I was talking to my bro' tonight and we were trying to figure out how cheap of an ATI card we could get away with to start sticking into our dedicated crunchers. I'm sure I've seen this info someplace but I've been looking for an hour and can't find it. Can somebody point me to list of which cards will work today with the 3rd party GPU app under Windows (and plez let a Linux app be on the horizon).

Or if I can't get all that info... lemme boil it down to a couple more specific questions

1) Will a HD3650 work with this app? (I assume 3x slower than a 38xx card of same clock speed due to 1/3 the shaders... er ah, I mean stream processors.)

2) Will a 575MHz 4830 really produce that much more output than say an 800MHz 3870?

3) Which of the 3000 & 4000 series cards draw the least juice?

Thanx in advance for any info you can provide.

Well it appears the HD3650 is off the list.

You want this post:

http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=925&nowrap=true#25910

Also gives you comparison of how each card does relative to each other.

As for your questions:

1) No.
2) Yes, because a 3870 only has 320 cores, whereas a 4830 has 640 and is more than half the speed than the 3870.
3) I go by the following chart I made:



These are from wiki and reviews. I can't be held liable if it's inaccurate :) Ideally a 700W PSU drawing 400W will perform better and last longer than a 450W drawing 400W. I've got a 550W powering 2x 4850s using a low power proc, and that just about works.

---

Borandi, a person after my own heart with the little spreadsheet action :-)


Thank you VERY much... just the sort of info I was after.

Looks like 4770s might be the best answer for my headless crunchers. But I'll have to do some poking around and factor initial cost into it also.

I wonder why the 4730 isn't down around 75w.

If you look at the efficiency charts for 80+ certified PSUs http://www.80plus.org/manu/psu/psu_join.aspx they nearly all produce their best numbers at just under 50% load. I'm told that non certified PSUs tend to have a sharper rise and fall off from that center and usually it's usually at full load where they can't meet this spec. I've got an 80+ bronze certified 530w in my desktop that I'd swear was the same company's 600w PSU with the LED fan swapped out and relabeled as a 530 so it can pass the 80+ bronze.

Anyway, I think the 80+ charts sorta bear out your general statement about PSUs.

"I wonder why the 4730 isn't down around 75w."

Well the 4770 is a new Chip (RV740) produced in 40nm; the 4730 is an old Chip (RV770) with other RAM, produced in 55nm.

Skip I'd rather do spreadsheets like that than my PhD atm :)

Yeah 4730 is basically a 4870 stripped with budget components. Also, it uses 2x 6-pin PCI-E according to Hexus, not 1 as I thought. It was made to fit a £60 price point, and still draws power.

My power calculations in that spreadsheet also may be a bit off, add about 10-15% to most. I took the values as rated by ATi, but review sites seem to suggest that they draw a bit more.

---

My power calculations in that spreadsheet also may be a bit off, add about 10-15% to most. I took the values as rated by ATi, but review sites seem to suggest that they draw a bit more.

Review sites I looked at with different cards showed total system idle/load wattage. Since I don't know what the system drew w/o any video card I can't really determine that starting point.

The point being that I think for the purposes of a factor in the table using the ATI number is better as at least it's likely to be consistent across all models.

I also suspect the review sites are putting a load on them that crunching on them might not... my 3870 will run at 98% sometimes but I have never seen it peg the meter. I suspect they are pushing the pedal through the floorboard and keeping the meter pegged with 3dmark and game clips.

---

- da shu @ HeliOS,
"A child's exposure to technology should never be predicated on an ability to afford it."

I also suspect the review sites are putting a load on them that crunching on them might not... my 3870 will run at 98% sometimes but I have never seen it peg the meter. I suspect they are pushing the pedal through the floorboard and keeping the meter pegged with 3dmark and game clips.

The ATI app uses the shaders of the GPU about as much as an average game would do. But those games will also stress the texture units and the memory interface much more than Milkyway does. So the power consumption while crunching MW should be a bit lower than in a game, especially if you underclock the memory while crunching MW (gives a lot of savings with GDDR5 based cards).

I also suspect the review sites are putting a load on them that crunching on them might not... my 3870 will run at 98% sometimes but I have never seen it peg the meter. I suspect they are pushing the pedal through the floorboard and keeping the meter pegged with 3dmark and game clips.

The ATI app uses the shaders of the GPU about as much as an average game would do. But those games will also stress the texture units and the memory interface much more than Milkyway does. So the power consumption while crunching MW should be a bit lower than in a game, especially if you underclock the memory while crunching MW (gives a lot of savings with GDDR5 based cards).

Good I just ordered a HD4770 to replace my HD3870 in my desktop and it appears to be the card that should be the easiest on the electric meter already.

I'm beating the bushes for somebody that could help ya port this over to Linux as this box is the only box I have that can boot WinXP. Everything else is running Ubuntu or some derivative of it. So right now I don't even have something to move the HD3770 to.

Is there any sort of time frame or is anyone working on porting the GPU app over to Linux 64b?

Review sites I looked at with different cards showed total system idle/load wattage. Since I don't know what the system drew w/o any video card I can't really determine that starting point.

There is equipment to do it, or an easy way without that equipment would be with two cards in the same system. Measure the full system at idle and load with one card, then add another and measure at idle and load again, compare the values.

---

Review sites I looked at with different cards showed total system idle/load wattage. Since I don't know what the system drew w/o any video card I can't really determine that starting point.

There is equipment to do it, or an easy way without that equipment would be with two cards in the same system. Measure the full system at idle and load with one card, then add another and measure at idle and load again, compare the values.

That was my point... that's what the review sites are doing is giving you idle/load total system wattages. They're not taking measurements of just the card.

Sorry, I didn't explain it clearly.

Here's my method of detecting power usage with an example:

[1] Idle with one card = 400W
[2] Full with one card = 450W

[3] Idle with two cards = 500W
[4] Full with two cards = 600W

Thus the one card uses 100W (thats [3] - [2]) on idle, 150W (thats ([3]-[2])+([2]-[1])) on full. So 150W would be the TDP in my table.

Review websites could do this, but most don't. This is just the way I would do it if I didn't have the appropriate equipment. Some websites do (Hexus, Toms Hardware), which are the ones I usually check.

---

Sorry, I didn't explain it clearly.

Here's my method of detecting power usage with an example:

[1] Idle with one card = 400W
[2] Full with one card = 450W

[3] Idle with two cards = 500W
[4] Full with two cards = 600W

Thus the one card uses 100W (thats [3] - [2]) on idle, 150W (thats ([3]-[2])+([2]-[1])) on full. So 150W would be the TDP in my table.

Review websites could do this, but most don't. This is just the way I would do it if I didn't have the appropriate equipment. Some websites do (Hexus, Toms Hardware), which are the ones I usually check.

Hmmmm.... Am I not understanding the ID schema here? [3]-[2] = 500w-450w=50w not 100w.

I agree using this example that one card is using 100w at idle but I isn't that [3]-[1] = 500w-400w=100w. Following the same pattern I agree with 150w on load but why isn't it just [4]-[2] = 600w-450w=150w ?

Yeah sorry, that's what I meant. Will teach me to drink 'n' surf :)

Hope your new card(s) help your BOINC total!

---

Yeah sorry, that's what I meant. Will teach me to drink 'n' surf :)

Hope your new card(s) help your BOINC total!

LOL... OK, thought my dumb hillbilly self was just getting denser with old age. Thanx for all.


Got a 4770 ordered... should be here next week. Couldn't get past the low power consumption after taking a look at my most recent electric bill. Think I found a home for the 3870 in my wife's Desktop (also dual boot).

---

- da shu @ HeliOS,
"A child's exposure to technology should never be predicated on an ability to afford it."

Hi, how important is the amount of RAM on these cards?
For example 4870 based cards are sold in 512meg and 1024meg versions.
Does this affect the RAC generated the card?

Mike.

Hi, how important is the amount of RAM on these cards?
For example 4870 based cards are sold in 512meg and 1024meg versions.
Does this affect the RAC generated the card?

Mike.

Not currently, but who knows in the future.....

You can even underclock the RAM without it effecting crunch times.

Someone, maybe Gipsel, said that the parameter that defines the crunching rate is the number of shaders, things like amount/speed of RAM play only a minor part. From my own experience, changing from a 3870 to a 4870 tripled the throughput.

---

Cheers,

PeterV

.

My 4850s, which normally run at 620 GPU Mhz core, 1000 Mhz mem. Currently overclocked GPU to 800 Mhz, downclocked the Mem to 750Mhz. No noticeable increase in throughput noticed, however my cards do use less of a % of the cores, thus run cooler.

Warning over/underclocking your hardware voids the warranty.

---

Warning over/underclocking your hardware voids the warranty.

And can cause warts on your private parts... unless you're GPU crunching for team Guru Mountain.... LOL

BTW, I sent my HD4770 back and am playing with a couple nvidia cards on other projects. I put the HD3870 in my wife's desktop as it's the only desktop machine that runs windoze all the time. Working like a champ over there on BOINC v6.4.7

---

- da shu @ HeliOS,
"A child's exposure to technology should never be predicated on an ability to afford it."