I've been playing around with various distributed compute projects in my office, and one thing I've noticed is that some of them really like their memory bandwidth and cache - and run poorly without it available. Plenty run purely in cache or nearly so, but CPDN is one that's really hard on the memory bandwidth (doing global climate models, I cut my Linux admin teeth partly maintaining a cluster doing that sort of stuff years ago).
For reasons I fail to understand, the whole "big L4 cache" thing hasn't really taken off - the handful of eDRAM chips Intel has released. Mostly for laptops, but there were a few that came out as desktop chips, and the late Broadwell era one is particularly promising looking for memory-heavy workloads. Onboard GPU (so no need for an external GPU on the board), 128MB eDRAM, 4C/8T, 12MB L3, 128MB L4 (effectively).
Fine, they're cheap on the used market.
I have one of these beasties on order (used) with associated hardware to make it run. I'm going to set it up as a little BOINCbox and do some testing of performance on the gnarly workunits I like to run, as well as some others.
One thing I've noticed is that on a lot of systems, especially with hyperthreading, full utilization of the system gets no more instructions retired per second than a reduced utilization - so I'm debating writing some sort of automatic system that fiddles with workunits to optimize for total system throughput. In a rather entertaining case, my old Xeon box, with some totally-fit-in-L3-cache workunits, retires almost exactly the same number of instructions per second (60G) with 12 threads or 24 threads - 24 threads only has a 10% boost to 66G instructions per second.
My office has an obscene amount of power now (I added another 8 panels in the fall, so I now have 5.2kW of solar on my shed), and I'm working towards being able to make better use of this power this summer. I'll probably be putting some of the compute outside as well - I've got a few things I want to try on that front, though I keep waffling between "adding vents to my office and venting indoor computers outside" vs "just putting the computers outside." I do like the heat in the winter, but I really just need to get the heat outside. If I could move all my compute outside in the summer, this would be really convenient for me.
Anyway, any experiences with the eDRAM based chips for compute? They seem like they should be really good at it.
For reasons I fail to understand, the whole "big L4 cache" thing hasn't really taken off - the handful of eDRAM chips Intel has released. Mostly for laptops, but there were a few that came out as desktop chips, and the late Broadwell era one is particularly promising looking for memory-heavy workloads. Onboard GPU (so no need for an external GPU on the board), 128MB eDRAM, 4C/8T, 12MB L3, 128MB L4 (effectively).
Fine, they're cheap on the used market.
I have one of these beasties on order (used) with associated hardware to make it run. I'm going to set it up as a little BOINCbox and do some testing of performance on the gnarly workunits I like to run, as well as some others.
One thing I've noticed is that on a lot of systems, especially with hyperthreading, full utilization of the system gets no more instructions retired per second than a reduced utilization - so I'm debating writing some sort of automatic system that fiddles with workunits to optimize for total system throughput. In a rather entertaining case, my old Xeon box, with some totally-fit-in-L3-cache workunits, retires almost exactly the same number of instructions per second (60G) with 12 threads or 24 threads - 24 threads only has a 10% boost to 66G instructions per second.
My office has an obscene amount of power now (I added another 8 panels in the fall, so I now have 5.2kW of solar on my shed), and I'm working towards being able to make better use of this power this summer. I'll probably be putting some of the compute outside as well - I've got a few things I want to try on that front, though I keep waffling between "adding vents to my office and venting indoor computers outside" vs "just putting the computers outside." I do like the heat in the winter, but I really just need to get the heat outside. If I could move all my compute outside in the summer, this would be really convenient for me.
Anyway, any experiences with the eDRAM based chips for compute? They seem like they should be really good at it.
