The NVIDIA L40S has been a popular universal data center GPU for inference, fine-tuning, and media work since 2023. OpenMetal carries its effective successor instead — the RTX PRO 6000 Blackwell Server Edition — which roughly doubles GPU memory (96GB vs 48GB), adds Blackwell-native FP4, and raises memory bandwidth. This comparison covers what changes between the Ada-generation L40S and the Blackwell RTX PRO 6000 for training and inference, and why teams cross-shopping the L40S should look at the RTX PRO 6000. See the RTX PRO 6000 spec page for full details.
Key Takeaways
- 2x GPU memory: 96GB GDDR7 on the RTX PRO 6000 vs 48GB GDDR6 on the L40S — larger models and batches fit on a single card for both training and inference.
- Newer generation: Blackwell (RTX PRO 6000) vs Ada Lovelace (L40S) — the RTX PRO 6000 adds native FP4 (NVFP4) on top of FP8, for higher low-precision inference throughput.
- Higher memory bandwidth: the RTX PRO 6000’s GDDR7 delivers 1.6 TB/s vs the L40S’s 864 GB/s — relevant for memory-bound inference and training.
- L40S advantages: lower board power (~350W vs 600W) and typically lower cost and broader availability — a fit when 48GB is sufficient and power/density is the priority.
- Same OpenMetal model: whichever GPU, OpenMetal delivers it as single-tenant bare metal with fixed monthly pricing and included egress — no per-GPU-hour “idle silicon tax” on sustained training or always-on serving.
Ready to Compare GPUs for Your Workload?
Tell us your model sizes and throughput targets and we’ll help you choose between the RTX PRO 6000, the larger-memory H200, or a multi-GPU cluster.
Spec Comparison
| Specification | NVIDIA RTX PRO 6000 Blackwell SE (OpenMetal) | NVIDIA L40S |
|---|---|---|
| Architecture | Blackwell | Ada Lovelace |
| GPU Memory | 96GB GDDR7 | 48GB GDDR6 |
| Memory Bandwidth | 1.6 TB/s | 864 GB/s |
| Lowest-Precision Tensor | FP4 (NVFP4) | FP8 |
| Max Board Power | 600W | 350W |
| NVLink | Not supported (two discrete cards) | No NVLink |
| Carried by OpenMetal | Yes — available now (Ashburn) | No |
*Both GPUs train and infer; the RTX PRO 6000 is the newer, larger-memory option, while the L40S is a lower-power Ada card.
GPU Generation: Blackwell vs Ada
The RTX PRO 6000 is a Blackwell-generation GPU; the L40S is Ada Lovelace (the prior generation). The headline functional difference is FP4 (NVFP4), a Blackwell-native 4-bit format that roughly doubles low-precision inference throughput over FP8 on supported stacks. For mixed-precision training (BF16/FP8) both are capable, but the RTX PRO 6000’s newer tensor cores and larger memory give it more headroom per card. Where the L40S still appeals: it draws less power (~350W) and is widely available at a lower price point, which can matter for dense, power-constrained inference fleets where 48GB per card is enough.
GPU Memory
Memory is the clearest separator:
- Capacity: 96GB vs 48GB. The RTX PRO 6000 holds roughly double the model and batch size per card. For training and fine-tuning, that means fewer GPUs to hold the same model state; for inference, larger KV caches and batch sizes.
- Bandwidth: GDDR7 vs GDDR6. The RTX PRO 6000’s GDDR7 delivers 1.6 TB/s vs the L40S’s 864 GB/s, which helps both memory-bound inference and training throughput.
Neither card uses HBM — for the highest memory bandwidth (largest-scale training), OpenMetal’s H200 with 141GB HBM3e is the step up. The RTX PRO 6000 sits between the L40S and the H200: more memory and a newer architecture than the L40S, at a lower cost than HBM-class cards.
Host Platform and Networking
On OpenMetal, the RTX PRO 6000 runs on a single-tenant bare metal host with dual Intel Xeon 6530P (64C/128T), 1TB DDR5-6400, and a 6.4TB Micron 7500 MAX NVMe data drive, with 20 Gbps private (up to 40 Gbps optional) and 10 Gbps public bandwidth. Full root access, IPMI, no hypervisor overhead, and included east-west traffic apply to every OpenMetal GPU server.
Security and Confidential Computing
Both GPUs run as single-tenant bare metal devices on OpenMetal — physical isolation, no shared hypervisor on the accelerator. Intel SGX is available on the host CPU. On the RTX PRO 6000, Intel TDX with single-GPU confidential passthrough is substantiated by NVIDIA’s confidential-computing documentation (one GPU per VM), with OpenMetal validation pending; it is delivered as an engineered build, not a self-serve toggle. The RTX PRO 6000 host is deployed in the HIPAA-compliant Ashburn (NTT DATA VA1) facility; OpenMetal offers BAAs at the organizational level.
When the RTX PRO 6000 Wins — and When an L40S-Class Card Suffices
When the RTX PRO 6000 is the right choice
- Training, fine-tuning, or serving models that exceed 48GB on a single card
- Inference where Blackwell FP4 throughput is a meaningful gain
- Workloads benefiting from higher GDDR7 memory bandwidth
- Consolidating multi-L40S deployments onto fewer, larger-memory cards
When an L40S-class card suffices
- Inference and fine-tuning that fit comfortably in 48GB
- Power- or density-constrained fleets where ~350W per card matters
- Cost-sensitive deployments where the lower-priced Ada card is sufficient
Cost and Value
OpenMetal prices the RTX PRO 6000 on a fixed monthly model with included egress — no per-GPU-hour metering. Because OpenMetal carries the RTX PRO 6000 (not the L40S), the practical decision is whether the RTX PRO 6000’s extra memory, bandwidth, and FP4 justify it over a smaller Ada card for your workload. For sustained training and always-on inference — high-utilization workloads — the fixed-cost dedicated model avoids the metered-cloud “idle silicon tax,” where every GPU-hour bundles elasticity premium you may not need. OpenMetal does not publish RTX PRO 6000 pricing; contact OpenMetal for a custom quote.
Ready to Compare GPUs for Your Workload?
Tell us your model sizes and throughput targets and we’ll help you choose between the RTX PRO 6000, the larger-memory H200, or a multi-GPU cluster.
Deployment Options
- Dedicated GPU server — a single RTX PRO 6000 (or dual-GPU) bare metal server with full root access and IPMI.
- Dedicated GPU cluster — multiple GPU nodes (all-RTX PRO 6000 or mixed with H200) on a private mesh.
- Attached to existing infrastructure — add RTX PRO 6000 nodes to an existing OpenMetal Hosted Private Cloud or bare metal deployment.
Where to deploy
The RTX PRO 6000 is available now in Ashburn, Virginia (US-East), with advance reservations available for Los Angeles, Amsterdam, and Singapore. Proof of Concept clusters are available for testing; ramp pricing is available for migrations from other providers.
Get an RTX PRO 6000 Quote
Ready to deploy? Tell us about your AI/ML training or inference needs and we’ll provide a custom quote for the NVIDIA RTX PRO 6000 — as a single GPU server, a dedicated GPU cluster, or GPU nodes attached to an existing OpenMetal deployment.
- Single GPU server: One or two RTX PRO 6000 cards with full root access and IPMI
- GPU cluster: Multi-node deployments (all-RTX PRO 6000 or mixed with H200) on a private mesh
- Attached GPU: Add RTX PRO 6000 capacity to your existing Hosted Private Cloud or bare metal footprint
All deployments include fixed monthly pricing, included egress, a 99.96%+ network SLA, and DDoS protection.
Related OpenMetal Answers
Product specifications, pricing, and availability may change due to market conditions and other factors. For the most current information, please contact the OpenMetal team directly.