GTX 460 and RAM bottleneck

[HellsMaster]

I plan on buying a GTX 460 but apparently my RAM could cause bottleneck due ot it being low latency, so my brother says. My RAM is 4gb 600mhz. would it caused bottlenecks? shoudl i buy higher latency RAM such as 800mhz or 1024?

[Nobody09]

I plan on buying a GTX 460 but apparently my RAM could cause bottleneck due ot it being low latency, so my brother says. My RAM is 4gb 600mhz. would it caused bottlenecks? shoudl i buy higher latency RAM such as 800mhz or 1024?

 

Video cards are first order bus masters so they take direct control over the memory system.

There's no real issue here as they tend to do rather large transfers and not tweak and peck at the memory subsystem. The clock speed of the ram affects the latency but that doesn't matter there's no massive difference between lowly 400mhz ddr and 2000mhz ddr3. The simply add more and more wait states and report a higher theoretical throughput which is fine for very small hunt and peck systems.

Most of the ram speed increasing is to help support mutl core chips. When you run multi core chips you have memory addressing problems as one thread can want to access the same memory locations. So it all just pauses and locks up waiting for access. The twitchy ram does much better at this. It's not a real problem unless you start getting past 4 cores but people want 8 and 16 core systems for servers which are of little benefit to the consumer. But the consumer gets dragged along for the ride to help pay for it.

Memory requires a FULL 3 shrinks to get any speed advantage. For every 15 percent speed increase you get from cpu you only get a 5 percent increase in ram. But you don't get it all at once you get it after the 3rd shrink. So they keep upping the clock speeds to try to squeeze those tiny little percents out on a wait state fraction. IE if your ram runs 3 wait states at 400 mhz running it at 800 mhz gives you 6 wait states. But since the ram has had a process shrink or two it's actually able to pull 5 wait states at those clocks.

 

Depending on your system you can simply drop your divider and up your bus speed to put your ram on edge. My cpu is 2500mhz at 12.5 divider. with 200 mhz bus speed giving me a 1000mhz hypertransport link speed. I set the computer up at 7 divider and 300 mhz front side bus speed giving me 1500mhz ht link speed. This puts my ram from 800mhz to 840 mhz because it's on a 1:5 clock to cpu. The divider I chose puts cpu at 2100mhz and my ram is on it's lowest divider of 1 clock for every 5 clocks of cpu. The actual clocks are 400mhz and 440 mhz. But since it's ddr2 they lie and say it's 800mhz and 880 mhz "theoretical". My ram max stable is 480mhz real at a 8 to 1 bus:cpu divider for 2400 mhz. Actually it's bleeding max is around 488 mhz where it will start getting flakey.

 

My cpu retires about 14 instructions a cycle on 2 cores at 2500mhz. That's too many as pentium 4's would retire upwards of 4 instructions per cycle at 2.8 to 3.6 mhz and those will run plenty of modern games. I set mine up to retire the same 14 instructions per cycle but at a lower clock. Plenty of kick. It's only 2 cores so it doesn't have much problem with memory locking and it operates at pretty much max throughput.

 

Simply configuring your system to run the memory at max clock while giving you whatever cpu power your comfortable with and maximizing your hypertransport link if you run AMD cpu or qpi link if you run intel cpu will give you the most your video card can do. I achieve a good benchable and frame rate provable increase in video speed by doing this as well as subjective "computer feels snappy".

 

All that being said. It's a GTX 460. It's not a monster but it's the redline of middle graphics market enthusiast market graphics card. It should eat just about anything you throw at it. And if you understand the kinds of things I'm talking about in this post then you understand the bottlenecks are all inside that card. Fermi was delayed because they tried to put a bunch of passthroughs in the silicon and make a system with a huge bus bar that wouldn't have a whole lot of internal bottlenecking. It resulted in a chip with terrible yields because they didn't put in enough reduntant passthroughs. After redesigning it throwing out 32 sp's and filling that space with more passhtrough redundancies they were able to get yeilds back up from nearly NOTHING. The GTX 460 simply represents failures of the GTX 480 chip. They turn off packs where the passhtroughs failed or the shader packs failed during the masking process and the result is a chip that operates very well on what works inside of it. Plus I doubt your ram runs at 600mhz. DDR 2 has 667 and 800 and 1033 clock rates. They really represent the 'edges' of timing though with 667 operating at 4 wait states, 800 operating at 5 wait states and, and 1033 operating at 6 wait states. Latency for inital transfer is around 15ms for 667 or 60ms with 4 wait states, 13ms for 800 or 65ms at 5 wait states, and 10 ms or 60 ms at 6 wait states though some 1033 ram can do 5-5-5-5 timings. The actual initial timing pulse improves though giving about a 58ish ms 667 system which can overclock down to about 54ish ms, a 54ish 800 system which can overclock down to around 49-51 and, and about a 52ish 1033 system which can overclock down to around the same 49-51 as most 667 800 and 1033 are generally the same 55 to 40nm process theres just not a whole lot of difference between them. Now those fancy 30nm chips which make up the 4gb single sticks they'll twitch a might faster but you can't get them as they are presently being swallowed whole by the server industry.

[urbex]

tl;dr = you're fine