Servizio · Physical servers on-premise
Physical bare-metal servers for critical workloads.
Despite the spread of cloud and virtualization, the bare-metal physical server remains the right choice for specific workloads: OLTP databases with very high concurrency, AI/ML training on GPUs, scientific HPC, and latency-sensitive applications that cannot tolerate hypervisor overhead. We work across all form factors (tower, rack 1U/2U/4U, blade) and the 5 main enterprise brands.
02 / When a physical server is the right choice
When the physical server is the right choice
- High-concurrency OLTP database: SQL Server, Oracle, PostgreSQL with direct storage I/O
- AI/ML training: NVIDIA H100/H200 GPUs with NVLink, LLM model training
- Scientific HPCsimulations, render farm, high-performance numerical computing
- Latency-sensitive applicationsalgorithmic trading, telco, VoIP, real-time analytics
- Workloads with socket-based licensing: some software has a licensing model that rewards physical hardware
- Compliance and isolation: regulated environments that require guaranteed single-tenant
03 / Form factor server fisico
Form factor server fisico
- Tower · ideal for offices, up to 25 users, quiet (e.g. HP ML, Dell T-series, Lenovo ST)
- Rack 1U · high density for data centers, mainstream cluster (HP DL360, Dell R660, Lenovo SR630)
- Rack 2U · mainstream workhorse with a density/expandability balance (HP DL380, Dell R760, Lenovo SR650, Fujitsu RX2540)
- Rack 4U / Convertible tower · maximum I/O, GPU and internal storage expandability
- Blade chassis · high-density multi-server in 5-10U (HP Synergy, Dell PowerEdge MX, Cisco UCS)
- HCI nodes · hyperconvergence with integrated storage (Nutanix, VxRail, Azure Stack HCI)
04 / Critical enterprise specifications
Critical enterprise specifications
- CPU: Intel Xeon Scalable 4th/5th gen or AMD EPYC Genoa/Turin, multi-socket
- RAM ECC Registered DDR5: error protection, scaling up to 8TB+ per server
- NIC: 1/10/25/100 Gbps, LACP bonding, dual out-of-band management port
- RAID hardware: dedicated controller with cache+BBU/Flash, NVMe RAID, levels 1/5/6/10
- Redundant hot-swap PSUs 80+ Platinum/Titanium with efficiency ≥94%
- BMC out-of-band: iLO (HP), iDRAC (Dell), XClarity (Lenovo), iRMC (Fujitsu)
Altri services
Explore the altri services AssistenzaServer.
05 / FAQ
Most frequent technical questions about physical servers.
Physical vs virtual server: when does bare metal still make sense?+
OLTP databases with very high concurrency (SQL Server, Oracle Exadata), AI/ML training on GPU (NVIDIA H100/H200), scientific HPC, latency-sensitive algorithmic trading, applications with socket-bound licenses. For everything else, virtualization wins for flexibility and cost.
Bare metal servers vs cloud (AWS EC2, Azure VM): TCO a 5 anni?+
For steady-state 24/7 workloads, an on-premise bare metal server typically costs 40-60% less than the equivalent cloud over 5 years. Cloud wins for variable workloads, burst, development, geographic distribution.
Tower server in ufficio: limiti pratici?+
Noise (a soundproof cabinet is recommended), heat (dedicated air conditioning), physical accessibility (security). For offices up to 25 users, tower units such as HP ML30/350 or Dell T550 are excellent. Beyond that, 1U/2U rack units in a closed cabinet are preferable.
NVIDIA H100/H200 GPUs for AI training: how to size them?+
Dual-socket server with Xeon Sapphire Rapids or AMD EPYC Genoa, 1-2TB DDR5 ECC RAM, 4-8 H100 SXM GPUs with NVLink, local NVMe storage + parallel filesystem (Lustre/BeeGFS), InfiniBand HDR/NDR networking for multi-node training.
Maintenance preventiva server fisici: cosa fare?+
Annual fan cleaning, SMART disk monitoring, BMC logs (iLO/iDRAC/XCC), periodic firmware updates, RAID controller BBU replacement every 3 years, preventive replacement of disks with > 50,000 hours.
Let's talk
Ready to talk with a systems engineer?
You will always be answered by a technician, never a call centre.
02 95 17 550