Why is memory on a graphics card so much faster than PC4400?

[TStep]

How can DDR memory on say an Nvidia 5950 clock at over 500mhz (1000mhz effective) and standard systems DDR memory struggle to reach 275mhz (550mhz effective). Is there a difference in how it operates that I am not understanding?? Are there latency issues??

I cannot believe its a cost issue as nuts like anyone asking a question like this will shell out big bucks for a video card with 256 megs of the high speed memory and I'm sure would do the same for systems memory if available.

 

[Matthias99]

The problem is that video card memory is both a) expensive, relative to regular DDR, and b) hardwired onto a custom PCB with a custom memory controller integrated into the GPU. It IS a cost issue; it's simply not feasable to produce general-purpose chipsets are both very expensive and non-upgradeable. In a limited situation like a video card the margins are there for it.

 

[TStep]

Well DDR II is going to debut at $455 for 256megs. So apparently there is a market at a high price.

See here

Performance will be on par with current standard memory. Now I understand its new technology, so higher pricing at the onset, but two Nvidia 128 meg 5900s are less than $455 and they incorporate a total of 256 megs of very high speed DDR along with the rest of the card. So say the cost of the ram is half the cost of the cards, or $200 for 256 megs. People are paying a premium price right now for Mushkin PC3200, 2 sticks of 256 meg for $300. An additional $100 is not that far of a stretch considering the price difference between a P43.2EE and a regular P43.2 is threefold and this is on the market.

I have to believe it is something else. Motherboard design constraints?? Too high of latencies?? Inability of current chipsets to handle memory at this speed?? Need for memory to be directly connected to the cpu, similar to Opteron??

 

[Matthias99]

*I* wouldn't buy DDR2 at those prices. Maybe not even at half those prices, unless it starts making SERIOUS performance enhancements. Intel's going to fall flat on their face if they demand you pay 4x as much for memory to buy a new motherboard from them (of course, they did the same thing with RAMBUS and the P4... apparently they're not learning). Very, VERY few people besides hardcore overclockers would pay $300+ for 512MB of memory. Let's say theoretically you could build a motherboard that can run 256MB of 500Mhz memory and a 3.0Ghz P4 processor, but you have to buy it with a fixed, non-upgradable memory and CPU configuration, AND it costs twice as much as a 'normal' MB + CPU + high-end RAM. How many of these do you think they would sell? Plus, they would have to put in all the design work just for a high-end product; it's not like there's a mainstream market for something like this.

I have to believe it is something else. Motherboard design constraints?? Too high of latencies?? Inability of current chipsets to handle memory at this speed?? Need for memory to be directly connected to the cpu, similar to Opteron??

Well, it's sort of all those things, but they all affect each other. It's mostly latency issues stemming from the fact that a more generic memory controller than can take differing sizes and types of memory is slower. Also, the chips are physically further away from the CPU/controller, and are interfaced through add-on slots, rather than being hardwired into the PCB. This lowers your signal/noise ratios and reduces the maximum speed you can run at. Memory in P4 systems is connected to the FSB, and there are limits to how fast you can run *that* as well. Connecting the memory right to the CPU (as the A64/Opteron does) will help, but it's still not as good as having it physically mounted to the same circuit board just an inch or two from the CPU and building in a customized memory controller. Also, it's a LOT easier to run small memory chips fast than large ones. You might be able to coax 32MB DRAM chips to run at 400-500Mhz (and use four of them in parallel for a 128MB video card), but you're not going to get 256MB or 512MB DIMMs anywhere near that speed (at least not today).

You have to understand video cards are pushing the limits of how fast you can run memory chips -- a 500Mhz clock rate means a cycle on the RAM is only 2 nanoseconds! The fastest chip-to-chip interfaces are only a few Ghz, and those are for serial lines between simple circuits, not 256-bit parallel connections to DRAM cells.

 

[TStep]

Thanks for the reply

 

[ReiAyanami]

nVidia and ATI have mass purchasing power, vs. a very small niche market.

the video ram may simply be the top of the crop from micron/hynix ect...

in other words, its not made high grade but rather is incidently high grade the same way AMD churns out a bunch of chips then hopes to yield more 3000XP+ than 2500XP+

 

[Peter]

All different environments.

Graphics cards have no RAM slots, the few chips connect through short traces directly on the board, and thanks to the wide busses used there is just one single RAM chip load per wire.

On a mainboard, you have a long trace to the DIMM slots, you have the slots and edge connectors, and you have to design for at least four chip loads per connection since you have at least two DIMMs (four DIMM sides) per RAM channel. All those things are doing no good to signal integrity.