Hi,
several years ago I ran BOINC on an i7-machine (930). When it started to crash when it was hot, I had to stop.

Now I have the chance to start contributing again with a rejected dual XEON E5530 2.4GHz server with 24GB of RAM. I am willing to add two affordable GPUs to the system. I think about two Geforce 760 TI. There will be no power costs, because it will run on photovoltaic during daytime only. The machine will be fully dedicated to BOINC. There will be Windows Server 2008R2 running.

The Dual CPUs (8 cores) should bring roughly 130 GFlops in double precision. The 760TI in contrast a hefty 2500 each.

Q1
Does it make sense at all, to use this server? The 130 GFlops do not look attractive at all. There is another core2duo E8650 3GHz Win7 64bit machine available also. But it has a lower bandwidth and only 4GB RAM. Does this bandwith and the several cores really matter? Or are the GPUs that dominant, that I will see no difference?

Q2
I looked up the speedfactors of several GPUs and the Geforce 760 TI seem to have the best relationship between speedfactor and price (120 Euros, Speed 5.0). Is there a better alternative for 250 Euros?

---

Kind Regards,
Horst

You're doing this the wrong way around. The very first thing you need to decide is what you are going to run on that machine. Only then can you choose the hardware that suits your needs best. You'll find that different projects have very different requirements and hardware that runs like hell at one project can show a disappointing performance at another.

Just some general remarks from me, but keep in mind they might not be valid for your particular choice of projects.
- GPUs need CPU support and some GPU applications need a lot of CPU support to run efficiently. You probably won't be happy with two cores only.
- If price is a concern to you but energy cost is not, AMD could be a better choice for you than Nvidia.
- Don't focus too much on a dual GPU setup. If there are no suitable medium range GPUs available you could be better off with a single more powerful GPU.
- Look at your chosen projects' sites for information on what runs well there. There's usually statistics sections.

floyd has given some really good advice. Without us knowing more details like how much power (electricity) you actually have at your disposal really puts a limit on things.

The basics: AMD is cheaper typically up front. But, it also tends to eat a lot more electricity and generate a lot more heat.

Dual core will actually probably be just fine for a dedicated box. Especially a faster processor like 3GHz+. The problem will be if you try to run CPU work units at the same time. I would not recommend running CPU work units on the same cores as the GPU's are fed by. So, if you do go dual core, just run GPU's only.

Now the next thing to look at is the power draw on the CPU's. The e5530 is I believe 80w each at max draw. I'm not familiar with a core2duo that is e8650. Others in that line are ~65w. https://en.wikipedia.org/wiki/List_of_Intel_Core_2_microprocessors So, again depending on how much power you have at your disposal, that could be another factor in this decision. If you are wanting to run both CPU and GPU, then I recommend the dual xeon. But, project selection is the second biggest factor here.

And to be clear, those xeons are 4 core 8 thread chips. So, it would be a total of 8 cores 16 threads. So, muuuuch more CPU power than a meager dual core.

---

Now the next thing to look at is the power draw on the CPU's. The e5530 is I believe 80w each at max draw. I'm not familiar with a core2duo that is e8650. Others in that line are ~65w.

I just want to add that "Thermal Design Power" of the CPU is not exactly same as how much power the CPU will draw from the wall socket. It refers more to thermal dissipation.

http://www.cpu-world.com/Glossary/T/Thermal_Design_Power_(TDP).html

Thank you very much for all your input. With a closer look at the problem I should have come to the result myself, that the workload will be most important to the hardware.

Energy is no problem at all. Most of the time I have 5 to 10kW excess solar power. The system will not use more than 500W in total.

The projects I am interested in are mostly astrophysics like milkyway, asteroids, universe, etc. but also some biology like malaria and proteins.

I think I need a more or less general purpose machine. What do you think of having a Nvidia and an AMD grafic card in parallel?

---

Kind Regards,
Horst

The projects I am interested in are mostly astrophysics like milkyway, asteroids, universe, etc. but also some biology like malaria and proteins.

There's several projects in that list that don't have GPU applications, so you'll not want to choose the dual core which would make this basically a pure GPU machine.

Milkyway: They have applications for both AMD and Nvidia but double precision performance is vital so this should clearly be AMD territory. Not all AMD GPUs will perform well here, though.

Asteroids: They have a CUDA application but I hear it's quite inefficient. I can't say if it's a good idea to run it.

Universe: CPU only.

Malariacontrol/FiND: CPU only.

Rosetta: CPU only.

GPUGRID: Nvidia only. They have short deadlines so you'll want to run a good middle class GPU or better if not running 24/7.

Poem: Both AMD and Nvidia but the application is designed for AMD.

And some projects you haven't mentioned but I think they could match your pattern.

Einstein: Both AMD and Nvidia. AMDs have been superior but Nvidias have caught up. I wouldn't want to predict who's ahead in the future. Whatever GPU you run here, you'll want really fast GPU memory.

POGS: CPU only.

DENIS: CPU only.

I think I need a more or less general purpose machine. What do you think of having a Nvidia and an AMD grafic card in parallel?

You'd need a steady supply of both types of work and there's not much choice. Personally I'd focus on one type.

It would be a total waste of computing power to run Asteroids with GPU. That GPU-application is really so slow (your computer would crunch almost the same amount of Asteroid tasks if you run that dual core full speed or run a single GTX 760...).

But for example, one nicely working combination is to run Asteroids with CPU and Einstein with GPU. At least with an Nvidia GPU their newest application does the crunching almost completely inside the GPU (only about 1-2% stress for CPU, per running task). Also, fast RAM speed doesn't make much difference with that application anymore, because there's not so much traffic between RAM, CPU and GPU as there used to be. Then, if "A" and "E" run simultaneosly with that configuration, the CPU resources would be nicely available for Asteroids... or any other CPU application.

4GB memory is enough for many projects. Maybe not for Milkyway, I don't know, but for example Asteroids and Einstein... you could load work for all the cores/threads and it still wouldn't use all that memory. A running task requires a relatively small amount of memory on those two.

Thank you again. I had a look into the machine and found only two pcie slots. And one is at the very end, so that no dual slot card fits. I have to look for one single-slot GPU. I found one Geforce 760 TI. The other may be dual slot.

I looked at the AMD GPUs and found most equipped with 2GB. But there are some 3GB out there as well at reasonable prices. Does GPU memory size matter? Or does just the speed and architecture?

The next question arises. GPUs are typically tested by their max fps at certain resolutions in different games. Do these figures go proportional with the number crunching abilities? Is there a list with actual AMD GPUs and their ability for BOINC? I found such a list for Geforce GPUs where I picked the 760 TI from.

---

Kind Regards,
Horst

I had a look into the machine and found only two pcie slots.

I think you should try to find more information on those PCIE slots, in particular if they're both capable of powering a GPU. It's not safe to judge that by socket size. And while the box is open, see if the PSU can provide enough power and if it has additional PCIE power connectors. But this is going off topic now.

I looked at the AMD GPUs and found most equipped with 2GB. But there are some 3GB out there as well at reasonable prices. Does GPU memory size matter?

1GB has always been enough for me.

GPUs are typically tested by their max fps at certain resolutions in different games. Do these figures go proportional with the number crunching abilities?

No. Some test sites do GPU computing tests but even those numbers can't be used to judge the suitability for a particular project. The applications are just too different.

By the way, you keep talking about a GTX 760 Ti. I don't think those exist. What you seem to have in mind is the computing power of a GTX 760 at the price of a 750 Ti. You'd better rethink that. :-)

Thank you again.

The board is an ASUS Z8NR-D12 server board.

I think you should try to find more information on those PCIE slots, in particular if they're both capable of powering a GPU. It's not safe to judge that by socket size.

There is a single PCIE x16 slot. Labeled Gen2 x16 link. Therefore it should support 150 Watts. There is an additional x8 slot with an x8 link directly adjacent to the x16 one. So it is useless when I have a double slot GPU. Then there are two further x8 slots which happen to have only x4 links. I doubt, that there are suitable x4 GPUs on the market?
At last there is a PCI 1 slot with 33MHz 32bit. This is the very last one on the board, so that there will only fit a one slot card.

And while the box is open, see if the PSU can provide enough power and if it has additional PCIE power connectors. But this is going off topic now.

The PSU has 1kW and several spare 6pin 12V plugs labeled PCIE (at least six)

1GB has always been enough for me.

Thx

By the way, you keep talking about a GTX 760 Ti. I don't think those exist. What you seem to have in mind is the computing power of a GTX 760 at the price of a 750 Ti. You'd better rethink that. :-)

Sorry, of course I meant the 750ti as a single slot GPU (Galaxy GeForce GTX 750 Ti RAZOR)

So two questions appear:
Q1: Does it make sense to look for an old 32bit 33MHz PCI1 GPU on eBay or alike?
Q2: Are there PCIE 2.x x4 or x8 GPUs on the market? Do they make sense at all?

---

Kind Regards,
Horst

So two questions appear:
Q1: Does it make sense to look for an old 32bit 33MHz PCI1 GPU on eBay or alike?

No

Q2: Are there PCIE 2.x x4 or x8 GPUs on the market? Do they make sense at all?

There were those cards in the past, but you don't have to think about them at all.

Just get a modern 16x card. They can be installed in whatever slot they physically fit into. PCIe cards and slots are downwards compatible.

Use the 16x slot if possible. If there is not enough room to install the card into that slot (if the card requires additional space for cooler... ), then 8x slots will be fine too. There wouldn't be much difference in crunching speed with 16x or 8x slot. Here you can see some info what effect the bus speed really has:

http://tpucdn.com/reviews/Intel/Ivy_Bridge_PCI-Express_Scaling/images/perfrel.gif

The board is an ASUS Z8NR-D12 server board.

Here is a picture of that board at asus.com.

So it turns out you have one PCIEx16 slot, several smaller and one PCI. That makes the decision easy. Put a good GPU in the x16 slot and ignore the rest. Yes, there are PCIEx8 cards, even PCI. No, they are no good for crunching. I can tell, I've tried. The main reason is not the reduced link speed, at least not for PCIEx8 or even x4. But all those slots are limited to 25W as far as I know. Whatever you can put in there won't have much crunching power. You'll get more for your money if you choose your main GPU half a size larger.

The board is an ASUS Z8NR-D12 server board.

Here is a picture of that board at asus.com.

Here's the rest of the specifications: https://www.asus.com/Server-Workstation/Z8NRD12/specifications/

---

Yes, I had this thought about a single GPU also. I found the Radeon R9 280X(3GB GDDR5) to have a good performance to the 220 Euro price tag. At primegrid they show some performance data. But there might be a problem regarding the length of these huge PCIE*16 cards. They might interfere with the cooler of the second CPU. I have to use a ruler to figure out, if those huge cards will fit at all.

There are some x1 cards on the market like the Zotac GeForce GT 730 (1GB DDR3). But they only have a tenth of the number crunching abilities of the Radeon.

Q1: I have not found x4 or x8 GPU cards for the other slots. Are there some out there at all?

Q2: What has been the strongest old PCI GPU single slot card on the market? Perhaps I should look for a used one. If they bring 10-20% of the big one, why not. They have to be cheap.

---

Kind Regards,
Horst

Arrgghh!
Full length PCIE cards do only fit, if the second CPU is not present.

With this I have to decide to use the machine without any GPU at all or with only a few short x1-cards (50EUR each).

---

Kind Regards,
Horst

Yes, I had this thought about a single GPU also. I found the Radeon R9 280X(3GB GDDR5) to have a good performance to the 220 Euro price tag. At primegrid they show some performance data. But there might be a problem regarding the length of these huge PCIE*16 cards. They might interfere with the cooler of the second CPU.

R9 280 are not bad either and they are a bit shorter. I purchased some recently, preferred them over 280X for their size. I'm very happy with their performance at Einstein.
How much space do you have available?

There are some x1 cards on the market like the Zotac GeForce GT 730 (1GB DDR3). But they only have a tenth of the number crunching abilities of the Radeon.

Q1: I have not found x4 or x8 GPU cards for the other slots. Are there some out there at all?

Those GT 730 are available for PCIEx8. Passive cooling usually, so you can't put two side by side. But you wouldn't want to.

I have run a GT 630 that really is a rebranded GT 730 like that one. Had it in a x4 slot just like you're planning. It isn't worth it. Performance at Einstein was bad. Milkyway not good either. Could run GRUGRID short tasks but wouldn't finish long tasks before deadline even when running 24/7. If you can get one for free, try it, but don't pay money for it!

Q2: What has been the strongest old PCI GPU single slot card on the market? Perhaps I should look for a used one. If they bring 10-20% of the big one, why not. They have to be cheap.

I have seen GT 430, 520 and 610. Have a GT 430 myself. They're useless. If you pay shipping that's already too much.

Speaking of cheap, I can't tell about the second hand market, but usually special solutions are not cheap. Not at all.

Look for "Low Profile" GPUs when searching. Either that or "Half-Length".

I found these, looking for AMD only thus far:
XFX R7-240A-CLF2 Radeon R7 240 2GB
XFX One ON-XFX1-STD2 Radeon HD 5450 512MB
XFX HD-545X-ZQH2 Radeon HD 5450 1GB
The latter two are passively cooled, which could be a problem.

Also available lots of R5 GPUs low profile.
Or a search for "R7 Low Profile" on Newegg.

No R9s available in low profile model.

Lastly, and explanation on the difference between full-length, half-length and low profile.

Oh and if you need the manual, look here. Don't let the German name for the manual throw you, the PDF is in English.

---

I've never before seen the terms "full length" or "half length" applied to PCI Express, so I won't use them. Let me describe in a few short words what I have in mind: About 160mm long, full height, two slots wide. Whatever is smaller than that is 95% sure not worth its money when it comes to crunching. A bit simplified but size really does matter here. Mind you, I'm not talking high end. Those figures are characteristic in the GTX 750 (Ti) / R7 250E class, that's lower middle class. You can't expect a reasonable output/price ratio from anything lower than that. So my question is, is this space available. If it is not, I'd not think about a GPU for this box and make it a CPU cruncher instead.

(Edit: 250X -> 250E)

About 160mm long, full height, two slots wide.

If this motherboard is sitting in a server rack, he may not want full height.

@Horst, what are the dimensions of the case or rack the motherboard sits or goes in?

---

Thx for the good ideas. Thd board is in a 4U box. So full height is no problem. I think I will stay with the 2 CPUs only. If I am able to get a fitting 'half length' double slot GPU I will add it.

Good news is, that there is a second nearly identical machine with only one CPU, which will be available in even shorter time. And with this I will go for the R9 280X.

---

Kind Regards,
Horst