Why?!!!! Video card w/ only 64M and nVidia just now releasing 128M

[brewerbob]

With Crucial retail (not bulk) being $60 for 256M, why are video cards only at 64M? Why aren't they at 256 or even higher?

 

[DisposableHero]

video ram is much faster than normal pc ram.. the gf3 has 3.8 ns ram.. thats rare and expensive

 

[Noriaki]

Your video RAM runs at 460DDR, your system ram runs at 266DDR.
Which do you think is more expensive?

 

[LukFilm]

Yep, unless we want Voodoo2's with 256MB RAM :Q They ran at 10ns, I think.

 

[thorin]

Given compression and the amount of data needed for today's graphics we don't need more then 64MB (unless you're 3dfx and want to keep multiple copies of everything locally...ala V5 6k).

Thorin

 

[paruhd0x]

how many nanoseconds are in a second? I know there's 1000 miliseconds to make a second....

 

[Killbat]

milli .000
micro .000000
nano .000000000 (one billion nanoseconds in a second)

 

[Hanpan]

it goes
milli 10^-3
micro 10^-6
nano 10^-9
pico 10^-12

 

[brewerbob]

This only leads to another question: Why is the PC ram slower? I thought ram was ram (except pc100, pc133, ddr, etc).So if video ram is so much faster, why not use it exclusively?

 

[br0wn]

<< This only leads to another question: Why is the PC ram slower? I thought ram was ram (except pc100, pc133, ddr, etc).So if video ram is so much faster, why not use it exclusively? >>



The answer is COST and AVAILABILITY.

If COST is not an issue, we can abadon RAM and HARDDRIVE all together and
use CACHE for storage as we know cache is much faster than ram and harddrive.
Heck, we can even only use registers since register can perform
load and store in 1 cpu cycle.

There is a reason by computer engineer to design all these different
types of storage:
register, cache, ram, harddrive, backup tape/dvd/cdrw/zip/floppy.

 

[Sohcan]

<< I thought ram was ram >>

...not quite. All DRAM (SDRAM, DDR SDRAM, RDRAM, etc) has a roughly 20ns page access time...after the address is applied and the CAS line goes high, it takes 20ns for the DRAM to return the data. SDRAM, DDR SDRAM, and RDRAM work by bursting a set amount of data after the initial page access. For example, if the burst length is 8 words, after the page access latency PC100 SDRAM can burst 8 words at one word per clock cycle (10ns in this case). DDR SDRAM can burst two words per clock cycle; once on the rising edge of the clock and once on the falling edge. So with a 3.8ns rating, the DDR SDRAM in the GF3 can burst two words per cycle, with a minimum cycle time of 3.8ns. The DDR SDRAM used in video cards is logically identical to PC2100 DDR SDRAM used in main memory, but the control and access logic is much faster. The speed comes at the expense of lower production yields, so 3.8ns DDR SDRAM is very expensive.

 

[nortexoid]

huh?..the voodoo 5/6 had so much ram cuz each proc. requires its own memory...i don't think they really had a choice...it's faster than their single proc. solution (v4)...

anyway, not to mention that video cards would be relatively big if they used 256mb of the same density chips that are found on single/double sided dimms.

 

[thorin]

&quot;This only leads to another question: Why is the PC ram slower? I thought ram was ram (except pc100, pc133, ddr, etc).So if video ram is so much faster, why not use it exclusively? &quot;

Ok so today we have:

PC RAM (PC133): ~7.5ns
DDR: ~6ns
Video DDR: 3.8ns
L1 Cache: ~ 1 to 1.5ns

Now as we go down the list we get smaller and faster and guess what happens to cost? It goes up.

Thorin

 

[Noriaki]

Because it's alot more expensive, and PCs need a lot more.

A video card really only needs 32-64MB of RAM, but it has to be very fast. Triple Buffered 32bit frame buffers and 32bit Z-buffer uses a lot of bandwidth...

Your PC needs 128-256..maybe more depending on what you do...but it doesn't need to be as fast. PC133 vs DDR266 really isn't a big difference.

They have 15000rpm 36Gig SCSI drives now...why do we still use IDE?

 

[Mookow]

I think the higher the speed of the RAM (PC66 vs PC133 for example), the shorter the trace lengths must be. If so, it would be really hard to engineer mobos for 500MHz DDR RAM system memory.

 

[AdamK47]

Not only does 3.8ns DDR SDRAM cost more, implementing more RAM on a graphics board would cost considerably more too. There would be more traces, capacitors, resistors, PCB, ect.

 

[brewerbob]

Now as we go down the list we get smaller and faster and guess what happens to cost? It goes up.[/i] >>

[/i] >>



Are there not people and corporations that say damn the cost, I want/need the fastest? The CPU, video ram and regular ram are all zero's and ones. The speeds should all be the same, the spped of light (or actually the speed of electromagntic energy using copper as a medium). This is an over-simplification but every thing in a computer is a logic gate.


It would come down to the age old question...if the computer is capable the software will exploit it. Who really needs a 1.7GHz for excel.

 

[Sohcan]

<< The CPU, video ram andregular ram are all zero's and ones. The speeds should all be the same, the spped of light (or actually the speed ofelectromagntic energy using copper as a medium). This is an over-simplification but every thing in a computer is alogic gate. >>

You're right, that's way too simplified. The absolute speed may be determined by the speed of the propogation of electricity in the wiring, but transistors have a delay time (and the length of signal traces is increasingly becoming more important), so the structure of the memory type has a huge influence on its speed.

Registers are used in the register file, typically about 1KB in size. They store each bit in a flip-flop (usually D-type positive edge-triggered), which in turn is comprised of two latches and an inverter. Since the flip-flop and the large amount of control logic and multiplexors for each port of the register file require many transistors, the cost/register is huge.

SRAM, used in L1 and L2 caches, uses complementory inverters to store a bit, and use six transistors per cell. They don't have high densities, so that's part of the reason the Alpha EV68 w/4MB of L2 cache costs $8000.

DRAM, used in main memory, uses a tiny capacitor to store each bit, and only requres 1 transistor per cell. Thus it can be produced in high-density, and costs much less than other memory technologies. But because each read operation wipes the charge stored in the capacitor, and because the charge must be periodically refreshed, DRAM is much slower than registers or SRAM.

As you go from registers -> SRAM -> DRAM, the the capacity goes up and cost/byte goes down, but so does the speed. Registers have access times of around .5ns, SRAM around 10ns, and DRAM around 100ns.

Also, it simply is not currently feasible to use high-speed DRAM for main memory. The speed of the memory FSB is severly limited by it's length (around 1 foot for SDRAM), and clock-skew becomes a huge issue at high clock rates. That's why raw FSB clock rates (discounting DDR signalling) has only doubled from 66MHz to 133MHz in the past seven years since the introduction of the Pentium 66. Video cards, on the other hand, have much shorter trace lengths, so it's (relatively) easier to use high speed memory.

There are corporations that say damn the cost, but they're not going to go with x86 processors...that's the market for Alpha, SPARQ, PA-RISC, MIPS, Power3, and Itanium, where is it feasible to have large amounts of fast SRAM caches.

Trust me on this stuff, I've taken a bunch of logic design classes. 😉

 

[nortexoid]

noriaki, i wouldn't consider pc2100 to be FASTER tahn pc133...they're the same speed (ns rating)...

i'm sure u know how ddr works....

i'm nto sure faster memory necessary costs more to manufacture (in pure manufacturign costs), though sicne it is faster, u pay the premium...sad.