We will conclude the current high-frequency search to focus on a lower frequency range with a new parameter setup. The app will stay the same, GPU memory requirement shall be around 1800MB. We'll switch to the new workunits ASAP.
We will conclude the current high-frequency search to focus on a lower frequency range with a new parameter setup. The app will stay the same, GPU memory requirement shall be around 1800MB. We'll switch to the new workunits ASAP.
Is it to be expected, that the new WUs need a lot more memory bandwidth? Even with all CPU tasks disabled I see DRAM Read Bandwidth in HWiNFO, that I've never seen before, basically it's nailed to ~38 Gbps, if I enable 4 CPU tasks I get even spikes over 40 Gbps. With the old tasks and 14 MCM WUs on the CPU cores it was somewhere around 32-34 Gbps with spikes to around 36 Gbps and the iGPU could complete a task in about 1 hour and 20-25 minutes, the estimate for my first new task is around 4-5 hours.
Is this the expected increase in runtime, or are the new tasks not very suitable for iGPUs with their limited memory bandwidth?
The setup is quite different. We split the original WUs in half and processed these one after the other, because that reduced the memory consumption. For the new WUs this isn't necessary (and actually more difficult). Also too the "recalc" stage at the end (mostly done on the CPU) is much shorter compared to the total runtime. I'm not entirely sure whether the new WUs run slower or faster on a specific machine, by how much and what properties influence that. On our machines the runtimes of the old an new workunits are comparable. So far I don't have enough results for a statistical evaluation.
The setup is quite different. We split the original WUs in half and processed these one after the other, because that reduced the memory consumption. For the new WUs this isn't necessary (and actually more difficult). Also too the "recalc" stage at the end (mostly done on the CPU) is much shorter compared to the total runtime. I'm not entirely sure whether the new WUs run slower or faster on a specific machine, by how much and what properties influence that. On our machines the runtimes of the old an new workunits are comparable. So far I don't have enough results for a statistical evaluation.
Thanks for the info. So, just to make sure I am understanding correctly, after the WU is initially loaded into the VRAM, there is no recalc step until the very end of the task?
The setup is quite different. We split the original WUs in half and processed these one after the other, because that reduced the memory consumption. For the new WUs this isn't necessary (and actually more difficult). Also too the "recalc" stage at the end (mostly done on the CPU) is much shorter compared to the total runtime. I'm not entirely sure whether the new WUs run slower or faster on a specific machine, by how much and what properties influence that. On our machines the runtimes of the old an new workunits are comparable. So far I don't have enough results for a statistical evaluation.
Thanks for the info. So, just to make sure I am understanding correctly, after the WU is initially loaded into the VRAM, there is no recalc step until the very end of the task?
yes that sounds like it, and that recalc step will be much faster than before.
I might need to revisit the 1.14 app again to see how that does.
FYI, you can differentiate the high frequency tasks from the low frequency ones in your task list by looking at the task name, high frequency have "HF" and freq numbers >1500Hz, and low freq have "Bu" in the name and freq values around 100-200Hz.
Bernd, are you guys abandoning the HF search altogether? or just refocusing on something else for now. can you remind us which frequency ranges have been searched so far and what ranges we can expect from the LF search?
If we had Elon's help I think
)
If we had Elon's help I think the project would be complete in a matter of days, leaving us nothing to do. I am currently working on 1680.00Hz
We will conclude the current
)
We will conclude the current high-frequency search to focus on a lower frequency range with a new parameter setup. The app will stay the same, GPU memory requirement shall be around 1800MB. We'll switch to the new workunits ASAP.
BM
Bernd Machenschalk wrote: We
)
Thanks for the heads-up, Bernd. Much appreciated!
Proud member of the Old Farts Association
Thanks for the update,
)
Thanks for the update, Bernd
Much appreciated!
Crunching O3 on the iGPU of
)
Crunching O3 on the iGPU of my Ryzen 5700G...
Is it to be expected, that the new WUs need a lot more memory bandwidth? Even with all CPU tasks disabled I see DRAM Read Bandwidth in HWiNFO, that I've never seen before, basically it's nailed to ~38 Gbps, if I enable 4 CPU tasks I get even spikes over 40 Gbps. With the old tasks and 14 MCM WUs on the CPU cores it was somewhere around 32-34 Gbps with spikes to around 36 Gbps and the iGPU could complete a task in about 1 hour and 20-25 minutes, the estimate for my first new task is around 4-5 hours.
Is this the expected increase in runtime, or are the new tasks not very suitable for iGPUs with their limited memory bandwidth?
.
I also see much longer
)
I also see much longer runtimes on the new tasks.
and it looks like they are no longer running in two stages, but rather one long stage.
Bernd, is that intended?
I suppose the validator isnt setup yet either
_________________________________________________________________________
Looks like work being sent
)
Looks like work being sent out is paused?
I watched a few of the new work units come through- definitely take longer. Still saw a pause at 49.5% and then at 99.0%.
EDIT: Just watched one NOT pause at 49.5%
Perhaps the one that I saw pause at 49.5% was from the previous generation.
The setup is quite different.
)
The setup is quite different. We split the original WUs in half and processed these one after the other, because that reduced the memory consumption. For the new WUs this isn't necessary (and actually more difficult). Also too the "recalc" stage at the end (mostly done on the CPU) is much shorter compared to the total runtime. I'm not entirely sure whether the new WUs run slower or faster on a specific machine, by how much and what properties influence that. On our machines the runtimes of the old an new workunits are comparable. So far I don't have enough results for a statistical evaluation.
BM
Bernd Machenschalk
)
Thanks for the info. So, just to make sure I am understanding correctly, after the WU is initially loaded into the VRAM, there is no recalc step until the very end of the task?
Boca Raton Community HS
)
yes that sounds like it, and that recalc step will be much faster than before.
I might need to revisit the 1.14 app again to see how that does.
FYI, you can differentiate the high frequency tasks from the low frequency ones in your task list by looking at the task name, high frequency have "HF" and freq numbers >1500Hz, and low freq have "Bu" in the name and freq values around 100-200Hz.
Bernd, are you guys abandoning the HF search altogether? or just refocusing on something else for now. can you remind us which frequency ranges have been searched so far and what ranges we can expect from the LF search?
_________________________________________________________________________