Charlie StrossAnd now for the world's silliest laptop: weighs 25kg, has up to six 24" screens, 192 CPU cores, 6Tb of RAM, and two GPU cards: the a-X2P mobile media workstation! https://www.mediaworkstations.net/systems/portable-amd-workstations/a-x2p/
Graham Sutherland / Polynomial@cstross if you're an LD, lugging that plus a GrandMA node is probably less weight and bulk than shipping a GrandMA full-size lighting console in a flight case. possibly even costs less too. and you can run additional stuff like ELM / Madrix for LED walls on the same physical system.
Janne Moren@gsuberland @cstross
Dual Epyc 9654 CPUs would mean 192 cores - but clocked relatively slowly (boost clock is 3.55Ghz).
They're great CPUs for HPC or running lots of VMs in a data center. I'm not sure this is the configuration of choice for any kind of media, though; fewer but faster cores with more memory bandwidth per core should be a much better fit?
@jannem @cstross exactly the opposite. video stuff is inherently chunkable, hence why GPUs (thousands of "cores", clocked slow) are so good at it. I run dual Xeon 8276L on my workstation for 3D renders and video rendering work. for LD stuff, solving pattern and effect parameters to channel outputs is easily parallelisable, and the same goes for pixelmapper stuff.
@jannem @cstross the core count is massive overkill for something like MA OnPC or MagicQ, but realtime VJ and pixelmapping is very demanding. using this as an all-in-one station would let you do all that and live broadcast transcodes in a single compute node, which is extremely useful if you're a live production contractor.
@gsuberland @cstross
I was mostly thinking that video is inherently I/O intensive - few OPS per memory access - so you won't be able to keep the caches filled for all those cores. Your threads will end up stalling waiting for memory.
I was mostly thinking that video is inherently I/O intensive - few OPS per memory access - so you won't be able to keep the caches filled for all those cores. Your threads will end up stalling waiting for memory.
@jannem @cstross mostly irrelevant because you wouldn't need all those cores just for video processing, and even if you did it wouldn't need to fully saturate the IMC, but with 12 memory channels per chip populated with DDR5 4800, and assuming a (fairly high) all-core clock of 2.5GHz, *and* somehow literally zero of the dcache lines being hits, you'll still pull in 2 bytes per clock tick per core. so in practice it's not really a limitation.
@jannem @cstross ultimately the huge core count is overkill for any of the tasks individually, but if you're being paid big money to run a live show you don't want a single unexpected background task hitting a core or two to cause instability, especially if you're using it as a turnkey solution for a range of tasks (light show, laser show, pixelmapping, VJing, livestreaming) all at once.