P4 predictions

[PiLsY]

A thought struck me just 30 minutes ago. At the moment with DDR memory were all trying to run it as fast as possible, and sacrificing cpu speed if necessary to get there (3:4, 3:5 dividers etc). Now as the P4 speed ramps up, and the multiplier follows suit I believe were going to see a case fo diminishing returns. For example, a P4 1.6a with its 16 multiplier is capable of running at 190fsb. The ddr memory is capable of 240 fsb. This means 180/240 is the best overall performance. Without a shadow of a doubt. Now as multipliers inctrease we'll start hitting problems. Theres going to be a multiplier thats a sweet spot for every system. IE if your ddr can hit 240 fsb you want a p4 that maxes out at 180 fsb. Something like a 2.4b or possibly even a 2.53b if you can keep them cool enough. However, now were starting to reach towards 3ghz, and thats going to cause problems for ddr. Were starting to see chips top out at lower and lower FSB due to higher multipliers. What this means real terms is that were having to back off our ddr clocks quite significantly. For instance a 2.8b maxing out at 170fsb - you lose a full 15mhz memory speed by being forced down into the 170/230 fsb:mem range. The implementation of new dividers may help this issue somewhat, but its certainly not the cure. Dual ddr also, fair enough it may provide enough bandwidth for the p4, but whos going to be happy having to underclock their memory, irrelevant of how good it is underclocked?

One thing is for sure, the only people benefiting from these higher and higher multipliers and lower and lower FSBs are Rdram users. 160 - 175 seems to be the top out stage for the high end rimm rigs (th7-ii - seen a few of those running 170 x 4 mode). With 16 and 32 bit rimms slowly making their way into production now things can only get better for rdram. Dont forget guys, dual channel 16 bit rdram was the starting block. 32 bit rimms are kicking ass now - imagine what would happen with dual channel 32 bit rimms? Im sure someone behind a desk somewhere is working along these lines right now.

Anyway, for now and certainly the next 6 - 8 months for me DDR is the way to go. By the time we start losing siginificant memory clocks the dual channel chipsets will be here. This should take care of the problem until something new and better comes out. Im predicting 6 - 12 months before rambus takes its crown back from the dual ddr rigs. My betting is on dual 32 bit Rimms for the next big rambus move. We all know about dual ddr, but the question to ask is whats next?

Its going to take a lot more than spit and string "get-by" solutions for current memory to keep up with the next generation of 4ghz - 8ghz P4s (not as far off as you think - 2 years or so). Manufacturers certainly cant keep ramping up the clock speed - current ddr yields in the 200mhz range are unimpressive to say the least. Most sticks arent truly 200mhz stable. Dual ddr may be a short term fix, but the long term fix can only be to change the whole architecture of the memory to allow the use of 400mhz plus BGA dimms as found on gfx cards.

Ah well, enough wittering, just a mornings worth of thoughts for you all to mull over.

PiLsY.

[QuadDamage]

Its going to take a lot more than spit and string "get-by" solutions for current memory to keep up with the next generation of 4ghz - 8ghz P4s

i would not agree with this, since p4's are very much maxed out at around 4ghz. i can see no headroom for p4 to hit 4ghz stable with current manufacturing process. except for this i agree with you. you made some valid points.

[docah]

The P4 Will probably need another process revision to reach 4ghz stable. 9nm processes? and i think getting past 5-6 will be moot, aren't the new cpu designs running lower clockspeeds? I thought i read that someplace. it's all hearsay though

[PiLsY]

C1 step has given intel the headroom to get to 3.8ghz by process end. Either another stepping or a new core design (much like the willamete to northwood change) would see maybe even 5ghz. The chip may not be called a P4 but that wasnt the point I was trying to make. Fact is because of the way our systems work a chip will allways have a multiplier, and the higher the multiplier the lower the FSB. For instance, a 30 x 133 multiplier for 4ghz. It leaves less ceiling for FSB clocking. Regardless of whether its a p4 or a p54 or whatever, this single fact remains. DDR architecture as is is having problems surpassing 200mhz (400 ddr) reliably, so were looking at moving to a new architecture. My overall point was where do we go from here?

Will it be higher efficiency processors (but that will mean that overall bandwidth internally will be higher, again necessitating higher bandwidth memory which tbh there just isnt the capability for.

There must be some new memory on the horizon as were already seeing chip bandiwidth being held back by memory. I for one say DDR is on limited time now. The bandwidth potential of dual DDR is around 6.4gb/s running at 200mhz. Dual 32bit Rdram (if thats the way rambus goes) has the potential for 8.4gb/s. Where next? 64bit memory? Quad ddr/rdram? The one thing for sure is thats omething has to evolve to replace our current memory technologies.

PiLsY.

[Obscure]

Hey Guys,

I remember reading somewhere that 4-5ghz will be max even at a 9nm die process. They said after that, they will need to get into Nano Technology because of the size and sheer quantity of transistors onboard. Ill see if I can find the link, it was a really good read.

[QuadDamage]

not necessarily PilsY. Intel will prolly move to 166fsb, therefore you would need only 24x multiplier to have your cpu running at 4ghz.

[Tony]

Intel will move to a 166fsb by late next year and there after a 200fsb will be on the cards.Amd's move to 166fsb is here now and I believe a move to 200 fsb will only be a year off.The thing with Amd cpu's is most use a 200fsb or close to it now..so the only benifit these will see is when the cpu's have 512cache.DDR's life for me as it is now is very short as I don't think we are going to see true ddr433 or ddr466 and 500.We may get chips that will run that fast but I think most of the manufacturers are moving to dual ddr333 and dual ddr400.I think we then will see a large jump with dual ddr but what this will be is only speculation as this is about 3 to 5 years off.

[PiLsY]

Wake up guys youre not listening :).

Current ddr technology is peaking at around 200mhz. I cant see it getting jedec approval much higher although we can o/c up to 250mhz. Now. This is the bit im concerned about.

We would either need to run 250mhz fsb or use dividers to run anywhere between say 150 and 200mhz. Take your 24 x 166 p4. Assume its going to be overclocked by 33% (quite conservative based on current yields). That gives you a bus speed of 220mhz. There is no way I can see this being divided up to 250mhz or this chip config achieveing 250mhz.

My point was that we are now seeing chips hitting such speeds that at stock theyre going to be pushing the memory to its max and beyond. How does this weigh up for us clockers? It means well need a move to a new memory architecture. DDR2 may provide some relief, but the cpu industry is much more quickly overtaking the memory industry in terms of bandwidth requirements compared to supply.

Look in terms of the last few years. Weve gone from pc133 sdram to pc266 ddr to pc333 ddr and now maybe pc400 ddr. And this is in 6 years. Yes thats right. 6. Now look at cpus. Weve gone from that athlon 1ghz @ 100mhz and are now seeing the xp @ 166mhz. This is in 2 years. The p4 has ramped up from 100mhz to 133mhz and now with talk of a 166mhz bus. This is over the course of 12 months. XP is double pumped, P4 is quad pumped. Matters are made worse when you overclock. With AMD apparently rmeoving multiplier adjustment very soon and intel having offered none for one hell of a long time we could be in for a hard time. Already were at the stage where the highest overclocked ddr cannot provide sufficient bandwidth for a p4 at STOCK speed (500mhz ddr compared to 533mhz quad bus @ 133mhz). The athlons wont be far behind. A multiplier locked AMD @ 166bus speed is bad news for DDR. espeically on a 200mhz fsb (only allows for a 25% overclock whilst retaining synchronous memory speed - dropping memory below bus speed will be disasterous for overall performance). With amd solely relying on DDR as its foremost memory architecture hard times are ahead for AMD. If intel keep on rambus as is now being suggested then they have a longer life ahead, but we all know how severely rambus hampers overclocking (4x versus 3x - a whole different argument). Either way, current memory will need to evolve significantly to carry forth the new generations of AMD and Intel cps whilst keeping us overclockers happy. Less "Xtreme" times are ahead unless something new memory wise is on the horizon.

PiLsY.

[docah]

If you check out micron's webpage and look at the DDR roadmap you see 300 and 400 mhz chips coming out q1 and q3 next year.

So yes, the short term future looks semi-bleak. But 300/400 mhz chips next year? that's ddr speeds of up to 800mhz. So, we'll see what happens i guess... http://www.micron.com/products/category.jsp?path=/DRAM/DDR+SDRAM

[DuraN]

This will prolly change some of this discussion...i think. Intel is not making any more RDRAM chipsets.

http://www.xbitlabs.com/news/story.html?id=1031787073

[sysfailur]

Still though, memory needs to start getting faster. They are making blazing fast DDR for graphics cards... why aren't we seeing the ram that's on our 4600's on our system chips. That's what I've always wondered. And I believe there are already 900mhz DDR chips being sampled. For the love of it all. Slap that on a stick of system memory and lets get the party started!

I know it's not as simple as that, obviously you've got to probably make a certain PCB board that will handle it but I'm sure it can be done. It's stupid that we're still using the old style of ram chips. Samsung should convert to BGA instead of TSOP like kingmax, but just do it BETTER than kingmax ;)

[DaGooch]

A simplified but rather quite comprehensive read on memory in the near future.
http://www.aceshardware.com/read.jsp?id=50000275