Bargain Hardware Thread - Page 20

Xyzzy · 2014-11-11, 22:23

Some thoughts:

At one time we had four GTX570 cards running. They were so power-hungry and threw off so much heat that we had to put each one into a separate computer. Each card was three slots wide, which posed a physical issue as well. Note that this was when the CPU did the sieving so that was another reason we were forced to run one card per box.

They churned out some serious work, and at one point we think we were at #1 for TF and #4 overall, both using the "lifetime" metric. We think we were outputting around 1,600GHd/d with that setup. The current draw was so high that we had to dedicate a branch circuit for them and the room the boxes were in used to hit 85°F in the summer, no matter what we set the thermostat to. However, the ridiculous cost to power and cool the boxes really hit us hard and eventually we had to retire them. This has cost us dearly in our overall standing on the leaderboard since nearly everyone else has passed us by.

So today we received two GTX980s. They both fit into one computer case and together they easily output 1,200GHd/d with little fuss. We have not measured their current draw but thermally they feel like they output the same heat as one 570. We are using a non-dedicated branch circuit with no problem as well. And, the CPU gets to run P-1 testing since GPU sieving is now thankfully an option.

The good:

They are relatively quiet.
They are very sturdy and do not sag at all, unlike our previous video cards.
They are energy efficient and relatively inexpensive.
One $600 980 does more TF work than a $1,000 Titan. (We paid that much for a Titan. They are probably cheaper now.)
One $600 980 does about the same work as a $1,000 690. (We paid that much for a 690. They are probably cheaper now.)
They have a three year warranty from a company known for being honest about warranties.

The bad:

The fan BIOS curve is weighted towards lower acoustics, so OOTB they throttle very quickly.
One of the cards was sent to us with a missing rubber SLI connector cover. We will have to contact EVGA about that.
We ran out of video card slots so our 750 is now homeless. (The 750 is good for around 175GHd/d.)
Our 570 cards were shipped in very exquisite packaging whereas the 980s came in ordinary packaging.

The ugly:

We have to run Windows to modify the fan curves and other GPU options. Plus, GPU-Z is a Windows program.
We have not figured out how to use the onboard (Intel) graphics for the display rather than one of the cards.

Random notes:

We spent several hours playing with all sorts of parameters to see where the "sweet spot" was for performance. After much testing, in mfaktc.ini we set "GPUSievePrimes=82486", "GPUSieveSize=128" and "GPUSieveProcessSize=32". We set the "power target" for the card to 125%, the temperature "limiter" to 79°C and the fan curve to "1:1". We did not boost the memory or GPU clock beyond the "stock" values, which on these cards are "superclocked" from the factory. The cards are very willing to pump out about 25-30 GHd/d more each if we allow the temps to go higher. Running the fans faster helps a little but get a bit loud over 85%. The voltage and power required to go a bit faster appears to increase rapidly. The cards will output around 430GHd/d each if we set the "power target" to 75% at which point the fans go near silent and the power and voltage drop to minimal levels. Various reviews show that there is a lot of headroom to overclock the GPU clock but we are happy with what we have. We let both cards run in the case for over an hour to make sure everything had stabilized WRT heat. With our settings, no matter how high the ambient temperature gets the cards will throttle to stay under 80°C. The fan curve adjusts the fans every second so the card temperature is very stable. In the pictures attached you may notice how each card uses the reference cooling design, which appears to almost be a totally enclosed design. Fortunately our "overclocker-friendly" motherboard has the cards spaced so there is room between them. The heat from the cards is almost entirely dumped outside the case, with a little leakage around the PCI connector and the missing SLI cover. GPU-Z tells us what metric the card is throttling with. In our case it is based off of the thermal sensors. When the cards first start up the limiter is the voltage sensor, which then switches to the power sensor as the card heats up a bit. Finally they switch to throttling via the temperature sensor. We have Prime95 running one P-1 test across four cores. During stage one the CPU temperature is around 63°C and during stage two it is around 55°C. We have 12GiB dedicated to P-1 testing. The onboard video on our motherboard has only DisplayPort and HDMI outputs. We have a HDMI-to-VGA adapter, which works, but we are unable to run the display off of the onboard video and have the graphics cards work with CUDA. If anyone knows a way around this we think we can get little extra work from the 980 that is currently running the system video.

VictordeHolland · 2014-11-12, 00:11

Quote:

Originally Posted by tha

On the Dutch version of E-Bay a Tesla C2050 3Gb GPU GDDR5 NVIDIA is being offered for € 695,- It is said to be in good condition, almost new and refurbished by a dealer. How do we value such an offer?

A C2050 is basically a GTX470 with ECC and more memory. For TF you will be much better of with a used GTX580 or new GTX970. For CUDALucas the ECC memory is nice, but GPUs should really be doing TF instead of LL.

xilman · 2014-11-12, 10:55

Quote:

Originally Posted by fivemack

https://software.intel.com/en-us/art...rocessor-31s1p

...

I don't know whether there are seven people on the forum who might want one, but Colfax offers a ten-pack for $1290 so if there are more than six people it would be worth putting in a bulk order.

Another nice link

http://www.phoronix.com/scan.php?pag...tem&px=MTgzNjY

I'm now getting keener on buying one, or just possibly two, of these.

Paul

ldesnogu · 2014-11-12, 11:52

Quote:

Originally Posted by xilman

Another nice link

http://www.phoronix.com/scan.php?pag...tem&px=MTgzNjY

I'm now getting keener on buying one, or just possibly two, of these.

I would be in if these cards ran on more than a few motherboards and didn't require a very significant airflow in the case (they are passively cooled but the TDP is 270W). Too bad, the offer is definitely very attractive

Xyzzy · 2014-11-12, 15:42

The link and forum thread there suggests that you can only use Intel's compiler. Wouldn't that be a major disadvantage?

xilman · 2014-11-13, 09:51

This Twitter conversation may also prove interesting and/.or discouraging.

https://twitter.com/9600/status/532798000918454272

pinhodecarlos · 2014-11-13, 09:54

Quote:

Originally Posted by xilman

This Twitter conversation may also prove interesting and/.or discouraging.

https://twitter.com/9600/status/532798000918454272

I think they need to use heat pipes from Spirax Sarco...lol

fivemack · 2014-11-13, 11:57

Quote:

Originally Posted by Xyzzy

The link and forum thread there suggests that you can only use Intel's compiler. Wouldn't that be a major disadvantage?

Intel's compiler is reasonably compatible and generates rather good code.

However it appears that Intel has removed the offer which used to provide their compiler free for non-commercial use

ldesnogu · 2014-11-13, 12:16

Quote:

Originally Posted by fivemack

Intel's compiler is reasonably compatible and generates rather good code.

However it appears that Intel has removed the offer which used to provide their compiler free for non-commercial use

So one has to buy a $2949 compiler (since you need MPI, don't you?) to go along an <$200 card. The offer is suddenly much less attractive.

VictordeHolland · 2014-11-13, 12:38

Quote:

Originally Posted by ldesnogu

So one has to buy a $2949 compiler (since you need MPI, don't you?) to go along an <$200 card. The offer is suddenly much less attractive.

Fail, Intel!
CUDA and OpenCL compilers are free and so is the GCC (obviously).

ewmayer · 2014-11-13, 22:10

Quote:

Originally Posted by ldesnogu

So one has to buy a $2949 compiler (since you need MPI, don't you?) to go along an <$200 card. The offer is suddenly much less attractive.

Oliver's || builds of my Mlucas code in generic-C mode (cf. here) indicate that pthread support is all one needs. (Unless Intel is using MPI to support pthreads - no clue about that, or even if it makes any sense, I've not used MPI, at least not knowingly.)

So I take it GCC cannot generate machine code for Phi?

2014-11-11, 22:23	#210
Xyzzy "Mike" Aug 2002 3×2,741 Posts	Some thoughts: At one time we had four GTX570 cards running. They were so power-hungry and threw off so much heat that we had to put each one into a separate computer. Each card was three slots wide, which posed a physical issue as well. Note that this was when the CPU did the sieving so that was another reason we were forced to run one card per box. They churned out some serious work, and at one point we think we were at #1 for TF and #4 overall, both using the "lifetime" metric. We think we were outputting around 1,600GHd/d with that setup. The current draw was so high that we had to dedicate a branch circuit for them and the room the boxes were in used to hit 85°F in the summer, no matter what we set the thermostat to. However, the ridiculous cost to power and cool the boxes really hit us hard and eventually we had to retire them. This has cost us dearly in our overall standing on the leaderboard since nearly everyone else has passed us by. So today we received two GTX980s. They both fit into one computer case and together they easily output 1,200GHd/d with little fuss. We have not measured their current draw but thermally they feel like they output the same heat as one 570. We are using a non-dedicated branch circuit with no problem as well. And, the CPU gets to run P-1 testing since GPU sieving is now thankfully an option. The good: They are relatively quiet. They are very sturdy and do not sag at all, unlike our previous video cards. They are energy efficient and relatively inexpensive. One $600 980 does more TF work than a $1,000 Titan. (We paid that much for a Titan. They are probably cheaper now.) One $600 980 does about the same work as a $1,000 690. (We paid that much for a 690. They are probably cheaper now.) They have a three year warranty from a company known for being honest about warranties. The bad: The fan BIOS curve is weighted towards lower acoustics, so OOTB they throttle very quickly. One of the cards was sent to us with a missing rubber SLI connector cover. We will have to contact EVGA about that. We ran out of video card slots so our 750 is now homeless. (The 750 is good for around 175GHd/d.) Our 570 cards were shipped in very exquisite packaging whereas the 980s came in ordinary packaging. The ugly: We have to run Windows to modify the fan curves and other GPU options. Plus, GPU-Z is a Windows program. We have not figured out how to use the onboard (Intel) graphics for the display rather than one of the cards. Random notes: We spent several hours playing with all sorts of parameters to see where the "sweet spot" was for performance. After much testing, in mfaktc.ini we set "GPUSievePrimes=82486", "GPUSieveSize=128" and "GPUSieveProcessSize=32". We set the "power target" for the card to 125%, the temperature "limiter" to 79°C and the fan curve to "1:1". We did not boost the memory or GPU clock beyond the "stock" values, which on these cards are "superclocked" from the factory. The cards are very willing to pump out about 25-30 GHd/d more each if we allow the temps to go higher. Running the fans faster helps a little but get a bit loud over 85%. The voltage and power required to go a bit faster appears to increase rapidly. The cards will output around 430GHd/d each if we set the "power target" to 75% at which point the fans go near silent and the power and voltage drop to minimal levels. Various reviews show that there is a lot of headroom to overclock the GPU clock but we are happy with what we have. We let both cards run in the case for over an hour to make sure everything had stabilized WRT heat. With our settings, no matter how high the ambient temperature gets the cards will throttle to stay under 80°C. The fan curve adjusts the fans every second so the card temperature is very stable. In the pictures attached you may notice how each card uses the reference cooling design, which appears to almost be a totally enclosed design. Fortunately our "overclocker-friendly" motherboard has the cards spaced so there is room between them. The heat from the cards is almost entirely dumped outside the case, with a little leakage around the PCI connector and the missing SLI cover. GPU-Z tells us what metric the card is throttling with. In our case it is based off of the thermal sensors. When the cards first start up the limiter is the voltage sensor, which then switches to the power sensor as the card heats up a bit. Finally they switch to throttling via the temperature sensor. We have Prime95 running one P-1 test across four cores. During stage one the CPU temperature is around 63°C and during stage two it is around 55°C. We have 12GiB dedicated to P-1 testing. The onboard video on our motherboard has only DisplayPort and HDMI outputs. We have a HDMI-to-VGA adapter, which works, but we are unable to run the display off of the onboard video and have the graphics cards work with CUDA. If anyone knows a way around this we think we can get little extra work from the 980 that is currently running the system video. Attached Thumbnails

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Hardware Benchmark Jest Thread for 100M exponents	joblack	Hardware	284	2020-12-29 03:54
Garbage hardware thread	PageFault	Hardware	21	2004-07-31 20:55
Old Hardware Thread	E_tron	Hardware	0	2004-06-18 03:32
Deutscher Thread (german thread)	TauCeti	NFSNET Discussion	0	2003-12-11 22:12
Gratuitous hardware-related banana thread	GP2	Hardware	7	2003-11-24 06:13

2014-11-12, 15:42	#214
Xyzzy "Mike" Aug 2002 3·2,741 Posts	The link and forum thread there suggests that you can only use Intel's compiler. Wouldn't that be a major disadvantage?

2014-11-13, 09:51	#215
xilman Bamboozled! "𒉺𒌌𒇷𒆷𒀭" May 2003 Down not across 10752₁₀ Posts	This Twitter conversation may also prove interesting and/.or discouraging. https://twitter.com/9600/status/532798000918454272