in either totally random memory locations or in a strictly sequential manner. If we had a benchmark that mimicked this real world use pattern, we could get a better measure of how each memory system performs on real applications and games.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
 
"At what point does this massive increase in memory bandwidth become a matter of white papers and specs and no longer actually results in any real world boosts"

When memory accesses become rare, this will happen. Because memory is the core of our systems, the place where both code and data reside, as you increase memory performance you increase the ability of the CPU to retrieve instructions and data more quickly. Past generations of CPUs had to rely on caches (L1, L2, etc.) in which to store both instructions and data because the memory subsystem was extremely slow. Imagine, however, if the memory subsystem was as fast as our cache currently is. We would have no need for cache at all. It would be the equivolent of up to 4GB of processor cache where all code and data would permanently reside. This is somewhat of a holy grail of memory design.

I am sure we have all seen the tremendous increase in performance of all CPUs when the cache size is increased. Applications and data continue to get larger and larger. They will inevidently overflow our small caches. This is when memory performance becomes the primary bottleneck in the system. As we increase memory performance, we further reduce the bottleneck and enhance performance.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
 
I apologize in advance for anyone who already knows this, but after reading through the thread it seems as if some folks are forgetting the following:

CPU designs are created with some certain parts in mind which are on the market at the time of the design. Tradeoffs are necessary regarding memory performance due to the fact that as technology marches forward, memory access frequency (or whatever it is called) is not keeping up with CPU operational frequency. Caching methods are what is used in PC's to get around that fact, and are directly responsible for any percieved performance differences between RAM types being used given a fixed CPU type. As has been shown: a design which was optimized for use with one type of memory may not be able to fully take advantage of some newer memory technology, without a complete redesign of the memory accessing interface and thus seem as if it didn't really need the faster memory.

For example: it may be true that if Intel's Pentium 3 design was optimized with PC100 memory in mind, with a reoptimization of the memory interface to be used with PC10,000 (or whatever), the very same pentium 3 would suddenly be a lot faster, and with no changes in the actual CPU itself.
 
Sorry I misunderstood perhaps, the actual vs theoretical performance of DDR memory. It's closer to 75-80% than 80-90%. What is 80-90% is the performance improvement of DDR RAM over SDRam memory performance with the KT266A as outlined in these two articles:


<A HREF="http://www.anandtech.com/chipsets/showdoc.html?i=1528&p=5" target="_new">http://www.anandtech.com/chipsets/showdoc.html?i=1528&p=5</A>
<A HREF="http://www.lostcircuits.com/motherboard/via_kt266a/4.shtml" target="_new">http://www.lostcircuits.com/motherboard/via_kt266a/4.shtml</A>

Mark-

When all else fails, throw your computer out the window!!!
 
"I've seen benches showing 80-90% usage of DDR Ram theoretical bandwidth"

Can you paste some URLs for me to peruse?

Here you go Raystonn

<A HREF="http://216.194.77.198/articles/2001/july/010720_Sandra_Bandwidth/010720_Sandra_Bandwidth.htm" target="_new">http://216.194.77.198/articles/2001/july/010720_Sandra_Bandwidth/010720_Sandra_Bandwidth.htm</A>
 
Besides being written by Van Smith, who is generally regarded as extremely biased for AMD, the benchmarking application that was used was actually testing the differences between the two processors and chipsets. The processors both have different sizes of cache, as well as radically different chipsets. The only way to actually test bandwidth efficiency between different types of memory are to use the same processor and chipsets as close in design and performance as possible. We will be able to do this at the beginning of next year when DDR SDRAM chipsets for the Pentium 4 become common in the marketplace.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
 
But it's still a Via chipset. To make Raystonn happy, it'll have to be an Intel chipset.

<font color=green>In memory of all the Americans that died 9/11/01
Rest in peace</font color=green>
 
THe K8 aka HAMMER.. has its memory controller and northbridge on DIE.... this has to reduce latency... can u explain what kind of bandwidth boost AMD can expect with that
 
Nope, because there aren't any actual Hammer processors in existance. There's no way of saying how well something new that's that radical will perform.

<font color=green>In memory of all the Americans that died 9/11/01
Rest in peace</font color=green>
 
I can only tell you a couple of things regarding that design at the moment. That processor will be completely locked into whatever memory type that controller uses. Additionally, you will no longer be able to control the FSB speed from the motherboard's BIOS. It will be controlled by AMD. You can say goodbye to one of the major methods of overclocking. Overclocking of the memory system will be entirely impossible. Should AMD decide to lock the multiplier on its processors as well, you will be unable to overclock at all. This would be a dark day for the hardware enthusiast scene.

-Raystonn


= The views stated herein are my personal views, and not necessarily the views of my employer. =
 
I have worked for 10 years as an assembly/C++ programmer. Some people on this forum seem to think they know it all and you can tell they don't really know what they are talking about.

A) 32 bit over 16 bit will never EVER mean double the performance. All that means is it can transfer twice the amount possibly but if you look at how packets are organized it will only increase the total USABLE bandwith between 60 and 65 percent.

B) Some people were mentioning that higher speed RDRAM has less latency. DUH cause it is running at higher clock speeds it isn't as apparent because more clock cycles are going on to fetch the data. Except you must remember this unless they are changing the entire structure of RDRAM this simply masks the problem. You still lose way to many clock cycles to the latency issues. Although it becomes less aparent.
 
So basically what you are saying Ray is that Hammer tweaking will be limited to a similar degree to P4 with RDRam?


When all else fails, throw your computer out the window!!!