Socket AM2 and DDR2 Timing Disync?

[vitaminc]

http://www.theinquirer.net/?article=31874

Socket AM2 screws up the DDR-II clock

Chip can't get the memory clock right

By Theo Valich: Monday 22 May 2006, 17:29

OH BOY. I can't think of many reasons why AMD would not like us to review the Socket AM2 CPU, but this reason could be a killer thing for all the reviews that will land tomorrow. Expect at least one "accidentally-leaked" marketing or sales presentation from Chipzilla where this thing gets blown out of all proportion.

It seems that AMD's memory controller has a bit of a maths problem. It has a habit of downclocking the system memory by a notch or two. Nothing serious, of course, it's just a matter of the fact that the HyperTransport divider is set at "5", and the number it is dividing is 1000. As we all know, 1000MHz is the actual clock of HyperTransport links. So, AMD works at "200" as a base, which is great if you have memory running at 200MHz, 400MHz, 600MHz, 800MHz, 1000MHz, and 1200MHz and so on.

But, if you have "shock-combo" DDR-I/II memory that runs at 333, 533, 667MHz, you could be in a bit of ruff'n'tumble. And just by accident, that DDR-II memory standard has several speed grades, including those at 533 and 667MHz.

AMD says: DDR-II 533 will run at... 255 MHz DDR max!

If you try running the FX-62 review sample with DDR-II 667 memory, you'll get DDR-II memory running at either 600 MHz or 624 MHz. It should of course, run at 666.66 MHz, but I guess folk at AMD aren't big fans of the triple-six number. If you try to run DDR-II at 533 MHz, you'll get something ranging from 243 to 255MHz, or DDR-II "486" and "510".

AMD says: DDR-II 667 will run at 312 MHz and not more!

Of course, this is just a sheer number in read speed, write speed and the latency is another thing. AMD still has a decisive lead in access time to the memory. Although the access time actually rose from 35-50ns to 50-65ns range, the fact still is that Intel cannot get hold of system memory under 90ns.

Best of all, we're not done here - it's not only the DDR-II 533/667 that get weird clocks. Same thing will happen with DDR-II 800 as well. If you try to run a straight comparison with the FX-60, by lowering the multiplier from default 14x200 (2.8GHz) to 13x200 (2.6GHz), you will get the DDR-II memory clocked at 780MHz instead of 800MHz.

You can, however, compare the FX-60 to the FX-62 only by overclocking the FX-60 to 14x200, e.g. having both CPU's running at 2.8GHz.

In the following table, here's an interesting comparison between Intel and AMD's approaches in DDR-II memory. We're talking about available bandwidth in dual-channel mode, e.g. two 64-bit DIMMs working together. In order to get the single-channel results, just divide the number by two.

The Battle of the DDRs

DDR-II 533 on AMD: 8.16 GB/s
DDR-II 533 on INTC: 8.58 GB/s
DDR-II 667 on AMD: 9.98 GB/s
DDR-II 667 on INTC: 10.67 GB/s
DDR-II 800 on FX-62: 12.80 GB/s
DDR-II 800 on "FX-60": 12.48 GB/s
DDR-II 800 on INTC: 12.80 GB/s

µ

Anyone running AM2 has similar problem that The Inquirer encountered?

[metro.cl]

people here told me it was a mobo problem but maybe wasnt mobo at all

here a couple of screenshots:

[***Deimos***]

OMFG

Has everybody completely forgotten that on ALL Athlon64's (regardless of DDR or DDR2), the memory clockspeed is derived from the CPU CLOCKSPEED.

The "symptoms" you describe of non-exact-official memory clockspeeds is precisely what happened with Socket754 and Socket939. Did you expect any different? Did you think this time they would have some magical floating point mulplier with 1/100th Mhz precision? So, once again, this is nothing new, and nothing wrong, and no news at all.

[Flox]

OMFG

Has everybody completely forgotten that on ALL Athlon64's (regardless of DDR or DDR2), the memory clockspeed is derived from the CPU CLOCKSPEED.

The "symptoms" you describe of non-exact-official memory clockspeeds is precisely what happened with Socket754 and Socket939. Did you expect any different? Did you think this time they would have some magical floating point mulplier with 1/100th Mhz precision? So, once again, this is nothing new, and nothing wrong, and no news at all.

Hehe...

Must be the same people that talk of "async" memory speeds on an A64 platform...

[***Deimos***]

Hehe...

Must be the same people that talk of "async" memory speeds on an A64 platform...

Quite frankly, when you use definitions like '1:1' for "sync", with the memory controller on the CPU chip, its difficult to image any other frequency than something like 200 Mhz. Afterall, this kind of "sync" mode is exactly how early Pentium 60 and 66Mhz worked. There was no dividers or ratios at all. FSB, cpu clockrate, memory clockrate and memory controller all running at the same identical speed.

Oh, and I will try to get a step ahead, and already mention there is no "magic" that Intel uses to get around this problem. Their memory controller/arbiter on the chipset works just like the ones for AthlonXP did.

[Victorshen]

Its not strange at all, the DDR2 divider is kinda similar to DDR1 divider

pls refer to the table below and you will get the idea how this divider works



source

[largon]

http://www.xtremesystems.org/forums/...d.php?t=100516
post #17 ->

(edit: the original thread was merged to AM2 faq)

[***Deimos***]

Now if only prefetch or at least anything would be able to make use of this bandwidth. Who knows..perhaps now ATI will make a nice integrated video for Athlon64... maybe even as high as X1300 performance. 12.8GB/s is almost decent. 9600pro - x700 kinda of levels if I'm not mistaken.
Maybe make the 1st 8 pipeline integrated video...

Better still imagine DDR2-1066.. although its doubtful budget minded consumer getting integrated video would get suchmemory.