Pushing a Gainward Geforce FX 5600 XT – Malaclypse The Younger
We see lots of reviews posted about what the high-end cards can do, and a fair number of reviews on the more budget-friendly models. Generally the budget cards will have their cores clocked (and cooled) less aggressively than the big boys, and have cheaper RAM with no cooling assistance. Which one is a larger cause for concern in a gaming situation? Also, are the chipmakers “binning” the chips all that much, or just testing a certain number for the top cards and pushing the rest down the pipe?
With a newly purchased Gainward Geforce FX 5600 XT, I decided to put this question to the test. It was purchased because it does DX9, supports two monitors, and was cheap. Gaming is secondary to me, and I don’t do all that much of it, but I want to at least be able to run them to see what all the fuss is about. Fast AA and AF are nice, but at four times the price? It’s hard to beat $99 for a modern video card and I don’t really have room in my post office box for a bulky OTES-type card anyhow.
As you are probably aware, a little Windows registry hack enables the “Coolbits” portion of nVidia’s driver, and when I made this change I was shocked to find out just how low the card is clocked by default. 400 MHz DDR made sense, the memory is 5 ns Hynix (right on spec). I didn’t figure I’d get too much more out of this but I figured I’d try.
The core, on the other hand, was clocked at an alarming 250 MHz. I guess that’s why it was cheap. I had a feeling I’d be able to push the core considerably higher (about 20%) even with the heatsink that looks like it’s made of a shredded tin can, and expected about 10% on the RAM, since that seems to be easily achievable out of most standard DDR400, which this probably is.
(I know all the controversy surrounding 3Dmark03, but I still figured it’d be good for one thing — comparing a card against itself, with no other changes to the system. If the driver cheats, it’s going to cheat at every clock speed, making it sort of irrelevant. It may be a highly suspect benchmark for apples-to-pears-to-oranges-to-tangerines sort of comparison, but I figured it’d do fine for raisins to raisins.)
The base system in all cases is an Athlon XP 2000+ (Tbred A) bumped up to 1900 MHz by pushing the FSB to 152 MHz. That’s all this old Abit KG7 can muster up, no matter what processor I drop in, so I may not have topped out the chip, but it doesn’t make budget sense to buy a new motherboard to see how far I can push a processor I paid $75 for. I figure if I need a little more speed I can just slap on some Type R stickers and a few more cold cathodes and aluminum fans. Also Win2k SP4 if you care about that. 🙂
So here’s what you all want to know if you’re still reading this:
I decided to bump up the memory first, since it was likely I’d hit the wall there first (and therefore faster).
Severe rendering errors (did not crash, I aborted the run)
With rendering errors
No errors, even on multiple repetitions
So my top memory speed is somewhere between 463 and 469, but it’s still January. I don’t want the system weirding out in the first heat wave of March because I pushed it to the edge when it was 20 degrees cooler. I opted to leave it at 463.
Then I went for core speed.
Lockup on second game test
After a reset, I started back at it again, but neglected to bump the memory back up to 463. So I got some initially odd-looking results (341 core was slower than 325?!) but it made sense once I noticed the error. It turns out it kinda fills out the box of data points anyhow, so my carelessness was useful in the end.
I didn’t notice the 400 MHz RAM, so this surprised me
OK, so I wasn’t overheating the core and making it throttle… hmm… Then I noticed the RAM speed. (fnord) 🙂
24.6% total increase
I could probably squeeze a few more MHz out of the core, but I didn’t because of that tin can cooler. I didn’t have anything smaller than a PIII heatsink laying around, so I decided to stop and write this up.
Breaking this down to percentages:
Max Memory Overclock = 463/400 = 15.75% increase
Speed increase from memory overclocking alone = 11.42%
Scales pretty darn well! Even small memory speed bumps translate to real speedups.
Max Core Overclock = 341/250 = 36.4% increase
Speed increase from core overclocking alone = 13.5%
A real benefit, but not nearly as good a correlation to total speed as with memory.
Max Total Overclock = 341/463
Speed increase from overclocking both = 24.6%
Ultimately I surpassed both of the marks I expected going in (10% memory and 20% core overclocks), and found that they both matter – it’s just that memory speed increases are about 85% effective. That is, for every 1% of memory overclock, you get 0.85% increase in total speed.
Obviously, this is the primary bottleneck with this inexpensive card. Core helps to some degree and is worth optimizing, but if you’re picking a budget card with speed in mind, you should be looking for the fastest RAM. Core overclock effectiveness was only about 0.37% — a 1% core speedup only produces 0.37% total speedup.
I have a feeling this will all come unglued if the card can run the memory at twice the core speed. At that point, memory may not be your bottleneck any more. Maybe as a follow-up I’ll underclock the core to see if this is true.
As you may recall, my last overclocking effort was on a Geforce FX 5600 XT, and I got slightly better than expected results. Problem is, that still left the card just dog-slow, mostly due to the 400 MHz RAM — the reason why the card was so cheap.
Well I done gone and stuffed it all up… I tried putting pieces of a hacked-up Pentium heatsink on as both GPU and RAM sinks. The operation was a success, but while I was doing this, my girlfriend came into the room and kissed me…
…and I thought several profane words as I felt the static jump from her to me, me with a video card in my hand.
Let’s just say we have named this “The $99 Kiss”.
I was able to use the card for 2D and old (like Quake) 3D, but anything more recent had horrible artifacts dominating the screen. Nothing to do about it but order another one, which we did. The only difference was that the replacement was supposed to be a MSI 5600XT.
A few days ago, the new card arrived, I powered the machine down and swapped them. I was surprised when Windows came up at 800 x 600 in 16 colors — I thought the change would be completely transparent. So I go through the routine of reinstalling the drivers and rebooting, and it’s all good.
Then I re-enabled the Coolbits hack and went to work. I expected the clock frequencies to be 250/400, but they were at 325/600. Odd, that doesn’t sound like a 5600XT… that sounds like a 5600. Further poking revealed that in fact it IS a 5600, not the lobotomized little brother. I checked the order page I had saved, and sure enough it doesn’t say 5600XT, it says 5600. I checked the invoice in the box, and it says 5600XT. Interesting.
I promptly informed my girlfriend that the $99 kiss had just been discounted to $59 or so, as the replacement card appears to be worth about $40 more than we paid for it. But that saved web page did say 5600, so that is in fact what I ordered, even if I didn’t mean to. I’m keeping it. 🙂
Once again I tested the card at stock speeds, top GPU/stock RAM, stock GPU/top RAM, and top of both. There was an auto-detect button, but it managed to both find lower settings than I did and induce artifacts at the same time. The artifacts went away on a reboot, so it appears the driver has a self-destructive streak.
OK here are the 3D Mark03 scores once again:
For comparison purposes, I set the clocks to the same as they’d been on the “kiss” card – 341/463. I expected it to behave like the XT when clocked down to that level, but it didn’t, by a good margin:
5600FX: 341/463 = 1615 (from previous article)
5600 : 341/463 = 2152
It appears the new card is 33% faster than the old one, at the SAME clock speeds. I have no explanation for this, but once again I’ll keep it. 🙂
Now for the overclock testing:
|This is the speed out of the box.|
|6.5% score increase, 15% core increase (0.43|
|2.75% score increase, 12.8% memory increase (0.21|
|10.7% score increase|
The GPU clock speed continues to scale at a rate similar to that of the 5600FX — for every 1% increase in clock speed, there was a 0.43% increase in score. Given the small sample size, this matches pretty well with the results the first time.
The memory is a different matter entirely. At these speeds, where the memory clock is approaching double that of the GPU, bumping up the memory clock does very little. It’s definitely not the bottleneck it was at 400 MHz.
I was a bit disappointed in the lack of overclocking headroom on the RAM of this new card, until I realized that it wasn’t going to accomplish much anyhow. The GPU headroom was also substantially less (relative to stock) than with the 5600XT, but that is hardly surprising and a 15% overclock isn’t bad.
My conclusion from these new results is that memory clock must stop being a bottleneck somewhere between the 1.36 ratio of the 5600XT and the 1.81 ratio of the 5600. I’d be interested in the shape of the curve between these two points, but I think I can safely say that if you want bang for your buck, you still want to be looking for the fastest RAM, up till a 3:2 or 7:4 ratio is reached.
After that, the bottleneck shifts to some other component. However, that doesn’t appear to be the GPU, since overclocking there seems to yield similar benefits on both cards.¹
My advice to readers would be to buy a 5600 at a 5600XT price if such an offer becomes available.
¹ Actually I think I know what’s up here: The new card has 128 MB, but it’s in 8 chips, while the old card had just 4. This makes me think the non-XT has a 128-bit memory bus while the XT has only a 64-bit memory bus. This would explain both the discrepancy at the “same” speed and the fact that cranking up the memory clock does very little on the non-XT board.
PS: Copyleft is in effect. Do whatever you want with this so long as it does not infringe on anyone else’s right to do as they want with it. I suggest using it to paper your bird cage.