On a node-per node basis, yes, but as a whole certainly not. Supercomputers usually use slower clocks an rely on parallelism at scale to provide throughput.
Nice video professor! I have one question. Do you think your cluster is more power efficient than, say, a rack mounted cluster like those in data centers, having the same computational performance as yours? Also, setting aside the power consumption, what about the price-to-performance relation, compared to a classical rack mounted data center cluster?
Just for fun, did you ever try to game on these workstations Professor ? (It's great cluster tho, would be interested to learn about the performance parameters in detail.)
What are you interested, specificly? I haven't tried gaming. I only have a mediocre GPU in the visualization node. My desktop, now, is a more capable beast. Anyway, for my next grant I'll build a similar cluster but this time with GPUs. Moving into AI territory.
Sir this informative sir please update aboute us HPC cluster how we can setup how mucn nood required and setup and for beginners which tools need to become HPC cluster engineer 😊
Excellent! The reason for only using CPUs is due to the advantage in accessing a greater amount of memory compared to a GPU? Or is there another reason? Best of luck with the simulations!
Genial. Disculpe Profesor, la Paralelización que se usa es nivel de ejecutar por ejemplo un modelo por cada nucleo, y asi poder correr varios modelos a la vez?. O por el contrario, a un solo modelo, descomponerlo y poder ejecutarlo de manera mas rapida?
Hola! Gracias por la pregunta! Ambas. En algunos casos queremos simular modelos "pequeños" miles de veces, por ejemplo para un estudio parametrico. Otras veces, simular modelos grandes que no caben en un solo computador.
Benchmarking is key. What is worrying is I/O mostly. The switches can handle 2.5GbE output from each compute node, but the raid storage server is getting fed through a 10GbE cables, so I can write from up to four nodes at full speed simultaneously. For my particular application, this setup has worked out great.
Ive been wanting to build a cluster of dual epyc 7742s to learn on and try to play around with data bots on the web to predict trends. Ive got 42 gpus i could throw into the leoop as well if i build my own cases
Amazing that this cluster probably has more performance than massive multi million dollar supercomputers from just a Decade ago
On a node-per node basis, yes, but as a whole certainly not. Supercomputers usually use slower clocks an rely on parallelism at scale to provide throughput.
this is so cool! the setup is crazyyy
It is pretty good for what I need.
Nice video professor! I have one question. Do you think your cluster is more power efficient than, say, a rack mounted cluster like those in data centers, having the same computational performance as yours? Also, setting aside the power consumption, what about the price-to-performance relation, compared to a classical rack mounted data center cluster?
I would love to see a detailed configuration vdo.
I’m want to build a cluster that can keep adding new node as my work load grows.
What specifically are you interested in?
Have u thought about using ubuntu server? Or u actually need a nice ui in the compute nodes?
Also what are the gpus model?
Yeah I thought. I don't need UI on the compute nodes. Honestly I just wanted to get it done. I'll learn to do a network install for the next cluster.
Just for fun, did you ever try to game on these workstations Professor ? (It's great cluster tho, would be interested to learn about the performance parameters in detail.)
What are you interested, specificly? I haven't tried gaming. I only have a mediocre GPU in the visualization node. My desktop, now, is a more capable beast. Anyway, for my next grant I'll build a similar cluster but this time with GPUs. Moving into AI territory.
Sir this informative sir please update aboute us HPC cluster how we can setup how mucn nood required and setup and for beginners which tools need to become HPC cluster engineer 😊
Excellent! The reason for only using CPUs is due to the advantage in accessing a greater amount of memory compared to a GPU? Or is there another reason? Best of luck with the simulations!
It's mostly because opensees doesn't run on GPUs, too much branching on complex nonlinear finite-element simulations.
Genial. Disculpe Profesor, la Paralelización que se usa es nivel de ejecutar por ejemplo un modelo por cada nucleo, y asi poder correr varios modelos a la vez?. O por el contrario, a un solo modelo, descomponerlo y poder ejecutarlo de manera mas rapida?
Hola! Gracias por la pregunta!
Ambas. En algunos casos queremos simular modelos "pequeños" miles de veces, por ejemplo para un estudio parametrico. Otras veces, simular modelos grandes que no caben en un solo computador.
Are these 2.5Gbps cables enough? I'm looking for building my own HPC but i'm concerned about bottlenecks
Benchmarking is key. What is worrying is I/O mostly. The switches can handle 2.5GbE output from each compute node, but the raid storage server is getting fed through a 10GbE cables, so I can write from up to four nodes at full speed simultaneously.
For my particular application, this setup has worked out great.
Ive been wanting to build a cluster of dual epyc 7742s to learn on and try to play around with data bots on the web to predict trends. Ive got 42 gpus i could throw into the leoop as well if i build my own cases
Good luck!
nice, but does it run doom?
😅 i'll make a video running doom on it
I've been wanting to do this video so bad!!
It would be boring, though. Since I only have one screen.
by the looks of I would assume you were a COMPUTER engineer, not a civil one 😂
I'm a huge all-around NERD!