Discrepancy in Creative's MX/DDR bandwidth spec...

[Leo V]

Creative advertises their GeForce2 MX/DDR as having 2.8GB/sec bandwidth. This requires 350MHz (5.7ns) DDR memory to accomplish, and some website claimed they actually used 7ns, 286MHz memory--yielding a measly 2.3GB/sec.

I've emailed Creative the following regarding the MX:

"Your webpage advertises its memory bandwidth as 2.8GB/sec. However, in order to achieve that specified bandwidth, the MX would require 350MHz DDR memory. Does your MX in fact use 350MHz memory? If not, then how fast is its memory, and how would you justify the 2.8GB/sec figure then?"

I'm very interested to know, because you can buy the sucker for $89 (after coupon) from Buy.com.

 

[Howard]

If I remember correcty, all MXs come with SDR memory?

 

[Goi]

LeoV, if I remember correctly, the GF2MX has a 175MHz core and memory speed. Hence the 2.8GB/s figure. There's no discrepency

 

[Leo V]

Guys, you're confusing the new Creative MX/DDR with all existing MX/SDR cards. The old ones have 128-bit SDR memory, and the new Creative MX has 64-bit DDR memory. This is why it requires 2x175 = 350MHz DDR memory to accomplish 2.8GB/sec. Let's see what Creative has to say...

 

[ruckb]

Hi LeoV,

I do not know the card, but why is creative going back from a 128bit Bus to a 64bit DDR bus ?
The nice thing about DDR is, that the infrastructure is nearly the same, so you can use your existing boards with a few modifications (termination, ...).
With a half databus wide they are just loosing the performance advantage of DDR.

ruckb

 

[Leo V]

ruckb, using 64-bit DDR is apparently cheaper than using 128-bit SDR, because fewer traces are needed by the 64-bit interface. (And DDR memory costs about the same as SDR, while using less voltage).

Of course, 128-bit DDR MX would rock the boat--but it would easily compete with the high-end GTS cards! That's why new MX cards are 64-bit DDR models, cheaper/lower-voltage than 128-bit SDR and about equally fast. $89 on Buy.com is a bargain for a GeForce2 MX, expecially because it has 32MB.

 

[shabby]

/me looks into his magical 8-ball

Creative Labs' response to Leo V "There is no such thing as 350mhz ddr memory, fastest memory is 230mhz ddr used in the gts2 ultra"

Think you should of stated 175mhz ddr memory(which is basically 350mhz) and not 350mhz ddr memory. If cl wants to be cocky, they will give you that response

 

[Leo V]

shabby, I signed the letter as "longtime Creative customer", so they'll just have to watch their manners. I assume they want more business, as most companies do.

 

[Goi]

OK, I didn't know you were talking about the new DDR models. These will definitely perform much poorer than the current SDR MXs, much like the G450 vs G400.
Anyway, you're right in that it needs DDR SDRAM running at 175MHz(X2) to achieve the same 2.8GB/s memory bandwidth as its SDR counterparts. Perhaps they do use that? Where did you see the claim of 7ns DDR SDRAM for the new MXs?

 

[Leo V]

Goi, here is a new link I found, which confims my fears.

The current 166MHz SDR MX boards achieve ~2.7GB/sec, and the 183MHz SDR boards ~2.9GB/sec. By common math, Creative must change their 143MHz DDR bandwidth spec to 2.3GB/sec (which really sucks).

 

[xtreme2k]

Leo V

you are correct pointing out a 143MHz 64bit DDR SDRAM does give a 2.3GB/s memory bandwidth
you can verify it by doing this calculation

143 * 10^6 * 2 * 64 / 8 / 1024^3 = 2.13GB/s actually even less than 2.3

anyway back to my point
EVEN IF the DDR has a theoritical bandiwidth of say 2.9G (183MHz 64bit DDR)
it will still be slower than a equvalent 183MHz 128bit SDR
as far as i know
DDR arent 100% effective, only about 70-80% (if i remember correctly)

personally i wont bother with creative's MX this time

 

[andylawcc]

Yucky: Creative's the only exception so far.

ixbt-labs just reviewed it btw.

 

[Leo V]

xtreme2k, I used base-10 gigabytes only for comparison to similar such measurements.

Regarding the relative efficiency of SDR vs. DDR, you're quite right (I actually forgot about that!) Due to CAS latency, a GeForce using DDR memory has about 80% effectiveness, and one with SDR memory has about 88.89% effectiveness (I've gone through the calcs before...) This leaves us the following "effective" base2 measurements:

GeForce2 Ultra/DDR230/128-bit...5.48GB/sec effective
GeForce2 GTS/DDR166/128-bit....3.97GB/sec effective
GeForce1 DDR/DDR150/128-bit....3.58GB/sec effective
GeForce2 MX/SDR183/128-bit....2.43GB/sec effective
GeForce2 MX/SDR166/128-bit....2.21GB/sec effective
GeForce2 MX/DDR143/64-bit.....1.71GB/sec effective*

* That's a whole 1GB/sec less than Creative advertises!! :|

 

[Leo V]

Worse yet, subtract the bandwidth used by the RAMDAC when running at 1024x768x32 @120Hz (.35GB/sec), and only this remains:

GeForce2 Ultra/DDR230/128-bit...5.13GB/sec for rendering
GeForce2 GTS/DDR166/128-bit....3.62GB/sec for rendering
GeForce1 DDR/DDR150/128-bit....3.23GB/sec for rendering
GeForce2 MX/SDR183/128-bit....2.08GB/sec for rendering
GeForce2 MX/SDR166/128-bit....1.86GB/sec for rendering
GeForce2 MX/DDR143/64-bit.....1.36GB/sec for rendering*

* This means the GF2/MX/DDR has a quarter of the Ultra's fillrate at this res!! It's only 2/3 the speed of a normal MX SDR/183! :| This makes it slower than even the original Prophet SE, which was a GeForce SDR with 143MHz memory!

OUCH

 

[xtreme2k]

hehehe told ya not to be bothered with the G2/MX 64bit DDR

anyway all these theories needs to be verified by real world benchmark
i am yet to see one that compares the follow which i will find interesting

G2/MX 64bit DDR at 143MHz vs G2/MX 128bit SDR at 143 (this is to prove how effective DDR/SDR is)
G2/GTS 128bit DDR at 300Mhz core at 120MHz vs Geforce256 DDR at 300 core at 120MHz(this is to prove how good the G2 is actually performing clock to clock)

 

[ruckb]

Xtreme2K:

Your calculation is right, but the problem is, that you will never see this numbers in the real world.

One reason is the calculation.
Your equation:
143 * 10^6 * 2 * 64 / 8 / 1024^3 = 2.13GB/s
is just
143 * 2 * 8 = 2.288 GB/s
for a base10 system.

For marketing there are two reasons to use the second one. It is easy to understand (even for a marketing guy ;-) ) and the resulting number is bigger.

Your equation is the exact one, but it will be never used. Just take a look at the confusion about the HDD sizes.


LEO V:
How have you calculated the effectiveness for SDR and DDR ?

As you mentioned, you will never get the full bandwith because CL, Read turnaround time, .., but I have no idea, how you calculated this numbers (80% and 88,89%) you gave just before.

ruckb

 

[xtreme2k]

ruckb

i understand about my calculation and the base10 one

but since we are kinda working out a specific memory bandwidth technically
that was why i used the 1024, since that is the true bandwidth

i also understand due to marketing BS they have to use the 1000 one so that the number looks bigger.

either way the difference are neglectable
2.13 or 2.288G

 

[Goi]

Either way, the DDR version is gonna suck.