Dear colleagues, would you be so kind to recommend a reasonable (price-wise) hardware configuration for a single workstation, looking forward to working with cryoSPARC for the next couple of years? We seek to maximize the number of GPUs, RAM, and processing cores.
Many thanks in advance for your input.
1 Like
Hi @ALBA,
I’m pretty sure you had a look to hardware and system requirements.
You can use a node as a master also, since the master is just the webserver and the database, I believe. The documentation says that for a node you need 4 cores per GPU and 64GB DDR4 per GPU, plus 1+ NVIDIA Tesla V100, RTX2080Ti, RTX3090, etc.
A good single workstation I believe will be having 4 cores and 256 GB DDR4 RAM, plus one GPU RTX3090. Double that (8 cores, 512 GB RAM) for 2 GPUs. I would add a 2 TB NVMA SSD configured as cache.
Unfortunately the price will depend on where you are… DELL can provide you with something like that on demand, but most of the commercial deep learning workstations also. Just check the specs 
Best,
Juan
3 Likes
Dear Juan, many thanks for this. Your reply is very useful in several aspects.
I understand now that 256 GB DDR4 RAM and 4x NVIDIA RTX cards are a staple looking forward.
However, I realize I posted the question a bit too simplistic regarding my goal, and apologize for that.
I am looking indeed for suggestions of maker+model for a motherboard and processor, compatible with the above, so I could buy the individual parts and custom setup the rig.
Thank you again.
Hi @ALBA
The problem with custom-rigs is that they may not be powerful enough (in watts) for what you want. For example this DELL 5820 tower has probably 2 PCIe slots for 2 GPUs but I’m sure it has no power (in watts & connections) to hold 2 RTX 3090.
Ideally you speak with your hardware provider about (or your shop) and buy the tower with a powerful PSU (power supply unit) and GPU power cables, and the maximum number of cores. Then later you will be able to buy the GPUs, a GPU power adapter (8 pin/6pin) and RAM modules (8/16/32 GB) according to your (expanding) needs.
I hope this helps! 
Best,
Juan
1 Like
Very good advice from @jucastil. The only thing I’ll add is make sure the GPUs you buy have at least 11 GiB of VRAM. The most common hardware related problem that we see is GPUs with not enough memory. The specific models mentioned by @jucastil all meet this requirement but if you decide to buy other cards, be sure to check this. Good luck!
–Harris
2 Likes
Hi All,
Is this post still active? I am interested too.
I am configuring a new workstation.
256Gb or 512Gb ?
Is it worth the extra money ?
-Gia
Hi Gia, it depends on your intended use. Some programs are RAM hungry, like polishing, relion’s implementation of motion correction, 3D classification, relion_reconstruct. These are CPU-based and not GPU based programs and they scale with MPI, so you need more cores to implement the ram. I have found cryoSPARC to not be RAM hungry. If you plan to get into tomography or need to use a large box with many particles then more ram becomes very helpful in speeding up these CPU-based programs with more MPI. With more RAM you can run more of these processes simultaneously as long as there is CPU available, so you could run relion’s motion correction in parallel with 3D classification.
Basically, if you plan to process multiple datasets simultaneously and implement these RAM hungry applications, then more RAM is very advantageous. If you plan to only use cryoSPARC on one or two datasets at a time, then I don’t think it is worth the very high cost.
Hi,
OK I see the point, having more RAM is a (costly) investment but that can be a good for longer term evolution of processing software and increasing size of datasets.
Can you me an advice for this configuration ?
2x Xeon 6248R processor
256Gb RAM (can be extended to 512)
2x Nvidia A5000 GPUs
1SSD system + 1SSD scratch
48Tb diskspace
Thanks,
GIA
Hi GIA, it sounds like a very good system. I am a novice in this department, but have talked to a few people happy with the A-series performance (A5000 and A6000). The ability to add more GPU cards and more RAM in the future is definitely a plus. One thing to note - you can optimize performance a bit by putting the operating system on a SATA SSD, rather than on one of the limited m2 NVMe SSD slots. This will allow you to use two (or three) m2 NVMe slots for scratch or movie processing. I was advised that putting the OS on the SATA will slow down boot but shouldn’t affect system performance once up and running. That’s not really a big deal. I would put the 48TB disk space in some sort of RAID configuration (RAID5, RAID6, or RAID10) as a low-level protection in case of drive failure
Hi
in fact there will be 5x12 Tb disks in RAID5.
actually what is a good amount of scratch space ?
1Tb is enough ? or should I put more ?
big NVME drives are quite expensive.
Thanks,
GIA
Hello @giax, I understand you want the 1 TB MVME as scratch.
I think it is too little, it will be full soon, specially if you don’t monitor constantly the storage. I would go for 2 TB, or 4 TB, depending on what type of data you analyse and how many people are going to use your installation.