2

I'm looking for hardware configuration pieces of advice. I'm having 2 computers, one hyperconvergent server, and one workstation for which I'm planning the hardware. The workstation will be for dev and research. For this one, in a nutshell I'm having 2 hardware possibilities, A and B. I'm having 2 important factors: 1) ethernet speed 2) cuda compute. Other factors: system stability at factory frequencies and overclocked ones, extensibility.

A. Maximus XI Formula Z390, two 2080ti with NVlink, and a connexion to a hyperconvergent server through a 10Gbit (switch and server) ethernet connexion. The Formula had only 5Gbit connexion. The two 2080ti will have 8x and 8x connexion with the motherboard.

B. Maximus XI Hero Z390, two 2080ti with NVlink, one 10Gbit PCIE 4x for the connexion to the 10Gbit server. I think I'll have 8x and 4x connexion with the 2080ti, and 4x for the NIC 10 Gbit.

In a nutshell I have 5Gbit and 32 lines vs 10Gbit and 28 lines.

Questions:

  1. can you confirm the PCI speeds in the B case, as well as the deep learning hardware requirements ?
  2. which configuration would you recommend, and why ?

Many thanks.

[0] http://timdettmers.com/2018/12/16/deep-learning-hardware-guide/

  • Based on your current config, I'm guessing budget isn't an issue at all? Are you doing anything besides deep learning? Also, why two separate workstations? – JMY1000 Jan 6 '19 at 22:12
  • Budget is relatively limited (I wish I could go for a SkylakeX or even few DGX2). It's a 1 WS and 1 server (headless). The WS will be dedicated to deep learning, computer vision and software development. The server will hold a hyperconvergent platform (many VMs) based on Windows server 2019 datacenter; and it will need to stay connected 24/7. The WS should be able to be rebooted or crashed. – Soleil - Mathieu Prévot Jan 6 '19 at 22:21
  • Okay. What's your budget? Also, odds are very good that Zen 2 drops during CES (this week), so I'll probably wait until after then to give you a completed answer. – JMY1000 Jan 6 '19 at 22:23
  • I updated, It's for a 9900k based WS, so no Zen2 for this guy. I'm using loads of code optimized for Skylake and the Intel compiler. Indeed, I wish I could go for a 9940x for instance. – Soleil - Mathieu Prévot Jan 6 '19 at 22:26
  • @JMY1000 Budget is fixed since the price(Formula) = price(Hero+NIC 10Gbit). In the first case I got 5Gbit and 32 lines, in the second 10Gbit and 28 lines. That sums up the situation. – Soleil - Mathieu Prévot Jan 6 '19 at 23:16
1

The option A (Formula) is more interesting (and less expensive) since we got:

  • the 5Gbit LAN (~90 EUR for a NIC)
  • the mosfet watercooling (~50-100 EUR for the waterblocks) both for about 100 EUR instead of 140-190 EUR.
  • more 4 PCIE lanes, available for instance for another NVMe

Switching to Skylake-X is interesting when:

  • latency+speed host-device is important (eg., inference service)
  • when there is more than 3 discrete GPUs
  • when more than 64GB (128GB since z390) is needed
  • when bandwidth is critical (in memory bound compute tasks)
  • when there is a need of AVX512 (eg. computer vision), which has the advantage to provide low latency SIMD with floats.

Intel platform is more interesting than AMD's due to the HPC support and ecosystem (both hardware and software).

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.