Intel going to integrated memory controller...

[Sacrificial]

Intel is going to an AMD-style CPU-integrated memory controller.

http://www.theinquirer.net/?article=14828

Integrated memory controller = reduced latency. Hopefully it offsets DDR2's supposed additions to latency.

It's not news but Intel sure has put a lot of eggs into the Mhz basket (Prescott and DDR2). So far it hasn't paid off (478 Prescott and DDR2 400 and 533) in terms of performance, but that was what everone thought about the Williamette.

(I'm sure *somebody* will come in some day and rant as if AMD copied Intel's unfaltering lead on this one too...)

 

[burningrave101]

An integrated memory controller should help alot in areas where timings fall into affect. DDR2 is suppost to have high latency at first but i suspect as time goes on they will be able to get it lower just like they did with DDR.

Deeper pipelines such as in Prescott are better for data-intensive applications like games and multimedia so once Prescott hopefully gets a cooling off some and climbs over 4Ghz along with maybe this on-die controller, it should be a really good gaming processor.

 

[xonik]

Originally posted by burningrave101
Deeper pipelines [...] are better for [...] games

I would like to know how you came to that conclusion. I am not arguing for the opposite of your statement; rather, I have not seen a direct correlation between game performance and pipeline depth. Cases in point:

http://www.hardocp.com/article.html?art=NTgzLDU=
http://www.anandtech.com/cpu/showdoc.html?i=1965&p=4

Neither the 'C' nor the 'E' model decisively appears to hold an advantage in gaming situations.

 

[burningrave101]

Well i have read it in a couple of articles online although i would have to look to try and find them and there is an article over Prescott in the April 2004 issue of MaximumPC that says exactly that word for word. I'm just going off what i've read.

If you look back at the Athlon XP's, the Pentium 4's are better gaming processors even compared to a 2500+ mobile at 2.4Ghz.

http://www.techreport.com/reviews/2004q1/athlonxp-m-2500/index.x?pg=1

Its not so much the case with the A64 for the reason of lower latency caused by the A64's on-die memory controller. That on-die memory controller is primarily the main thing thats giving the Athlon A64 its performance in gaming and office applications. If you look at the K8 architecture its really not much different then the K7 except for the 64-bit extention addon. The K8 instruction piepline is a little longer than the K7 design. One extra cycle in the instruction-decode sequence and one more for integer math relieve some constrictions in the K7 design. But since gaming performance responds greatly to low latency its a definite current advatage for the A64 in gaming.

Sticking with PC3200 2-2-2 instead of going to PC4000 will save you 3-5% gaming performance right there. The Pentium 4 having its controller in the northbridge just slows things down in quick access data applications like games.

You wont see Prescotts increase in performance for gaming and multimedia till it climbs to higher clock frequencies. It was the same way with the Pentium 4 vs Pentium III and Athlon when the Pentium 4 first released. The Athlon was of course a better gaming processor then. Northwoods 20-stage pipeline kept crawling further and further ahead though as it ramped to higher frequencies. It was really the exact same situation then as it is now with Prescott vs A64.

Is it not much the same way with video cards and their number of pipelines?

 

[xonik]

So what you're saying is that it's not the pipelines, but the higher MHz gained from the deeper pipelines? So basically, Prescott's great gaming performance in the future is reliant upon its ability to scale to very high frequencies. Did I read that right?

 

[burningrave101]

Originally posted by xonik
So what you're saying is that it's not the pipelines, but the higher MHz gained from the deeper pipelines? So basically, Prescott's great gaming performance in the future is reliant upon its ability to scale to very high frequencies. Did I read that right?

According to Maximum PC, “The main problem, however, lies in optimizing DirectX 9’s High Level Shader Language (HLSL) for different chipsets. GPU’s have very deep pipelines, much like the Pentium 4. If shader code isn’t optimized to keep a pipeline busy throughout each clock cycle, performance will suffer. Right now, the Half-Life 2 shader code is not optimized for nVidia hardware.” MPC Vol. 8 No. 11, pg 11. And thus the finger pointing began. nVidia screamed that the game was not optimized for their hardware and the tests were performed with the wrong drivers (the 45.23 detonators were used). Valve said the sad-sack performance was nVidia’s fault for designing poor hardware that doesn’t meat the DX9 specifications.

http://www.techimo.com/articles/index.pl?photo=138

Just like video cards and their pipelines, a processor must keep its pipelines full and busy in order to receive the performance gain from having more stages. If it doesn't do this then it hurts performance. In order to keep its pipelines full it must be at a fast enough clock frequency. Prescott currently is not. Prescott was designed to scale to very high speeds.

Games and Multimedia tend to favor deeper pipelines as long as there isn't a latency issue. The Pentium 4 and Athlon XP both had their memory controllers in the Northbridge so the latency was nearly the same for both. Its different for the A64 but probably not for long if Intel goes ahead and adds an on-die memory controller to their processor.

According to this research the P4 pipeline is actually not deep enough. That paper concludes that P4 performance could be improved by up to 90% by increasing the pipeline depth to around 50 stages and increasing the cache size.

http://systems.cs.colorado.edu/ISCA2002/FinalPapers/Deep Pipes.pdf

I'm sure that as long as Intel has been at this, they know the consequences for increasing the pipeline depth and they have plans for taking advantage of it.

 

[Josh_B]

"But since gaming performance responds greatly to low latency its a definite current advatage for the A64 in gaming. "

Not necessarily. The P4 benefits little from reduced latency, and prefers bandwidth to latency.

 

[burningrave101]

Originally posted by Josh_B
"But since gaming performance responds greatly to low latency its a definite current advatage for the A64 in gaming. "

Not necessarily. The P4 benefits little from reduced latency, and prefers bandwidth to latency.

Um, not really. If your motherboard supports dual channel there isn't a need for more bandwidth. Lowering the latency on just the RAM alone affects overall performance 3-5%. Its just like having 2GB of RAM, if your not running a program that needs that much then your not benefitting any more if you have 2GB than if you had 1GB. Games are NOT bandwidth intesive. Heavy applications like CAD are but with dual channel you have almost 6.4GB/s of bandwidth. Dual Channel on a P4 rig is like running DDR800 in your machine.

Good article on Memory Bandwidth vs Latency here -

http://www.pcstats.com/articleview.cfm?articleID=873

Latency has the same affect on any processor. The Athlon XP, Pentium 4, and A64 all run the exact same games and games are quick access applications.

 

[DDrasus]

games are not bandwidth intensive? try gaming with a nice vid card on aa pc100 p4 setup.

 

[xonik]

Results show that games are a little memory bandwidth intensive. As shown in these game benchmarks,

http://www.hardocp.com/article.html?art=NjAyLDQ=

The Athlon 64 FX-51 has a 1-5% performance lead over the Athlon 64 3400+ of identical clock frequency. Note that the FX-51 is also held back by registered DDR latency, so this figure is probably more like 3-5% on average.

Interestingly enough, the Pentium 4 platform doesn't seem to respond much at all to memory bandwidth increases:

http://www.hardocp.com/article.html?art=NDk1LDM=

 

[burningrave101]

Originally posted by DDrasus
games are not bandwidth intensive? try gaming with a nice vid card on aa pc100 p4 setup.

No games are not bandwidth intensive. That means that once a certain amount of bandwidth is met you really aren't gaining anything from more bandwidth. Take a look at the A64 3400+ vs FX-51. There is hardly a difference at all between the two even though the FX-51 has over double the bandwidth from a dual channel on-die memory controller.

As long as you've got around 3.2GB/s of bandwidth your fine. Dual Channel on P4 boards just about doubles that. So buying PC4000+ is just a waste of money unless you run bandwidth intensive programs.

The Athlon 64 FX-51 has a 1-5% performance lead over the Athlon 64 3400+ of identical clock frequency. Note that the FX-51 is also held back by registered DDR latency, so this figure is probably more like 3-5% on average.

Yea but alot of the time they are tied and sometimes the A64 3400+ wins in the gaming benchmarks cause of using unbuffered memory. The A64 3400+ only has 3.2GB/s bandwidth. It doesn't have dual channel. I believe the maximum current games can make use of is somewhere around 4.0GB/s or something of that nature. I've seen the numbers posted on other forums before.

The FX-51 vs A64 could have to do with other things besides just bandwidth also. It could be the FX-51 motherboards are better or the fact one is on socket 754 and the other on socket 940. Doesn't the FX-51 draw more juice then the A64's? I dunno. Lots of factors to figure in besides just a dual channel controller.

 

[xonik]

Originally posted by burningrave101
Yea but alot of the time they are tied and sometimes the A64 3400+ wins in the gaming benchmarks cause of using unbuffered memory. The A64 3400+ only has 3.2GB/s bandwidth. It doesn't have dual channel. I believe the maximum current games can make use of is somewhere around 4.0GB/s or something of that nature. I've seen the numbers posted on other forums before.

More often than not, as shown in the benchmarks, the FX-51 holds a marginal lead. Just like lowering the RAM latency, there is a small but consistent advantage to dual channel configurations over single channel setups in the AMD64 case, in games. Here are some other benchmarks to show that this is not a fluke:

http://www.anandtech.com/cpu/showdoc.html?i=2002&p=6
http://www.aceshardware.com/read.jsp?id=60000320
http://www.tech-report.com/reviews/2004q1/athlon64-3400/index.x?pg=5

Once registered memory is gone, there will be no exceptions.

The FX-51 vs A64 could have to do with other things besides just bandwidth also. It could be the FX-51 motherboards are better or the fact one is on socket 754 and the other on socket 940. Doesn't the FX-51 draw more juice then the A64's? I dunno. Lots of factors to figure in besides just a dual channel controller.

Sorry but I don't buy any of these reasons.

 

[burningrave101]

Originally posted by xonik
More often than not, as shown in the benchmarks, the FX-51 holds a marginal lead. Just like lowering the RAM latency, there is a small but consistent advantage to dual channel configurations over single channel setups in the AMD64 case, in games. Here are some other benchmarks to show that this is not a fluke:

http://www.anandtech.com/cpu/showdoc.html?i=2002&p=6
http://www.aceshardware.com/read.jsp?id=60000320
http://www.tech-report.com/reviews/2004q1/athlon64-3400/index.x?pg=5

In the Anandtech review, the 3400+ won 3 out of 6 of the game benches against the FX-51. If bandwidth made much of a difference in current games with the current processors which it doesn't after a given point then the FX-51 would have a much greater lead in performance because of having more then double the bandwidth and dual controllers. Almost always when the FX-51 is ahead of the 3400+ its only 2 or 3 FPS which can be attributed to other things. Registered RAM alone will not negate that much bandwidth if excessive bandwidth affects gaming performance. But like i said before, the maximum games can benefit from is somewhere around 4.0GB/s or so i've read. So since the A64 only has 3.2GB/s by default it can probably make use of a little more.

Once registered memory is gone, there will be no exceptions.
Sorry but I don't buy any of these reasons.

Unless you have proof there is nothing to say those reasons aren't it. The Abit IC7 Max 3 is faster then the Asus P4C800-E in almost every benchmark. It could be nothing more then the FX-51 Asus SK8V is a little faster motherboard. Most of the Registered RAM out today from Mushkin and Corsair is rated for 2-3-2 timings. It just has to spend one extra clock cycle for error checking. Sometimes they use PC3500, PC3700, or even PC4000 with much worse timings then that in the A64 and P4 systems. Motherboards have as much to do with performance as RAM or anything else.

I'm sure the extra bandwidth helps a little but not nearly enough to warrant the price tag on an FX-51 or FX-53 processor unless you work specifically with programs that can use a ton of bandwidth.