I'm building an on-prem GPU cluster—what vendors should I compare?

For an on-prem GPU cluster, it helps to compare vendors across a few layers, not just the GPU card itself:

1) GPU/platform vendors

These are the core options for the accelerators and reference platforms:

• NVIDIA
  Dominant for AI/ML and HPC; strong software ecosystem (CUDA, NCCL, Triton, DGX systems).
• AMD
  Competes with Instinct GPUs; often worth comparing on cost/performance and open software stack (ROCm).
• Intel
  Less common for large-scale GPU training, but worth a look depending on workload and supply/pricing.

2) Server OEMs / integrators

These vendors build the actual cluster nodes and often provide design support:

• Dell Technologies
• HPE
• Lenovo
• Supermicro
• Cisco
• Inspur (availability depends on region/policy)
• Atos / Eviden (more HPC-oriented)
• Fujitsu (strong in some HPC markets)

3) Interconnect/networking vendors

Critical for multi-node training clusters:

• NVIDIA Networking / Mellanox (InfiniBand, Ethernet)
• Arista
• Cisco
• Juniper
• Broadcom-based ecosystem via OEM switches
• If HPC-focused, compare InfiniBand vs high-speed Ethernet offerings

4) Storage vendors

For data pipelines, checkpointing, and shared datasets:

• Pure Storage
• Dell PowerScale / EMC
• NetApp
• WekaIO
• VAST Data
• DDN
• Lustre-based HPC integrators

5) Cluster software / orchestration

Especially important if you’re running a private AI platform:

• NVIDIA Base Command / DGX software stack
• Red Hat OpenShift
• VMware
• Kubernetes distributions (e.g., Rancher/RKE2, OpenShift)
• Slurm for HPC scheduling
• Bright Cluster Manager
• SchedMD (Slurm commercial support)

6) If you want end-to-end “turnkey” comparisons

Consider comparing these system-level offerings:

• NVIDIA DGX (often the benchmark for AI cluster performance and support)
• Dell AI Factory
• HPE Cray / HPE ProLiant GPU systems
• Lenovo AI/HPC systems
• Supermicro GPU servers
• Lambda (AI-focused infrastructure)
• Pinecone? Not for on-prem hardware, so usually not in this category

What to compare beyond the logo

For each vendor, evaluate:

• GPU availability and roadmap
• Server density and power/cooling requirements
• Network fabric support (InfiniBand vs Ethernet)
• Storage architecture
• Driver/software maturity
• Support SLAs and on-site service
• Total cost of ownership
• Lead times and supply chain reliability
• Compliance/security requirements
• Proven references for your workload (LLM training, inference, HPC, VDI, etc.)

If you want, I can turn this into a vendor comparison matrix for your specific use case (e.g., LLM training, inference, CAD/VDI, or HPC) and budget range.