NVIDIA DGX-1 / DGX-2 Maintenance
side by side.

DGX-1 · Pascal generation (older)

DGX-1 with 8x Tesla P100 (2016-2017)

DGX-1 · Volta generation

DGX-1 with 8x Tesla V100 16GB / V100 32GB (2017-2019)

DGX-2 · 16x V100 with NVSwitch

DGX-2 (16x V100 32GB, 2 PFLOPS, NVSwitch Gen 1)

Components (refurbishing pool)

P100/V100 GPU boards · NVSwitch Gen 1 · power supplies · mainboards

EOSL status of DGX Legacy generations

DGX-1/-2 maintenance is not trivial. Unlike standard server EOSL (Supermicro X9, F5 s-series, etc.), DGX hardware is proprietary — NVLink, NVSwitch, P100/V100 GPU boards with NVIDIA-specific firmware. Refurbishing pools for DGX-1/-2 are structurally thinner than for standard servers, and some critical components (NVSwitch Gen 1, early V100 boards) are scarcely available on the open market. NVIDIA offers standard support for DGX-1/-2 only limited or not at all — those who stay here often have no OEM service option. TPM is a bridge service solution with clear risk disclosures.

We recommend DGX-1/-2 TPM in two scenarios: (1) AI inference workloads that don't strictly need state-of-the-art hardware — DGX-1/-2 are technically still usable for many inference use cases, TPM extends productive lifespan economically. (2) Bridge service between DGX-1/-2 EOSL and a planned refresh phase to DGX H100/H200 or GB200 — TPM typically bridges 6-18 months until refresh hardware is available and installed. For very critical productive AI training workloads on DGX-1/-2 we recommend refresh instead of TPM bridge — modern DGX generations are 10-50x more powerful and energy efficient.

The full OEM-vs-TPM matrix (spare parts, SLA tiers, language, sub-outsourcing, compliance docs, reporting) is the same for every Nvidia line — see the Nvidia battle card.

BAIT/MaRisk documentation package, DORA Art. 28 contract template, ISO 27001 and GDPR-compliant media handling apply identically to every Nvidia line — details on the Nvidia battle card.

Same four steps for every Nvidia line: inventory → fixed-price quote in 48 h → contract (BAIT/MaRisk/DORA hooks) → hand-over with spare-parts provisioning. Details on the Nvidia battle card.

Which DGX Legacy models do you service?

DGX-1 and DGX-2 in all configurations: DGX-1 with 8x Tesla P100 (Pascal, 2016-2017), DGX-1 with 8x Tesla V100 16GB or 32GB (Volta, 2017-2019), DGX-2 with 16x V100 32GB and first generation NVSwitch (2018-2020). Including refurbishing pool coverage for GPU boards (P100, V100 16/32GB), NVSwitch Gen 1, power supplies and mainboards. Pool depth varies per component — we make status transparent before contract conclusion.

What does TPM cost for DGX-1 / DGX-2?

30 to 70 percent savings (where NVIDIA conditions still available at all). For a DGX-1 V100 32GB with 24×7×4, TechCare is typically roughly 55–60% below NVIDIA premium support (if NVIDIA still offers it); for the DGX-2 (16x V100) roughly 60%. For many DGX-1/-2 fleets NVIDIA no longer offers support — TPM is then the only service option.

How is component availability for DGX-1/-2 really?

We're honest: structurally weak but addressed with certified refurbishing sources. P100 GPU boards (DGX-1 Pascal) are rare on market today — we have pool coverage but for multi-board failures lead times are possible. V100 16GB/32GB boards (DGX-1 Volta, DGX-2) are still available in certified refurbishing sources. NVSwitch Gen 1 (only DGX-2) is the most critical component — on NVSwitch failure the DGX-2 can no longer operate in full topology, reduction to without-NVSwitch config is possible but performance loss significant. Power supplies and mainboards uncritically available. We create a risk assessment per DGX model and component before contract conclusion.

When should we refresh to modern DGX (H100/H200/GB200) instead of TPM?

Three clear refresh indicators: (1) Active AI training workloads (LLMs, computer vision, multi-modal) on DGX-1/-2 — modern DGX generations are 10-50x more powerful, ROI on refresh investment is typically under 12 months performance gain. (2) Energy efficiency: DGX-1/-2 have significantly higher power-per-performance than H100/H200 — with rising electricity prices refresh pays off faster. (3) AI software stack requirements: newer PyTorch/CUDA versions, newer Triton inference features, Transformer Engine only available on H100+. TPM is right choice for: inference workloads without state-of-the-art requirements, test/dev environments, bridge service until refresh hardware is available (often 6-18 months lead time for H100/H200/GB200).

Do CUDA and AI software continue to work on DGX-1/-2?

Yes, with restrictions. CUDA toolkit, GPU drivers for P100/V100 and standard AI frameworks (PyTorch, TensorFlow) continue license-free. CUDA versions are still supported by NVIDIA for older compute capability (P100=6.0, V100=7.0) for a while, but newest CUDA releases (12.x with Hopper optimizations) bring no advantage on P100/V100. Transformer Engine, FP8 operations and some newer Triton inference features are hardware-limited to H100+ — not available on DGX-1/-2. For standard AI inference and smaller training workloads, software stack coverage is sufficient.

Which SLA levels do you recommend for DGX Legacy?

Strongly depends on workload criticality: DGX-1/-2 in productive AI inference workloads (customer-facing pipelines): 24×7×4 with German-speaking onsite engineer is standard. For DGX-1/-2 as backup compute or in test/dev environments: 5×9 NBD typically sufficient. For DGX-1/-2 in refresh bridge phase (parallel to newly ordered H100/H200): Parts Only or 5×9 NBD sufficient — refresh hardware takes over productive load successively. For very old DGX-1 Pascal fleets running as pure lab servers: Parts Only with 14-day delivery time as most economical option.