|
|
本帖最后由 David_R 于 2026-5-7 15:52 编辑
The RTX 5000 Pro Blackwell GPUs are very expensive, and, to ensure it's worth paying the premium for these cards, it's a good idea to double check that your workflows require the specifications they have over less expensive options. The main ones here are the two-slot cooler design (so they'll fit the GPU server form factor), the 72GB VRAM capacity, and ECC memory.
For many GPU-accelerated scientific computing applications, double precision is required, and this generation of workstation/professional class GPUs do not offer any huge advantages over the consumer class cards—you'll have to look at datacenter class GPUs for big improvements in performance here.
These days it's fairly straightforward to get modified consumer class cards with two-slot blower coolers, which offer great value for money when building GPU servers. I have quite a few of these in my cluster and have had good experiences with them, but I can understand why one might prefer to go the more official route.
The large VRAM capacities of these cards are really only useful for training or inferencing large ML models (like LLMs, at a scale much larger than is typically encountered for scientific applications), or MD simulations with a very large number of atoms. There are of course exceptions in this rapidly developing field, but I haven't come across any examples that exceed the typical VRAM capacities of consumer class cards (16-32 GB).
Also bear in mind that GPU parallelisation is also a challenging thing to get working properly, and I presume the two specified GPUs will be communicating via regular PCI-e (4.0?) connectivity, which can easily bottleneck some parallel code implementations. I wasn't entire sure what your immediate need for GPU acceleration was, but perhaps you could look at installing a single, inexpensive GPU first, test your workflows, and add more powerful accelerators later?
The other thing that I could foresee as a potential issue in this configuration is the relatively small capacity of fast storage. I couldn't tell whether these are regular SATA SSDs or M.2/U.2 NVMe drives, but in any case, I would want to configure more than 480GB of usable fast storage (especially as much of this will be taken by the OS) - in fact this is even less than the specified RAM capacity! Using the 16TB spinning HDD for continuous random read/writes in various workflows will be way too slow. I strongly recommend having a 4TB U.2 NVMe drive for scratch storage. |
|
楼主 Author
发表于 Post on