I have been using CUDA for a few weeks now and have noticed that the granted credit is only about 75% of the claimed credit.Have I missed something.
Are hosts that use CUDA being penalised? or is there someting more significant that I am missing.
Setup is an amd9650 quad and 2X nvidia 9600gt's. I seem to have a problem getting my average above 3000. Nothing else running for 18 out of 24 hrs.
Dsve

---

The WU processed with CUDA appear to be over claiming. You are receiving the credit claimed by your wingman with a CPU processed WU.

OK understood

---

'Overclaiming'?

Or maybe we see in future that the CUDA-WUs will get the higher Credits?
Because the GPU do more FLOPS than the CPU for the same WU..

---

sounds more realistic.

---

'Overclaiming'?

Or maybe we see in future that the CUDA-WUs will get the higher Credits?
Because the GPU do more FLOPS than the CPU for the same WU..

CUDA may do more work in less time, but the amount of work done on an individual WU is the same. Just as the optimize apps are faster than the generic apps, its all the same analysis and results or they wouldn't validate. You get more credits per hour but the per job rate remains the same.

---

Boinc V7.2.42
Win7 i5 3.33G 4GB, GTX470

'Overclaiming'?

Or maybe we see in future that the CUDA-WUs will get the higher Credits?
Because the GPU do more FLOPS than the CPU for the same WU..

CUDA may do more work in less time, but the amount of work done on an individual WU is the same. Just as the optimize apps are faster than the generic apps, its all the same analysis and results or they wouldn't validate. You get more credits per hour but the per job rate remains the same.

AB is right.

---

'Overclaiming'?

Or maybe we see in future that the CUDA-WUs will get the higher Credits?
Because the GPU do more FLOPS than the CPU for the same WU..

CUDA may do more work in less time, but the amount of work done on an individual WU is the same. Just as the optimize apps are faster than the generic apps, its all the same analysis and results or they wouldn't validate. You get more credits per hour but the per job rate remains the same.

AB is right.

Hmm.. really?

Hmm.. the current credit system is with FLOPcounter.

The GPU-FLOPcounter counting more FLOPS than the CPU-FLOPcounter for the same WU.
So the GPU app make more work as the CPU app for the same WU.
So why not reward the GPU because of more work she has done?

---

But how can the G.P.U. make more FLOP'a, then the C.P.U., for the same WU?
Something don't add up here . . .
And, NO, credit is NOT simply the amount off (M/G)FLOP's.
Because, there are more factors responsible for the final result, a correctiom-factor is used, for each host, look at your own host, except IP external and internal adress, more is visible, for your eyes only { :) }. Also the correction factor, ;)
Size is usually < 1, (0,1001 on one of my hosts).

---

I'll try to simplify it a bit. Since my electronics goes back to vacuum tubes, this is just a generalization of how I perceive it and I stand to be corrected and others, I'm sure can give a more precise explanation.

The credit is calculated only in a general way to the actual FLOP count. At a higher level we could say 2*2=4 is an operation. Different CPU require more or less movements (FLOPs) of bits to perform this function. We get paid for the number of operations that are performed not for moving bits around inside the computer.

The correction factor you see in your account is just used to indicate how fast you process work compare to a gold standard, and is used to estimate how much work you have on hand and how many seconds of work to fetch to replenish your cache. It has nothing to do with credits.

I'll try to simplify it a bit. Since my electronics goes back to vacuum tubes, this is just a generalization of how I perceive it and I stand to be corrected and others, I'm sure can give a more precise explanation.

The credit is calculated only in a general way to the actual FLOP count. At a higher level we could say 2*2=4 is an operation. Different CPU require more or less movements (FLOPs) of bits to perform this function. We get paid for the number of operations that are performed not for moving bits around inside the computer.

The correction factor you see in your account is just used to indicate how fast you process work compare to a gold standard, and is used to estimate how much work you have on hand and how many seconds of work to fetch to replenish your cache. It has nothing to do with credits.

Same generation, but vacuum tubes still in use, saw new radio a few months ago. Uses TWT for power output in transmitter and lighthouse tubes at front end of receiver, to be used where they suffer lots of electrical storms.

But back to topic, situation I believe is similar to RISC Vs CISC. Google for explanation.

'Overclaiming'?

Or maybe we see in future that the CUDA-WUs will get the higher Credits?
Because the GPU do more FLOPS than the CPU for the same WU..

CUDA may do more work in less time, but the amount of work done on an individual WU is the same. Just as the optimize apps are faster than the generic apps, its all the same analysis and results or they wouldn't validate. You get more credits per hour but the per job rate remains the same.

AB is right.

Hmm.. really?

Hmm.. the current credit system is with FLOPcounter.

The GPU-FLOPcounter counting more FLOPS than the CPU-FLOPcounter for the same WU.

You're confusing a more efficient architecture that can perform more FLOPs to mean that it gets more credit. Remember that your credit is based on a unit of time. The fact that your GPU can perform more FLOPs than your CPU simply means that your GPU should be able to complete more work in the same amount of time, earning it a higher RAC, not more credit per workunit.

Imagine we both worked in a kitchen, and we were paid per plate to serve, no matter what is on the plate, as long as the end result of what is supposed to be on the plate is there.

Imagine that I am able to complete 5 plates per hour (I'm slow). Imagine that you are able to complete 50 plates per hour. You'd earn more simply because you're faster, not because you prepare the plates the same as I do, since we are required to come up with the same result.

So the GPU app make more work as the CPU app for the same WU.
So why not reward the GPU because of more work she has done?

The GPU doesn't "make" more work, in fact the work is the same because it has the same workunit. It is rewarded by it performing more work per unit of time, not for coming up with the same answer.

---

In past SETI@home needed to reduce the Credts/WU.

Because to be protected for Credit-Wars with other projects.

All projects -> same Credits


So something new at SETI@home.. CUDA..


A GTX260 would have a RAC of ~ 10,000 at GPUGrid.
Here at SETI@home maybe a RAC of ~ 8,000.

So if every project would give the same Credits then GPUGrid need to reduce or SETI@home need to rise the Credits/WU on GPU.

There are lot of people out there which crunch only for Credits.
And if they get more Credits at GPUGrid, they will not crunch here at SETI@home with their GPUs.
So.. SETI@home must be 'smart' to hold- or recruitment new members..

---

SETI@Home has been very considerate of keeping the value of a Cobblestone close to the original standard. Taking this into consideration, that would mean GPUGrid would need to lower theirs because they are granting too high.


...and if it came down to it, I'd rather see those who are only in it for the credits leave SETI than for SETI to participate in any sort of credit inflation just to keep users. Since SETI has plenty of dedicated crunchers who are in it for the goal, and those in it for credits are in the minority, I doubt it would hurt SETI too much for them to leave - it would simply take us longer to sift through data which is already older than anyone alive anyway.

---

I totally agree with OzzFan. The credit hound may be very vocal, but are only a small fraction of the 180,000 active Seti crunchers.

Seti is the gold standard by which other projects are suppose to calibrate their credits. A few projects (mostly still in Alpha development) currently have exaggerated credit levels. From dialog on the projects mailing list some already are starting to reduce credits to be more in line with Seti.

...
Hmm.. the current credit system is with FLOPcounter.

The GPU-FLOPcounter counting more FLOPS than the CPU-FLOPcounter for the same WU.
So the GPU app make more work as the CPU app for the same WU.
So why not reward the GPU because of more work she has done?

The stock 6.03 CPU app has various routines for doing many of the more difficult operations, on x86 systems there are 12 different Transpose functions, for instance. The FLOPcounter does not vary depending on which function is chosen, instead it 'counts' floating point operations which would be used if the most generic function were chosen. That's been true for all the setiathome_enhanced apps, the intention is that all hosts doing the same work will claim very nearly the same credits. Optimized apps use the same system for 'counting' to match claims.

For the 6.08 GPU app the count is meant to be the same, but all the 'counting' is done in the CPU portion of the code and that doesn't know all the factors needed. So there are some estimates involved which have not yet been tuned to get a close match.

Rewarding an app for doing extra work or penalizing an app for doing work more efficiently is not what the FLOPcounter is for.

                                                                 Joe

wondering what the credit is for?

wondering what the credit is for?

First, welcome to the Seti boards.

It is just a measure of how much work your computer(s) have done for projects that use BOINC. It should, but isn't, be equal for all projects.

The formula(s) used to calculate credit, use the number of Floating Point operations used to process a task.

If you have lots of time on your hands and are very, very bored you might search for things like cross project parity, but I wouldn't advise that you do.

The WU processed with CUDA appear to be over claiming. You are receiving the credit claimed by your wingman with a CPU processed WU.

So if my wingman has processed his WU with cuda too, then we will receive both higher credit that if my wingman is a CPU-cruncher?

In other words. The credit I get for a WU can vary according to the wingman I have?

The WU processed with CUDA appear to be over claiming. You are receiving the credit claimed by your wingman with a CPU processed WU.

So if my wingman has processed his WU with cuda too, then we will receive both higher credit that if my wingman is a CPU-cruncher?

In other words. The credit I get for a WU can vary according to the wingman I have?

At the moment that is the case. I'm sure that the dev will eventually recalibrate things to rebalance the credits.