Matching FSB and Memory

Yeah, DDR2 runs at double the bus speed while the CPU runs at 4x. So a 1333MHz bus would be 1:1 with DDR2-667 RAM. You can forego the 1:1 ratio if you get fast enough RAM to overcome any penalties of not running synchronous.

Here's an older article that is still good:

http://www.madshrimps.be/?action=getarticle&articID=472

Thats not exactly accurate, or I would even say its just wrong.

Ram runs at 2x bus speed only when 1:2 ratio is applied

System clock combined with fsb:ram ratio gives end Dram frequency, it is not system clock x2. Data rate is the Dram frequency x2(but thats irrelevant really, created for marketing purposes) So system clock 333MHz with 1:1 ratio = 333Mhz Dram frequency = DDR 667

Cpu end frequency is the system clock x Cpu multiplier

FSB is the system clock really, and yours doesnt run at 1333MHz. Again that was created for marketing purposes and to create confusion. Its called quad pumped fsb because it can transfer data 4 times per clock cycle instead of 1. So it has bandwidth "as if" it ran at 1333Mhz with one data transfer. Now, when it comes to selling a product, FSB 1333 sounds better that 333MHZ, doesnt it?

Same goes for DDR memory, but its two data transfers per clock.

The situation is further complicated because everyone seems to define the terms a little differently.

True that.


But that aside, for running your memory 1:1 you need to set it to 333Mhz/ddr 667, although if i were you I would do it differently - im sure your cpu is more than capable of running at fsb 400+ with ease, so instad of running it at stock 333x9 you could go for something like 400x7.5 or 433x7 which would make your memory run faster and closer to spec while still keeping the 1:1 ratio. That is of course if you can be bothered to experiment a little.

or set the FSB to 533, and use DDR2-1066  

MaDMagik has got the topic covered well.

What I originally did was get 1066, so I could overclock without worrying about the memory. I ran a 460Mhz FSB with the RAM running 1:1 (DDR2-920). This also let me reduce the CAS timings which gives a performance boost on its own.

Also lots of boards have custom memory multipliers (e.g. 1:2.6), however I prefer to run native and use lower CAS timings than rely on these, as I've found they can cause instability when overclocking and the speed increase isn't always as expected.

I opted for a 533 FSB, and lowered my multiplier to 8 (9.5 was 5.06Ghz, and it wouldn't even turn on ehehe, boots Windows at 4.75Ghz but is as stable as OJ with a knife). I'm told using this FSB on a quad core is more difficult (seems to be some motherboards can, some motherboards can't). (Also side note, anyone who runs off to try 533, watch your NB temp, I had to replace my NB cooler).

Also do not forget Dual Channel mode. In dual channel mode (using two identical DIMMS), twice the data can be read at once. Wikipedia has a good article on dual channel memory. I beleive this doubles the memory throughput again (on top of the double data rate technology). Performance isn't actually double though, because it's only useful when requesting blocks big enough.

I need to re-read that article. But I was lead to beleive that DDR2-667 in dual channel can fill Intels FSB1333 quite nice.

333Mhz bus x 2 (Double Data Rate) x 2 (Dual Channel) = 1333
333Mhz bus x 4 (QDR Intel Bus) = 1333

Back in the P4 (200Mhz bus) days, DDR-400 in dual channel was common place.

Running RAM at 1:1 and tightening the timing will give you more of a performance boost than just overclocking the memory.

For another thread on the topic:
http://www.tomshardware.com/forum/251715-29-ratio-myth

Thanks to everyone for their help. I think what was confusing is I thought you had to match the actual memory clock to the system clock so in the case of DDR2 667 the memory clock is 166Mhz and the bus clock is 333Mhz which would make sense. Trying to match the memory clock to 333Mhz would mean finding Ram that ran at a bus speed of 667 theoretically DDR21333 which I'd never heard of. However if its the bus speed that needs to match and not the memory clock itself then I understand.

Thanks guys

I think you still do not understand. If you had SDRAM, a 333 MHz FSB:333 MHz mem clock would be a 1:1 ratio.

DDR2 RAM transfers a word on both the leading and trailing edge of a bus cycle, so the corresponding memclock would be double the FSB freq for the same 1:1 ratio.

I know what you mean. Its because the memory clock and the bus clock are 2 different speeds on DDR2 at least they are according to this link:

http://en.wikipedia.org/wiki/DDR2_SDRAM

And the following sentence in particular:

'The key difference between DDR and DDR2 is that in DDR2 the bus is clocked at twice the rate of the memory cells, so four bits of data can be transferred per memory cell cycle'

And its the memory cells/clock that I thought had to be 333Mhz to match the system clock and not the bus clock. I understand what Double Data Rate means. Basicaly the explanation from the above link means that A DDR1 module would run at the same bus speed as the memory cells/clock but DDR2 doubles the bus speed of the memory clock so a 133Mhz memory clock would actually be 266Mhz for DDR2-533. The 533 coming from applying the double data rate to the bus clock.

I know this sounds like a mess and I'm not great at explaing things sometimes but if you follow the link hopefully you'll see what I mean....

 

The total bandwidth for the cpu (theoretical) is 3.2 gbs for every 100mhz physical on the fsb so at 200mhz (fsb800) that makes 6.4gbs while the bandwidth from standard ddr2 and ddr3 isn't any ware near that good at those speeds. In the end Intel screwed up it's I/O on it's post socket 370 up to s775. The only time when the ram to cpu fsb bandwidth was the same was socket 423 with rdram but lantancies ate into that so only 2.7gbs was achievable but 45ns was the cause. One can look this stuff up but will have to hunt through more than a decade worth of articles and reviews while experimenting first hand. Socket 7 best I have ever seen was 205mbs while the best for native slot1 was 550mbs while socket 370 with the 815 got the max for pc100 sdr at 800mbs at 99% efficiency. So after testing several s775 rig nvidia got 60~80% efficiency so a typical user got less than what the ram could perform while Intel at got better results. AMD however has a much better I/O performance and performs well but their weak fpu (5~10% slower per clock) holds them back.