dazhuang001 发表于 2026-5-7 16:06
各位大佬，这个配置在这个价格是偏贵嘛？

正如sobereva老师和David_R所说，如果GPU不需要不建议买，vasp主要要求的是双精度，这时候rtx 6000 ada好像还不如5090，要不然就用个消费级游戏显卡，要不然就计算卡（v100？a800？不是很熟悉）
硬盘也是系统盘可以4tb，然后数据盘可以是hdd，Gaussian最好应该单独分出来一个盘做临时文件，不过似乎可能dft影响不大？可以参考http://sobereva.com/397
如果不打算扔自己旁边的话显示器不用那么好，但是这么有钱区别也不大，零头中的零头
PSU电源最好明确一下品牌，虽然问题不大吧

David_R 发表于 2026-5-7 15:51
The RTX 5000 Pro Blackwell GPUs are very expensive, and, to ensure it's worth paying the premium for ...

嗯嗯好滴，感谢大佬

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.

tonganlhy 发表于 2026-5-7 13:46
配置非常顶，但是显卡似乎没有必要，价格也算是比较高的，性价比一般。

感谢大佬

dazhuang001 发表于 2026-5-7 12:52
这个是一个供应商给出的报价，各位大佬能帮我看看这个配置和价格如何？
感谢各位大佬

配置非常顶，但是显卡似乎没有必要，价格也算是比较高的，性价比一般。

tonganlhy 发表于 2026-5-4 12:51
只要溢价不是太离谱，整机买还是可以接受的。一个配件一个配件买的问题是真有问题扯起皮来非常耗时耗力。

是的，确实有这方面的担忧

睡觉不打鼾 发表于 2026-5-3 12:36
都是组装的，应该没有真正意义上的“品牌机”，即使所谓的DELL什么的，毕竟他们的机器也是各种零件拼起来 ...

哦哦原来如此。感谢大佬的解释

这个配置很赞啊，有米之人，问价之后来分享一下啊，也在想攒机，就是动作慢了，

tonganlhy 发表于 2026-5-4 12:51
只要溢价不是太离谱，整机买还是可以接受的。一个配件一个配件买的问题是真有问题扯起皮来非常耗时耗力。

一般也是整机的，挑好店家其实溢价可以接受，主要还是单位采购的问题，大宗固资手续巨麻烦

睡觉不打鼾 发表于 2026-5-3 12:36
都是组装的，应该没有真正意义上的“品牌机”，即使所谓的DELL什么的，毕竟他们的机器也是各种零件拼起来 ...

只要溢价不是太离谱，整机买还是可以接受的。一个配件一个配件买的问题是真有问题扯起皮来非常耗时耗力。

dazhuang001 发表于 2026-5-2 22:08
好的，感谢大佬。店家提供详细的测试结果是指单个配件的嘛？我看淘宝上搜相关配件的时候，官方旗舰店都没 ...

都是组装的，应该没有真正意义上的“品牌机”，即使所谓的DELL什么的，毕竟他们的机器也是各种零件拼起来的，比如cpu是AMD或Intel的，主板是超微或技嘉的，内存是三星或者镁光或者海力士的。

也没有“官方旗舰店”的概念，因为cpu和主板、内存厂商几乎没有toC的业务。从这个角度，无论是品牌机还是组装机都是BtoC。当然，你愿意花更多的钱去买DELL什么的“品牌机”机器也完全没有问题，也许（注意只是也许）它们的主板、cpu、内存都是全新的，但花超过三倍的价格买性能差异不到10%的配件不是那么地合理。淘宝的各种卖服务器的资质都是够的，只要避雷那种做一两单就跑路的店家，正常的都不会有问题。

至于测试结果，就是在工作条件下不死机、重启等硬件层面的问题，这个需要做一些调试比如限制cpu电压等。现在很多做服务器组装的店家对各种专业计算软件都会基本的测试，实在不会的你提供软件和算例也会乐意帮你测试好再发货。搜肯定是搜不到的，搜也没有用，每台机器都不完全一样。

要是课题组没有人懂一点点硬件，那就忍痛花超过30%的溢价给“品牌机”厂商购买全方位的售后服务吧。当然我相信如果这笔钱中的一小部分发给学生做劳务，能做的比大多数所谓售后好一百倍。