Memory performance, number & size of chips ?

Snap

n00b
Joined
Feb 28, 2009
Messages
12
Hello together

Today i bought two video cards (Asus 9800 GT, 1024 MB DDR3). One card is for my Mom's computer and the other will be needed to build a pc for a friend. However... ;)

I've unpacked both cards and looked on them. I've noticed that one card has 8 memory chips on both sides (totaly 16) and the other one has only 8 chips on the upper side. I've checked the description on the chips and found out, that one card has 16 hynix 512Mbit (H5RS5223CFR-N0C) chips and the other one has 8 hynix 1Gbit (H5RS1H23MFR-N0C) chips on it. According to the hynix datasheets, the chips have the same clock- and speedrates. Both video cards have the same pcb-layout (identical layout- and revisionnumber).

My question now is, does this difference affect the performance ? Is the card with 8 1Gbit chips faster because of the double capacity of each chip ? Or is the fully loaded card with 16 512Mbit chips faster, because the card uses more data lines to the chips (on both pcb sides), so the data can be stored / accessed faster. Or is there no difference ? From the power consumption and thermal aspects, i think the card with less bigger chips is more economical and generates less heat. But maybe the card with more smaller chips has more oc potential ?

Thank's for your opinions, Snap.

P.S. Sorry, i'm writing here in bad english (i'm from switzerland), but i think this is one of (maybe) the best forum about hardware. I've learned a lot of things, by watching Kyle's reviews and read the postings here.
 
Without the model numbers on the cards we can't tell you much if anything. I'd hazard a guess that the card with more memory chips has double the bus speed, doubling it's memory bandwidth. Someone, correct me if I'm wrong.
 
Without the model numbers on the cards we can't tell you much if anything. I'd hazard a guess that the card with more memory chips has double the bus speed, doubling it's memory bandwidth. Someone, correct me if I'm wrong.

I wont say you're wrong but you're probably not right. After all 6 sticks of memory in a triple channel X58 motherboard isn't any faster than 3 sticks.

I suspect they are identical in speed the only difference is the amount of physical chips. I would take the lesser amount of chips.
 
True. But I was going by the bus width (64bit vs 128 vs 256bit etc...) these do make a difference in performance.
 
The card is this one :

http://www.asus.com/product.aspx?P_ID=trOKSJujEvAyNXRb

Large pictures :

http://www.brack.ch/aspx/abdl.aspx?W=114502374
http://www.brack.ch/aspx/abdl.aspx?W=114502384


My both versions have the same pcb- and revisionnumbers. Also all visible components on it are the same, except the memory chips. One card has both sides filled with 16 (8 on each side) 512Mbit chips (hynix H5RS5223CFR-N0C). The other card has 8 1Gbit (hynix H5RS1H23MFR-N0C) chips only on the upper side, on the lower side the empty places, where the chips can be fitted, are visible.

The card on the pictures from Asus has the 512Mbit (Samsung K4J52324QH-HJ1A) chips. Because there is no picture from the lower side, it's not possible to see if it's the 512 MB or the 1 GB Version - or in other words if there are also chips on the lower side.

Theoretically the card can be fitted with 16 1Gbit chips to reach 2 GB memory. In a review from the 8800 GT i've read that the manufacturers just used the first upcomming 1Gbit chips from Qimonda to build cards with 1 GB memory. They wrote it was just that simple to use bigger chips - no redesign of the pcb or another memory controller was required. Same goes on the GTX285 - some manufacturers offered 2 GB versions. They just put larger memory chips on nVidia's reference layout. But in this case the number of the chips on the cards were not affected. Also the bigger chips were running at the same speed, so the memory has the same performance but the capacity were doubled.

I'm just interessted about the technical details. How does work the memory-controller, -bus, -addressing, -bandwidth, etc.

Thank you, Snap.
 
True. But I was going by the bus width (64bit vs 128 vs 256bit etc...) these do make a difference in performance.

9800GT's all run with a 256bit bus.

OP: both cards will run at the same speed. There is a possibility one MAY OC better than the other, but this cannot be determined until you actually try. The type of memory used is not a clear indicator of OC potential, due to the binning process of chips. The card with more chips may produce less overall heat IF they are properly heatsinc'ed, but that is not guaranteed.

TL;DR - Probably little to no difference.
 
The performance for 16x512 and 8x1024 will be nearly identical.

Assuming your card has a 256-bit memory bus, then the target number of chips is eight. Each chip is 32 bits wide, and thus you need a minimum of 8 chips to fully fill a 256-bit bus (8 x 32 = 256).

But you can add more layers of chips than that. In the 16x512 configuration, twice as many chips are on the same bus. This means you get more reflections on the shared bus, and you have to use slower timings on the memory (much like you see with 4 DIMMS in a 2-channel motherboard). But the memory clock will be the same, so overall bandwidth will be almost unaffected.

So, performance should be similar.
 
@LeninGHOLA

Why you think, the card with more chips may produce less overall heat ?

Maybe the bigger chips produces more heat, because they have more capacity, but perhaps they are produced in a smaller nanometer process, so it's difficult to say. And does the heatsink not just moves the heat away ? I think it does not affect the formation of the heat ? The chips on the lower side of the card are naked (without a heatsink), same like the chips on the upper side, but they will be cooled by the gpu fan.

@defaultuser

It's not written in the specs on the asus-site, but the card has a 256-bit memory interface (according to informations from other sites (i.e. http://www.overclock.co.uk/product/ASUS-1GB-GEFORCE-9800GT-VGA-/-HDMI-/-DVI-I_24859.html)

Can you explain me, why the target number for a 256bit-interface is 8 chips (32bits / chip) ? That would mean, that the both-side-filled card with totally 16 chips has only a 16bit per chip connection ? Can be generally said, that less of chips with more bits connected to a bus are better, because they generates less 'reflections' ?

Thank you very much, Snap.
 
@LeninGHOLA

Why you think, the card with more chips may produce less overall heat ?

Maybe the bigger chips produces more heat, because they have more capacity, but perhaps they are produced in a smaller nanometer process, so it's difficult to say. And does the heatsink not just moves the heat away ? I think it does not affect the formation of the heat ? The chips on the lower side of the card are naked (without a heatsink), same like the chips on the upper side, but they will be cooled by the gpu fan.

I meant, if there were ramsinks, there would be more airflow per unit. It looks as though there are none, so for my statement to be true, you would have to add them aftermarket.
 
I've read some technical informations about the internal functions of the 1Gbit chip :

"The chip is programmable into two different configurations. In the default mode the architecture is organized as two 512Mbit memories of 8 banks, each (two CS mode). In an alternate configuration, it behaves as a conventional, 8-bank 1Gbit DRAM (one CS mode)."

Does this mean, if the chip is working in the default mode, it's visible as two 512Mbit chips on the Bus ? So there would be no advantages (like the described reflections) over the 16 x 512mbit chip configuration ?

And what about all the data lines to the empty places on the lower side of the card ? Has this no impact ? Has each chip its own fixed datalines to the 256bit memory-interface, or does they all share the available 256bit on the bus - independent from the number of chips ?

Hum, the things are too complicated for me...
 
I think this is being way too overthought. The two cards probably differ in performance and heat generation by less than 1%.
 
I think this is being way too overthought. The two cards probably differ in performance and heat generation by less than 1%.

Yeah, i think so too. Maybe i run some Benchmarks when i build the new machine with both type of cards.

It's more the technical stuff which is interessting for me. Which chip / interface configurations are possible. How ist the bus designed and how does it work ? Which solutions are clean and which are tinkered ? Any advantages, even if they only existing in theory...
 
Look at it this way - if there was in any way a performance difference between them, you would have paid more for one than the other. Being from the same company, thats how business works. if they can make one card seem better than another, they will do it and sell it for just a little more.

There would be absolutely no difference as far as you could tell between the two because the specifications are the same. And I doubt there is any difference anywhere other than the physical difference. It likely just saves them money not having to mount two times as many chips.
 
@LeninGHOLA
Can you explain me, why the target number for a 256bit-interface is 8 chips (32bits / chip) ?

Thank you very much, Snap.

There are three aspects to describe a memory chip's configuration (this is just in-general):

1. data bus width
2. address bus width
3. access time

data bus width * address bus width = capacity of the memory

Thus, the 512Mbit chip you mentioned above is 16Mbit x 32 (address x data)

You access data in memory by providing it with a specific address. The data is output on the data bus, which in this case would be 32 bits of data.

The BANDWIDTH of a memory chip (Gigabytes per second) is limited by the access time (speed), and by the width of the data bus. A wider data bus means the memory can transfer more data in a given period of time.

For reasons of easy mass production, memory is not typically made with data buses wider than 32 bits. This means that, if you want to get more bandwidth, and you can't increase your memory clock speed, then you have to stack multiple memory chips in-parallel.

This is how the 8800 GT works, with it's 256-bit bus. When using 8 1Gbit chips, each memory chip takes up it's own 32-bit wide slot on the 256-bit bus, and the total bandwidth is eight times that of a single chip.

The reason you need multiple chips on the 8800GT is because a single memory chip with 32-bit data bus would not be enough to keep the card fed (not enough bandwidth).

That would mean, that the both-side-filled card with totally 16 chips has only a 16bit per chip connection ?

That's a simple way of thinking about it. But the reality is this: the 16 chips are still 32-bits wide. How you do this is, you have 8 pairs of chips, and they will be connected to the same data bus output. Each pair is now stacked on it's own 32-bit section of the 256-bit bus. To avoid contention (both chips trying to output at the same time), you use known tricks with the address ranges to enable only one chip at a time.

Can be generally said, that less of chips with more bits connected to a bus are better, because they generates less 'reflections' ?

On a perfect point-to-point connection, you don't get reflections from data transmitted.

On a shared bus, you get reflections in data transmitted from one point to another along the bus. This is caused by multiple "taps" (connections) to the shared bus.

I can't think of a really good analogy to describe it, but here goes anyway:

Consider the point-to-point case: a shower that feeds from the head right into a huge drain. The water flows freely from head to drain. Now, stick your body into that shower stream. Some of the water still reaches the drain directly, but much of it bounces off your body in all directions - this is much like a reflection caused by a "tap" on a shared bus.

In a bus with multiple taps, it's like a bunch of people trying to catch that shower stream. Not only is the strength of the original signal reduced, but you can also misinterpret one of the reflections as correct data (this is bad).
 
the 16 chips are still 32-bits wide. How you do this is, you have 8 pairs of chips, and they will be connected to the same data bus output. Each pair is now stacked on it's own 32-bit section of the 256-bit bus. To avoid contention (both chips trying to output at the same time), you use known tricks with the address ranges to enable only one chip at a time.

Many thanks for your explanation.

In the technical description of the 1Gbit chip is mentioned :

"The chip is programmable into two different configurations. In the default mode the architecture is organized as two 512Mbit memories of 8 banks, each (two CS mode). In an alternate configuration, it behaves as a conventional, 8-bank 1Gbit DRAM (one CS mode)."

Does this mean, in default mode the chip is working in the same way like a pair of 512Mbit chips ? So it makes no difference if there are 8 x 1Gbit chips or 16 x 512MBit chips on the bus ?

If so, why the chips have this function ? Maybe for the reason that 8 chips can fill a 512bit memory-interface (8 chips in default mode (two CS) for 1 GB Ram or 16 chips in alternate mode (one CS) for 2 GB Ram) ?
 
Many thanks for your explanation.

In the technical description of the 1Gbit chip is mentioned :

"The chip is programmable into two different configurations. In the default mode the architecture is organized as two 512Mbit memories of 8 banks, each (two CS mode). In an alternate configuration, it behaves as a conventional, 8-bank 1Gbit DRAM (one CS mode)."

CS refers to Chip Select. This is the electronic signal you set high to enable the chip. When I mentioned above for the 16-chip configuration, "use known tricks with the address ranges to enable only one chip at a time," I was referring to the Chip Select.

This just gives the designer two operational modes to use the chips. The aggregate bandwidth is still the same, but you can treat the chip as two separate 512Mbitx16 devices on one package.

Does this mean, in default mode the chip is working in the same way like a pair of 512Mbit chips ? So it makes no difference if there are 8 x 1Gbit chips or 16 x 512MBit chips on the bus ?

Yeah, but there's no reason to use this mode unless your memory controller has very finely-grained access (think 16 bits at a time).

Your 8800 GT doesn't just access memory one 256-bit chunk at one time. Instead, it has a crossbar memory controller that can access 4 64-bit chunks of memory independently. This improves memory access efficiency because, if you access only large blocks of data, you have to throw away the portions of that huge block that you don't need.

Since the data types in GPUs today are typically 64-bit, having four memory controllers that access 64 bits each is a good setup. You don't want to cut that down to 16-bits, so there's really no reason to run the memory in that mode.

If so, why the chips have this function ? Maybe for the reason that 8 chips can fill a 512bit memory-interface (8 chips in default mode (two CS) for 1 GB Ram or 16 chips in alternate mode (one CS) for 2 GB Ram) ?

I can think of one reason why you might want the chip to behave as two 512Mbitx16 chips instead of one 1Gbitx32:

In Smartphones today, you want to pack as much into the device as you can, but you are limited by PCB space. Thus, a high-density memory device like the 1Gbit chip is a good idea.

But what if you need memory for two devices? You have three options:

1. add a second memory chip to the PCB (not enough space for this).
2. add control logic to map the memory to multiple devices (also adds complexity).
3. use the 512 x16 x 2 operating mode, and have each of your two devices use a different Chip Select.


#3 requires no more complexity on the PCB, and although you can't customize how much memory goes to which device like #2, it's a quick and dirty way to get you up and running.
 
@defaultluser

Thank you. I begin to understand. You have a great knowledge. I love it to talk shop. :)

It's questionable why the default mode of the chip is the two CS mode. I hope the chips on my card are running in the alternate one CS mode, so i have an advantage versus the non optimal 16 chip (full filled) version.

Or maybe the manufacturer purposely runs the Gbit chips in the two CS mode, so there is no advantage versus the 16 Chip version. But then the 512 MB version with 8 x 512Mbit chips is in the advantageous. :rolleyes:

So, when i'm on the Dating Game, which model should i prefer ? The fully loaded or the naked below ?
 
On the upper side, the places of the 8 chips are marked with M1A - M8A. The other 8 chips on the bottom side are marked as M1B - M8B.

Maybe this means the fully loaded configuration works in groups (group A and B) ?

So, if the memory interface can only transfer 256 bits at once, then maybe these two groups of 8 chips will be handled alternately (one clock cycle for group A and one for group B) ?

The only disadvantage here would be possibly in timing, when the next group of the cycle interval should be accessed, then a clock cycle would be lost. Does it work like this ?
 
Back
Top