4870 of memory bandwidth naming?

[gulogulo96]

Having a very limited knowledge (and hoping to improve it) I'm a bit confused about how different companies define their memory for cards.

(similar to DDR/DDR2/DDR3 SDRAM?)

GDDR/GDDR2 half the I/O memory bus clock of GDDR3/GDDR4

GDDR/GDDR2 a fourth the I/O memory bus clock of GDDR5

therefore if we memory clock (lets say 900MHz) the total memory frequency will be the following:

GDDR/GDDR2: 900MHz
GDDR3/GDDR4: 900MHz x2 = 1800 MHz
GDDR5: 900MHz x4= 3600 MHz

Why are some manufactures using 1800 MHz when specifying the 4870 then? example MSI http://global.msi.com.tw/index.php?func=proddesc&prod_no=1502&maincat_no=130

 

[mactronix]

I think you are getting confused by how they market the cards DDR stands for Double Data Rate. This means that a card that has a speed of 900Mhz is in fact running 1800 effective.

This might help http://www.extremetech.com/article2/0,2845,2309870,00.asp

Mactronix

 

[gulogulo96]

thanks for reply and link it was very useful.

Just for clarification then, Is a standard 4870 running at 3600Mhz (effective DDR speed)?

 

[mactronix]

Yes that would be correct, just for clarification purposes. http://www.anandtech.com/video/showdoc.aspx?i=3341

Mactronix

 

[gulogulo96]

thanks again for the reply.

Still don't get the naming though (and being nit picky). From the link you just sent the memory clock for the 4870 is 900MHz (3600MHz data rate) GDDR5. So this goes back to my original question why are manufactures using 1800 MHz when specifying the memory clock for the 4870 then when it should be either 900MHz or 3600MHz?

 

[nottheking]

So this goes back to my original question why are manufactures using 1800 MHz when specifying the memory clock for the 4870 then when it should be either 900MHz or 3600MHz?

Not quite... Going from memory here, this is roughly what each type of memory runs at,, for three different types of things sent: commands from the memory controller, (i.e, commands to read/write) addresses for the memory operations to take place, and the actual data transfer for the memory operation. For this example, the clock speeds may not match what such memories are available in (or even capable of) in real life, but that is purely so that you can see how they relate in their differing clock rates for different parts; all of these are clocked so that their effective DATA transfer rate is identical. I also may be a bit off here, but this is the general idea... (someone WILL correct me where I'm wrong)

DDR
■Memory commands - 1000 MHz, double-pumped for 2000MHz effective.
■Memory addresses - 1000 MHz, double-pumped for 2000MHz effective.
■Memory data - 1000 MHz, double-pumped for 2000MHz effective.
DDR2
■Memory commands - 1000 MHz, double-pumped for 2000MHz effective.
■Memory addresses - 1000 MHz, double-pumped for 2000MHz effective.
■Memory data (internal) - 500 MHz, double-pumped for 1000MHz effective, double bit width.
■Memory data (external) - 1000 MHz, double-pumped for 2000MHz effective.
GDDR3
■(same as DDR2)
GDDR4
■Memory commands - 500 MHz, double-pumped for 1000MHz effective, double bit width.
■Memory addresses - 1000 MHz, double-pumped for 2000MHz effective.
■Memory data (internal) - 500 MHz, double-pumped for 1000MHz effective, double bit width.
■Memory data (external) - 1000 MHz, double-pumped for 2000MHz effective.
GDDR5
■Memory commands - 500 MHz, 500MHz effective, quadruple bit width. (not double-pumped so as to improve reliability)
■Memory addresses - 500 MHz, double-pumped for 1000MHz effective, double bit width.
■Memory data (internal) - 500 MHz, double-pumped for 1000MHz effective, double bit width.
■Memory data (external) - 1000 MHz, double-pumped for 2000MHz effective.

As you can see, as each type progresses, they find a way to scale down the clock speed of other parts of the RAM module while keeping the clock rate for the actual external data the same, allowing it to keep the same bandwidth without reducing its own number of pins... but by cutting the speed on the others, it allows the whole thing to be clocked much higher; the speeds I mention above are, of course, completely unrealistic for DDR and DDR2; the highest reasonable speeds for them tend to top out around 300MHz and 700MHz or so. (600MHz and 1400MHz effective) And because of their clock rate cuts, GDDR4 and GDDR5 can reach even higher external data clock rates; with respective speeds of 1400MHz (2800MHz effective) and 2250MHz (4500MHz) reportedly being available for volume production.

 

[mactronix]

thanks again for the reply.

Still don't get the naming though (and being nit picky). From the link you just sent the memory clock for the 4870 is 900MHz (3600MHz data rate) GDDR5. So this goes back to my original question why are manufactures using 1800 MHz when specifying the memory clock for the 4870 then when it should be either 900MHz or 3600MHz?

I can see where you are comng from but hopefully what nottheking has posted clears it up a bit for you. As its the bandwidth thats increased in GDDR5 and mot the cycle count then it is correct to report the speed as either 900 which is then physically double pumped to the effective(if we were talking about GDDR3) of 1800. the actual effective out put is double that thanks to the higher bandwidth. Its not tecnically correct but you could look at GDDR5 as sort of quad pumped.

Mactronix

 

[nottheking]

Its not tecnically correct but you could look at GDDR5 as sort of quad pumped.

One could suppose that... But yeah, VERY technically incorrect.

Basically, one could properly describe GDDR5 as having the different pins running at different clock speeds. The data pins on the 4870's GDDR5 do, in fact, run at an actual 1800MHz, double-pumped for an effective 3600MHz. However, it's listed as 900MHz because both the command and address pins run at an actual 900MHz; the command pins are not double-pumped at all, though the address pins are double-pumped to run at an effective 1800MHz.

 

[gulogulo96]

Right! Thanks a lot for your replies it clears that question up