Radeon 9600: ultimate SilentPC video card?

[wumpus]

I recently ordered this generic Radeon 9600 256mb video card, which arrived exactly as pictured:

http://www.newegg.com/app/ViewProductDe ... 320&depa=0

I knew that the Radeon 9600 ran cool and performed well, but I had no idea it ran this cool. For reference these are the system specs:

• Athlon XP t-bred @ 2100mhz and 1.75v
• MSI nForce1 mobo, onboard sound/lan active
• WD 80mb SE 7200rpm hd
• Fortron 120mm 300w PS
• Radeon 9600 per above link, running at 390/200 (out of box 325/185)

Kill-a-watt LCD says:
- at Windows XP desktop: 108w
- folding@home: 125w
- 3dMark2001: 131w

Therefore , from 0% --> 100% load for..
- cpu: draw of 17w
- video: draw of 6w

converting from "wall" to "inside computer voltage".. adjust for power supply efficiency @ 67.5%.. 6w * 0.675 = 4w

FOUR WATTS UNDER LOAD?! And remember, this is a video card that has full DirectX9 support, runs current brand new games at very respectable frame rates, and delivers near Ti-4600 performance in 3dMark2001 for older games (~8500).

Furthermore, My Mk. I finger confirms what the kill-a-watt says: I can barely get this card to get WARM under any kind of 3D testing, and that's with an extremely modest heatsink. For comparsion purposes, the Ti 4200 in my wife's computer with the ginormous Zalman HP80A heatsink/heatpipe gets borderline hot all over, at the windows desktop, much less under load.

These are amazing numbers.

More anecdotal evidence to support the ultra-cool-running nature of the 9600 exists in this tech-report article:
http://www.tech-report.com/etc/2003q1/a ... dex.x?pg=1

the Mobility and desktop versions of the Radeon 9600 are very similar. Like the desktop Radeon 9600, the Mobility Radeon 9600 is a 4x1-pipe design featuring two vertex engines and support for vertex and pixel shaders 2.0. The chip also features full DirectX 9 compatibility and all the 128-bit internal precision of ATI's DirectX 9-compliant Radeons. The desktop Radeon 9600's lossless color and Z-compression schemes are also featured in the Mobility Radeon 9600s along with up to 6X antialiasing and 16X anisotropic filtering.

Like the desktop Radeon 9600 family, the Mobility Radeon 9600s are built on a 0.13-micron manufacturing process that ATI claims it has had no problems migrating to. Though ATI has yet to confirm how many transistors make up the desktop Radeon 9600, the Mobility Radeon 9600s weigh in at only 60 million transistors; quite a remarkable transistor count considering that the Mobility Radeon 9600s not only support all the features of their desktop siblings, but also a few Mobility-specific features that aren't present on desktop versions of the chips.

the desktop and laptop versions of the 9600 are EXTREMELY similar. Same tech, same process, same featureset.. etc.

Can anyone else with a kill-a-watt and a Radeon 9600 confirm these numbers?

[wumpus]

Here's the card.

Now imagine looping 3dmark2001 for two hours, opening the case, and finding the heatsink slightly warm to the touch.

[Rusty075]

Note: This thread is actually a continuation of an off-topic tangent that wumpus and I got into in a different thread, so here's my copy/pasted commentary on his numbers, just for the sake of continuity:

If we go by this chart that we've been using: VGA card power dissipation with the 9700Pro being listed at 54 watts, and extrapolate that data into this chart: Computerbase's Radeon 7500 - 9800 Pro charts (the second chart on the page, titled Codecreatures, since we're talking max load):

Take the 192.1 watts that the 9700 Pro system is drawing, and subtract out the 54 watts for the VGA card, you get 138 watts for the rest of the testbed system. (double checking that with the 8500 listed on both charts seems to confirm that methodology, I think)

Since they use the same testbed for all the tests, you can subtract that 138 watts from the result for the 9600, from the Codecreatures chart, to arrive at it's wattage.

165.1-138= 27.1 watts.

The fact that your card is running at a 25% OC should bump your number up even higher....

Any theories on this?

I'm not trying to be argumentative, and I'm sure that the kill-a-watt is actually saying what you're reporting, I'm just trying to reconcile it with the readings from the other cards.

I just can't see how the die-shrink that took the 9500 to the 9600 could have resulted in a 88% drop in wattage, as your numbers appear to claim. (it mainly was a pure die-shrink, little else changed architecturally) If you look at the "Codecreatures" chart, the 9600 is lower than the 9500, but only by 12 watts, which seems more reasonable.

I see two immediate flaws in your "4 watts" theory:

1. You seem to be saying that at idle the VGA isn't using any power, which is obviously false.
2. You assume that the CPU load, and hence wattage, stays at 100% for 3DMark. 3DMark is specifically designed to test VGA cards, not load CPU's. The CPU is probably not running at full wattage then, so the increase in wattage for the VGA is being "hidden" by that reduction

Try running CPUBurn and 3DMark at the same time, and see what the kill-a-watt says.

[wumpus]

1. You seem to be saying that at idle the VGA isn't using any power, which is obviously false.

No, just that I expect the lion's share of power usage to be when the video card is doing 3D work. I can't think of a good way to baseline the "loading into windows" power profile of the video card.. maybe if I had a totally ancient 2D PCI video card like a Matrox Millennium, which presumably uses as little power as a video card imaginably could? Then I could boot into that and compare the kill-a-watt numbers?

2. You assume that the CPU load, and hence wattage, stays at 100% for 3DMark. 3DMark is specifically designed to test VGA cards, not load CPU's. The CPU is probably not running at full wattage then, so the increase in wattage for the VGA is being "hidden" by that reduction

I tried running each of the following during 3dMark2001:
A) Prime95
B) Folding@Home
C) nothing
Results..

- at Windows XP desktop: 108w
- folding@home: 125w
- prime95: 128w
- 3dMark2001: 131w
- 3dMark2001 + prime95: 131w
- 3dMark2001 + folding@home: 131w

131w is the highest observed during test 1 (the car scene), both low and high detail, for consistency's sake.

HOWEVER. I did notice that a few of the tests seem to be more intensive than others, notably Dragothic. Dragothic peaks at 134w, which is the highest value I see during the four scenes. I loaded up Far Cry (fantastic game) which does use a lot of DX9 features, and I could get that game to peak at 134w as well-- oddly enough mostly when I was swimming around underwater.

So the other lesson is, dragothic is the best way to load the video card. The real max power usage isn't 131w, it's 134w.

Then again the 100% CPU baseline is somewhat variable too. Do we use 125w folding as the baseline? or the 128w prime95? So it ends up being kind of a wash.

Conservative, worst case scenario: 134w - 125w = 9w * 0.675 = 6w. A whopping six watts, plus whatever the idle power draw is for the video card.

For what it's worth, I did test monitor sleep, which reduces the baseline PC power draw from 107w to 106w, very consistently. Monitor sleeps, 106w, monitor wakes up, 107w.

[wumpus]

looking at the windows desktop (idle) numbers here

http://forums.silentpcreview.com/viewtopic.php?t=7753

Again I am struck by how cool the stock heatsink on this Radeon 9600 stays, even under load. Clearly this RV350 chip is absolutely comfortable with a modest passive cooling solution.

I find myself asking: how different is a GPU from a CPU? In other words, what was the last CPU we could buy and install where the CPU vendor approved passive cooling? That's got to give us a decent rule of thumb for guesstimating how much power the RV350 chip is dissipating!

http://users.erols.com/chare/elec.htm

The first PC I owned was a Micron 486dx2-66 and I remember that having a heatsink and fan-- although the fan eventually died on it and I never fixed it, so technically it wasn't required. According to the chart linked, a 486dx2-66 dissipates 6w.

I also remember using a 486dx2-50 PC at work around the same time, and I distinctly remember that processor not even having a heatsink. Why? Because I was messing around with that PC once and touched the CPU die after I opened the case-- and nearly burned my finger off! 486dx2-50 is supposed to dissipate 5w according to the same chart.

For Intel, I think the transition to heatsinks was around the 486-66 (6w), and the transition to heatsinks and FANS, was around the early Pentiums (15~20w).

Aren't those reasonable guidelines for video cards too? Heat is heat-- and a GPU die isn't that different than a CPU die. These numbers seem consistent with the TEC Channel wattage rankings, too.. any video card with a desktop idle of >15w appears to come "standard" from most vendors with an active heatsink/fan assembly.

[wumpus]

OK, the Via C3 "Eden" CPUs-- which VIA approves to cool passively up to I think 800mhz?-- dissipate a lot less power than I thought:

http://www.via.com.tw/en/viac3/pcb.jsp

733mhz -- 11w
800mhz -- 12w
1000mhz -- 15w

However, anecdotally, the CPU heatsinks on both of my Eden boards (I have a 533mhz and a Nehemiah 1ghz; note that the 1ghz board shipped from VIA with a heatsink/fan) get MUCH hotter under 100% load than the Radeon 9600 heatsink does. Not even close, really.

[Ralf Hutter]

Well, I don't know about all the fancy numbers stuff and the bickering, but my 9600 Pro Ultimate runs very, very cool. It's heatsink barely gets warm while running 3DMark or actual games. It doesn't heat up the case at all, compared to the ATi 8500 that it replaced. To me, that equates to a cool-running card, even without firing up the Cray to do all the math.

Dis be da one:

[wumpus]

Bickering? who is bickering? yeah, that heatpipe is probably overkill for the 9600 pro. I put one of the older zalman zm17-cu models on mine, and I can barely get the ends of the fins to warm up. nexfan has them for $7.99, which is an incredible deal.

Some clock rates for different 9600 models.

Code: Select all

rv350 9600          325mhz / 200mhz DDR
rv350 9600 "pro"    400mhz / 300mhz DDR
rv360 9600 "xt"     500mhz / 300mhz DDR

rv350 and rv360 are very similar. Same 0.13 micron process, except the rv360 uses a "high k insulators" to achieve higher clock speeds, and it offers on-die temp monitoring.

Not sure how much more heat the 9600 xt would produce with its higher clock speeds, but I can clock my 9600 to Pro core levels (well, 390mhz actually) with good stability and heat generation is still virtually nil.

Paging through some reviews, looks like the XT is about 10% faster than the Pro. Probably not worth the cost differential, as that's getting into 9700 pro and 9800 se pricing territory, both of which offer dramatically better performance-- at a cost of massive increase in heat output. Might still be worth it for silentpc enthusiasts, though!

prices from newegg (least expensive model shown):

128mb 9600 xt - $153
128mb 9600 pro - $126
128mb 9600 - $96

[dasman]

Wumpus,

Just out of curiosity -- what card did your 9600 replace. Do you access to any of the cards on the list Rusty is using -- possibly you can swap cards and see what your kill-a-watt reads?

Assuming that the wattage numbers from the german site are correct (big assumption), then only question I have with Russ's methodology is where is PSU efficiency get accounted for?

Assuming 70% efficiency and the 9700pro, isn't the testbed really drawing:
192.1*0.7-54=80.5 watts

That puts the 9600 at:
165.1*.7-80.5=35.1 watts

Of course, this assumes that the 54watt number is the draw of the card itself (how they came up with that, I have no idea).

While I believe Ralf's statement that his 9600 is cooler than his 8500 was, I also find my 8500 only warm to the touch using the Mark I finger test -- I've got a Zalman NB cooler on my 8500 without any fans blowing in it's direction.

If the 54watt number from the charts is suspect, then this entire discussion is suspect as well.

Dave


PS I still question the leap that taken when we say that since the feature sets and similar and the 0.13 manf process is the same, that the architecture and power consumption of the mobility vs the desktop chips must be similar. I'm not saying that they may not end up being similar, but I don't think we can necessarily extrapolate that from the Tech Report quote.

[Nowhere_man]

Let's talk frames per second, what are you getting out of that 9600 during games and benchmarks Wumpus?

[wumpus]

I switched in my Ti 4600 (all stock) for comparison. Same system, numbers observed on kill-a-watt LCD:

Idle at Windows desktop: 135w
Idle at Windows desktop with Ti4600 fan disconnected: 132w
folding@home: 154w
prime95: 157w
Max observed during 3dMark2001 test 1: 157w
Max observed during 3dmark2001 test 2: 164w

We have to factor in the 3w of power drawn by the Ti4600 fan (!), I verified this a few times by plugging and unplugging it and watching the kill-a-watt LCD.

So, basically, the Ti4600 appears to draw near load power all the time, eg it doesn't ever throttle or power down on idle. We can conclude this by comparing the 3dmark2001 results with the prime95 results.. almost identical (!)

windows desktop (idle) load:
Radeon 9600: 108w
Ti 4600: 132w (+24w)

Increase over folding load, during 3dmark2001 test 1:
Radeon 9600: +6w
Ti 4600: +3w

Increase over folding load, during 3dmark2001 test 2:
Radeon 9600: +9w
Ti 4600: +10w

I tested again and there is no change in wattage used during 3dmark2001 if prime95 or folding is running in the background.

Let's talk frames per second, what are you getting out of that 9600 during games and benchmarks Wumpus?

Nothing special about this card.. just do a google search "radeon 9600" review and pay attention to the clock rates of the video card, and the CPU/mobo that it is installed in. On this particular system, Ti 4600 generates 3dmark2001 of 9134 at stock speeds, and the Radeon 9600 generates 3dmark2001 of ~8500 at 390/200 (from 325/185). And remember the Ti4600 has no DX9 features..

[wumpus]

I remembered this nForce1 board has onboard GeForce2MX, so I tried that too. This is with NOTHING in the AGP slot:

Idle at Windows desktop: 103w
folding@home: 120w
prime95: 122w
Max observed during 3dMark2001 test 1: 122w
Max observed during 3dmark2001 test 2: 123w

So here's what we have.. all numbers in watts from the same system, the only variable is the video card.

Code: Select all

         gf2i    r9600        ti4600 
         ----    -----        ------ 
idle     103     108 (+5)     135 (+32)
fold     120     125 (+5)     154 (+34)
prime    122     128 (+5)     157 (+35)
3d2k-t1  122     131 (+9)     157 (+35)
3d2k-t2  123     134 (+11)    164 (+41)

And the resulting 3dMark2001 scores from this system:

gf2 integrated, stock: ~2000
r9600 overclocked: ~8500
ti4600 stock: ~9100

[wumpus]

OK, now that I've gathered all the data, I am totally comfortable concluding that the (semi-overclocked) Radeon 9600 draws..

~3.5w (5w * 0.675) at idle
~7.5w (11w * 0.675) under 100% load.

Pretty easy to see how this card can be cooled passively without even breaking a sweat. The amazing thing is the PERFORMANCE you get for that minimal level of heat! Now if only AMD and Intel could do the same for CPUs..

[sthayashi]

Hmm... I might have to consider the Radeon 9600 vanilla as my next video card, since I need a new quiet video card w/ DVI.

[dasman]

Ok, if we assume the Ti4600 roughly matches the Creative Labs 3D Blaster Ti4600 in Russ's link (24w idle & 40w load), then maybe we can come up with some rough comparisons (I'm gonna assume your Fortron is 70% efficient because I'm lazy and don't want to find the real number )

With Ti4600:
idle = 135w*0.7-24w=70.5 system draw
load = 164*0.7-40w=74.8 system draw (assuming 40w load for Ti)
load = 164*0.7-24w=90.8 system draw (assuming no increase in Ti load)

With the radeon 9600:
idle = 108w*0.7-70.5 system draw=5.1w for radeon @ idle
load = 131w*0.7-74.8 system draw=16.9w for radeon @ load
load = 131w*0.7-90.8 system draw=0.9w for radeon @ load

Now, I have a bunch of problems with this entire exercise -- something, obviously, doesn't add up. If we assume the kill-a-watt can be interpreted to mean the Ti doesn't increase @ load, that translates into the radeon draw actually reducing power @ load . I think everyone can agree that isn't going to happen.

If we assume that Russ's link numbers for the Ti are correct, then we're coming up with entirely believable numbers for the radeon at load (16.9w) and less than the 29w of Ralf's and my Radeon 8500's (which jives with Ralf's assertion that the 9600 is cooler than the 8500). However, these numbers mean that the CPU only draws 4.3w more power at load than at idle. I don't know enough about the Tbred to know if that's realistic, but I have to imagine that it is not.

What's this all mean? That our testing method is somehow flawed -- each result has problems.

Question, if you let the power saving features of XP turn the monitor off, does that shut down the video card processing? It may make sense to see if the kill-a-watt draw changes when the video/monitor powers down. If it does change, then we need to run F@H or Prime with the XP powering down the video to get a true CPU only load.

If the kill-a-watt draw doesn't change w/ XP's power saving, then I don't know what the deal is, something in the methodolgy is fouled up that is beyond me.


Dave


EDIT Sure, now you throw in the onboard stuff into the equation. I'll look at that tomorrow -- I'm going to sleep...

[Rusty075]

Wumpus, can you disable the onboard VGA in the BIOS, take out the VGA card, and then boot the machine?

With no video it won't POST, but it should run. You can take a kill-a-watt reading while listening to the beep code

That "should" give you a number with no video at all.

[dasman]

Wumpus, can you disable the onboard VGA in the BIOS, take out the VGA card, and then boot the machine?

With no video it won't POST, but it should run. You can take a kill-a-watt reading while listening to the beep code

That "should" give you a number with no video at all.

It doesn't get more "idle" than that...

Dave

EDIT One thing that could be changing things -- I'll how much extra power does a HD take idle vs load? I'll be F@H and Prime leave the HD pretty much alone -- does 3dmark access the HD? (I've never run it, so I don't know).

[Rusty075]

The difference between idle and load for an HDD is more than 10 watts, since wumpus is watching for Max load on the Kill-a-watt, you should probably deduct that form the calculations as well.

(Seagate tech doc's list Idle as 0.5 watts, and Seek as 12.5 )

[Rusty075]

There's another significant flaw in the readings: The CPU

If you take the lowest F@H number, which is with the on-board VGA: 120watts.

Multiply that by the PSU's efficiency to get what the system is actually using: We don't know what that is, so I'll use a range from 60% to 80%, it's most likely in there somewhere.

120*60%=72 watts
120*70%=84
120*.80%=96 watts

The CPU that Wumpus is using is a T-bred OC'd to 2100Mhz, at 1.75vcore. That means that at full load it is drawing somewhere in the neighborhood of 80-90 watts, all by itself

Now include the 10-12 watts for the HDD, plus whatever the RAM, the mobo, the soundcard, the NIC, the on-board VGA, etc are using, and you very quickly run into trouble with Law of Conservation of Energy.

[wumpus]

Now include the 10-12 watts for the HDD, plus whatever the RAM, the mobo, the soundcard, the NIC, the on-board VGA, etc are using, and you very quickly run into trouble with Law of Conservation of Energy.

Remember this is an nForce board so sound and LAN are integrated.

All my kill-a-watt testing over the last 6 months or so has confirmed one thing: once you subtract CPU and video card usage, the remaining power draw is really quite small.

[wumpus]

OK, kids, don't try this at home, but..

Radeon 9600 at windows desktop (idle): 107w

after unplugging hard drive while system is running (!) 100w

So 7w for the 80gb WD 800JB from the wall. Ironically, or not, that matches exactly with their datasheet for this product..

http://www.wdc.com/en/products/products.asp?DriveID=32

Power Dissipation
Read/Write 7.00 Watts
Idle 7.00 Watts
Standby 1.00 Watts
Sleep 0.70 Watts

[wumpus]

[combined]

[wumpus]

Wumpus, can you disable the onboard VGA in the BIOS, take out the VGA card, and then boot the machine?

Can't. The "disable" is to plug in an AGP video card. Otherwise you're always booting up onboard video.

This is certainly an experiment that someone else with a kill-a-watt and a different motherboard could do.. although I did find that the watts used in the BIOS screen (110w) is higher than the watts used at the windows desktop (108w), so bear that in mind.

[wumpus]

A few more misc things I determined power for via the "ghettotech(tm)" before and after method:

Code: Select all

WDD 800JB HDD:                         7w
CD-R and DVD-ROM:                      3w
Vantec Stealth case fan, 3-pin:        1w
Vantec Aeroflow TMD 70mm fan, 3-pin:   5w (!)
512mb DDR DIMM:                      0.5w (estimated*)

Really nothing else to unplug or disconnect at this point.

* I booted with the only change being half the memory (one 512mb DIMM instead of two 512mb DIMMs). The desktop idle bounced between 104w and 103w, which it had not done before, so I'm guesstimating <1w per DIMM.

[Rusty075]

.....after unplugging hard drive while system is running.....

Wumpus, it will never be said that you're not willing to take risks for the advancement of SPCR community knowledge.


I'm about out of things to ask you to do. The questions aren't all answered, but we're probably as close as we'll get for now.

My testbed machine is occupied at the moment, but once it's free I'll use its on-board VGA and my assortment of cards to do some testing to compare with your's. (a bit of semi-on-topic trivia: the CPU cooler its running right now draws 30 watts at full load )


All this work you've gone through does illustrate one point very well:

We need more information on the power draw of Video Cards

The question is how to get it. Ati and Nvidia are really lacking in the technical documents department. Even the "technical documents" that they release to the press are nothing more than the usual techno-marketing garbage.

It seems reasonable that you could use an ammeter to measure the draw through the AGP slot, since the information on which pins are the power leads is public knowledge. And you could do the same for the supplemental molex power connectors on the higher-end cards. But that sort of investigation is beyond my electrical knowledge. Any electrical engineers out there want to step up?

[Nowhere_man]

wumpus wrote

just do a google search "radeon 9600" review and pay attention to the clock rates of the video card, and the CPU/mobo that it is installed in.

cute.

[dasman]

All this work you've gone through does illustrate one point very well:

We need more information on the power draw of Video Cards

Agreed! My head is spinning right now...

Dave

[wumpus]

Is this really all that complicated?

Idle at Windows desktop: 108w

.. subtract, per kill-a-watt
7w for HDD
5w for Radeon 9600 (using baseline over integrated gf2mx)
5w for aeroflow 70mm TMD fan
3w for CD-R and DVD-ROM
1w for two 512mb DIMMs
1w for Vantec Stealth 80mm case fan

.. that leaves 108w - 22w or 86w for the motherboard itself and the CPU.

86w * 0.675 = 58w

I think the mobo (northbridge, southbridge, etc) draws probably ~5w, certainly less than 10w. That leaves us with ~50w for an idle Athlon 2100mhz @ 1.75v.

According to this chart..
http://users.erols.com/chare/elec.htm

the 2.13ghz T-bred XP-2600+ @ 1.65v dissipates 62w "typical".

Maybe "typical" heat dissipation isn't so typical? Maybe it's actually variable depending on the CPU sample? Anecdotally, we know that certain chips run "cooler" than other chips even of the same model # and designation-- I've certainly experienced this myself..

Also according to the chart, the range for a 2600+ is 62w - 68w or +6w. On this machine, I observed a peak from idle --> folding of triple that.. +17w! So maybe that's where some of the "missing" watts went. It dissipates less watts at idle than anticipated, but also dissipates more than anticipated under load.

[dasman]

But what does that tell us? It appears to tell us that the 5w calc'd value for the Radeon 9600 @ idle is somewhat accurate. I don't think anyone was necessarily disputing the idle draw -- it was the at load draw that was the issue.

Your link for the CPU wattages is helpful though -- it lists 62w for "typical" (idle?) T-bred and 68.3w for "maximum" (f@h?). A difference of only 6.3w.

From my previous post:

If we assume that Russ's link numbers for the Ti are correct, then we're coming up with entirely believable numbers for the radeon at load (16.9w) and less than the 29w of Ralf's and my Radeon 8500's (which jives with Ralf's assertion that the 9600 is cooler than the 8500). However, these numbers mean that the CPU only draws 4.3w more power at load than at idle. I don't know enough about the Tbred to know if that's realistic, but I have to imagine that it is not.

6.3w vs. 4.3w -- probably a difference in the PSU efficiency I used (70%) vs what your particular PSU actually is.

To me, this all comes to together as the Radeon 9600 drawing ~5w idle and ~16w @ load. The numbers in Russ's link appear to jive with what you've measured.

Can we put this to bed now? This much thinking makes me tired

Dave


EDIT The +17w when folding can probably be attributed to:
17w*0.65=11w-6w (from your CPU link) = a 5w discrepancy I'm sure this could be used up by mobo components, dimms, etc -- they're all doing something F@H, at idle they are not doing much.

[Ralf Hutter]

1w for two 512mb DIMMs

I don't know how accurate this is, and I'll bet it applies to regular SDRAM, but it specifies 8W per 128MB which would math out to 64W for a gig of mermory! I believe that I've seen similar numbers on other "power use" sites as well but I can't find the appropriate bookmarks right now.