Discussion Thread for the Continuous GW Search known as O2MD1 (now O2MDF - GPUs only)

To answer my own question, I was running two on one GPU (as it got them done quicker - not waiting on the CPU so much).  I noticed the GPU RAM was full when running two at once (a WU needs 2.4GB, the card has 4GB), so it was using main system RAM, which presumably is a lot slower.  Changing it to run one at a time on the GPU has speeded it back up again.  I guess the WUs recently got large in RAM size.

Peter Hucker wrote:

To answer my own question, I was running two on one GPU (as it got them done quicker - not waiting on the CPU so much).  I noticed the GPU RAM was full when running two at once (a WU needs 2.4GB, the card has 4GB), so it was using main system RAM, which presumably is a lot slower.  Changing it to run one at a time on the GPU has speeded it back up again.  I guess the WUs recently got large in RAM size.

I see from your RX 560 host that those increased times were for a few tasks in the VelaJr data series, which have been known to bog down run times.  It's good to know though that changing task runs to 1x can resolve the issue.

For the O2MDFG3a_G34731 data series that my host (two RX 570s) is currently running, GPU memory of 4 GB doesn't appear to be a limiting factor for 2x runs on a single card.  Across different run configurations, here are my average GPU memory loads, sampled for 10 min with readings every 2 sec:

Run config.   GPU mem. load
1 GPU@1x = 25%
1 GPU@2x = 64%
2 GPU@1x = 20% across both cards
2 GPU@2x = 18% across both cards

So for 2x tasks on one card, there seems to be plenty of memory headroom.

Peter Hucker wrote:

Your numbers don't make sense, why were you getting no increase at all for 2 GPU@2x over 2 GPU@1x?  (I assume that means 2 GPUs with two WUs each).

Good question. That's the topic of my earlier post here where I was guessing that GPU utilization for 2GPU@2x (yes, 4 concurrent tasks) is throttled by limited CPU resources. If it's a matter of number of cores and their speed, then a major CPU upgrade should much improve daily task yields. If it's a matter of limited PCIe bandwidth, as Zalster suggested, then I (we) may be stuck because all the consumer Intel CPUs that I've looked at list  "Max # of PCI Express Lanes" as 16 and "PCI Express Configurations: Up to 1x16, 2x8, 1x8+2x4". Meaning that two GPUs will get no more than 8 PCIe lanes.  That hasn't mattered for FGRBPG1 work, which can get by fine with 2 lanes, but it MAY be a limiting factor for boosting GW productivity.

EDIT: I just saw over in Problems and Bug Reports that Gary Roberts has been running tasks at 3x on an RX570 with 12 min task times for non-VelaJr tasks. I'll give that a shot because, based on those times, it looks like a single card at 3x would increase daily task productivity over that of two cards at 1x.

... If it's a matter of limited PCIe bandwidth, as Zalster suggested, then I (we) may be stuck because all the consumer Intel CPUs that I've looked at list  "Max # of PCI Express Lanes" as 16 and "PCI Express Configurations: Up to 1x16, 2x8, 1x8+2x4".