Markdown

RTX PRO 6000)¶

Scope: NVIDIA's consumer (GeForce RTX 50/40) and professional workstation/server (RTX PRO 6000 Blackwell, RTX 6000 Ada, L40S) GPUs, and how they differ operationally from datacenter SXM parts. The point of this page is the differences that bite in a cluster: drivers and their licensing, the absence of NVLink, MIG/vGPU/ECC availability, power connectors, and cooling. Specs shift with each datasheet revision; re-check the cited NVIDIA pages before relying on a single number.

Figures verified against NVIDIA product pages, datasheets, and licence text as of June 2026. Consumer board-power and connector behaviour are safety-relevant; verify against the current product page and PSU guidance before deployment.

What it is¶

Two distinct product lines share the same silicon generations as the datacenter parts but are built, driven, licensed, and cooled for a different job:

GeForce RTX (consumer): RTX 5090 (Blackwell GB202), RTX 4090 (Ada AD102). Gaming/creator cards. GDDR memory without ECC, a GeForce driver, no NVLink, no MIG, no vGPU, no GPUDirect RDMA, and a driver licence that prohibits datacenter deployment.
RTX PRO / workstation & server (professional): RTX PRO 6000 Blackwell, RTX 6000 Ada, L40S. ECC memory, the RTX Enterprise / production-branch driver, vGPU, and (on the Blackwell PRO 6000 only) MIG. Still no NVLink on any current part; multi-GPU is PCIe peer-to-peer.

The single most important operational fact: none of these GPUs has NVLink, and the GeForce parts additionally lack MIG, vGPU, ECC, and GPUDirect RDMA, and may not lawfully run datacenter drivers. They are excellent single-GPU dev and inference boxes; they are not drop-in replacements for SXM parts in multi-GPU training or multi-tenant clusters. See GPU generations for the datacenter line (Hopper, Ampere, Blackwell platform).

flowchart TB
  Q["Choosing a GPU tier"] --> DC{"Multi-GPU NVLink, MIG, vGPU, or datacenter compliance needed?"}
  DC -->|"Yes"| SXM["Datacenter SXM/HGX: A100 / H100 / H200 / B200 / B300 — NVLink + NVSwitch, MIG, datacenter driver"]
  DC -->|"No, but want ECC / vGPU / MIG-on-one-GPU"| PRO{"Need MIG partitioning?"}
  PRO -->|"Yes"| P6000["RTX PRO 6000 Blackwell: 96 GB GDDR7 ECC, MIG up to 4 instances, vGPU, no NVLink"]
  PRO -->|"No, ECC + vGPU only"| ADA["RTX 6000 Ada / L40S: 48 GB GDDR6 ECC, vGPU, no MIG, no NVLink"]
  DC -->|"No, single-GPU dev or inference, cost-driven"| GEFORCE["GeForce RTX 5090 / 4090: 32 / 24 GB no ECC, GeForce driver, no NVLink/MIG/vGPU, not licensed for datacenter"]

Lineup & specifications¶

All figures cited below; verify on the linked NVIDIA pages. "—" means the feature is absent.

Spec	RTX 5090 (GeForce)	RTX 4090 (GeForce)	RTX PRO 6000 Blackwell	RTX 6000 Ada	L40S
Architecture	Blackwell GB202	Ada AD102	Blackwell	Ada Lovelace	Ada Lovelace
Memory	32 GB GDDR7, no ECC	24 GB GDDR6X, no ECC	96 GB GDDR7 ECC	48 GB GDDR6 ECC	48 GB GDDR6 ECC
Memory bus	512-bit	384-bit	512-bit	384-bit	384-bit
Bandwidth	1,792 GB/s	~1,008 GB/s	1,792 GB/s (WS); 1,597 GB/s (Server)	~960 GB/s	864 GB/s
CUDA cores	21,760	16,384	24,064 (per datasheet)	18,176	18,176
Tensor cores	5th gen (FP4)	4th gen (FP8)	5th gen (FP4)	4th gen (FP8)	4th gen (FP8)
NVLink	—	—	—	—	— (No)
MIG	—	—	Yes (up to 4 instances)	—	— (No)
ECC	—	—	Yes	Yes	Yes
vGPU	—	—	Yes	Yes	Yes
GPUDirect RDMA	— (GeForce)	— (GeForce)	Yes	Yes	Yes
Board power	575 W	450 W	up to 600 W	300 W	350 W
Connector	1x 12V-2x6 (600 W)	1x 12VHPWR (16-pin)	1x CEM5 16-pin	1x CEM5 16-pin	1x CEM5 16-pin
PCIe	Gen5	Gen4	Gen5	Gen4	Gen4
Driver	GeForce	GeForce	RTX Enterprise / production	RTX Enterprise / production	datacenter / RTX Enterprise
Cooling	open-air (consumer)	open-air (consumer)	blower/double-flow; Server passive	blower	passive (datacenter)

Notes: RTX 5090: 32 GB GDDR7, 512-bit, 1,792 GB/s, 575 W, 12V-2x6, PCIe Gen5, 21,760 CUDA, 5th-gen Tensor with FP4. RTX 4090: 24 GB GDDR6X, 384-bit, 450 W, 12VHPWR, PCIe Gen4, 16,384 CUDA; the RTX 4090 spec page lists NVLink as "No". RTX PRO 6000 Blackwell ships in Workstation, Server Edition (passive, FHFL air or single-slot liquid for the datacenter), and Max-Q (lower-power) editions; Workstation bandwidth is 1,792 GB/s, Server Edition is 1,597 GB/s. The RTX 5090's 21,760 CUDA cores sit across 170 SMs on the GB202 die. The RTX PRO 6000 CUDA count (24,064) is from the datasheet; verify on the current PDF.

Operational differences¶

This is the section that matters. Each line is a concrete divergence from a datacenter SXM part.

Drivers and licensing (the compliance differentiator)¶

GeForce RTX 50/40 run the GeForce Game Ready / Studio driver. The NVIDIA Driver License Agreement (the GeForce/consumer "License For Customer Use of NVIDIA Software", v. February 25, 2025), section 2.8, states: "You agree that GeForce or Titan SOFTWARE: (i) is licensed for use only on GeForce or Titan hardware products you own, and (ii) is not licensed for datacenter deployment." A long-standing carve-out permits blockchain processing in a datacenter, but not general datacenter compute. This is a real, contractual restriction, not a technical lockout: the silicon runs, but running the GeForce driver in a datacenter violates the licence. RTX PRO / datacenter parts use a different licence without this clause.
RTX PRO 6000 / RTX 6000 Ada use the RTX Enterprise (production-branch) driver, the professional analogue of the datacenter LTS branch, with longer support and ISV certification.
L40S is a datacenter card and uses the datacenter driver (and is licensed for datacenter use).
See GPU software stack for the driver/CUDA/Fabric-Manager versioning rules; the consumer parts simply omit Fabric Manager entirely.

No NVLink: multi-GPU is PCIe peer-to-peer only¶

RTX 5090, RTX 4090, RTX PRO 6000, RTX 6000 Ada, L40S: none has an NVLink connector. Ada Lovelace removed NVLink from GeForce; the RTX 3090 (Ampere) was the last GeForce card with NVLink. NVIDIA's rationale was that PCIe Gen5 peer-to-peer is "sufficiently fast" to replace it on consumer parts.
Consequence: multi-GPU on these cards is PCIe peer-to-peer (P2P) or staged through host memory. NCCL falls back to PCIe / SHM / host transports rather than NVLink, so collective bandwidth is bounded by the PCIe link (and by the topology, since separate root complexes cannot do GPU P2P). There is no NVSwitch domain, so Fabric Manager is not used and IMEX is irrelevant. Tensor parallelism that assumes NVLink bandwidth will be communication-bound. See tensor parallelism and networking fabric.

No GPUDirect RDMA on GeForce¶

GPUDirect RDMA is not supported on GeForce parts. The NIC cannot DMA directly into GeForce GPU memory, so the nvidia-peermem path used by InfiniBand/RoCE for zero-copy GPU-to-NIC transfer is unavailable; multi-node traffic stages through host memory. The professional parts (RTX PRO 6000, RTX 6000 Ada, L40S) do support GPUDirect RDMA. This is decisive for any multi-node job.

MIG only on RTX PRO 6000 Blackwell¶

Among the cards on this page, only the RTX PRO 6000 Blackwell supports MIG, partitioning the 96 GB GPU into up to four fully isolated instances (each with its own memory, cache, and compute). The published profile is four 1g.24gb slices; the datasheet/MIG user guide also list larger partitions (e.g. 2×48 GB, 1×96 GB); verify the exact profile set on the current datasheet.
GeForce (5090/4090) has no MIG. RTX 6000 Ada and L40S have no MIG. For hard multi-tenant isolation on these, the only options are whole-GPU allocation or MPS (no fault isolation). See security & multi-tenancy and the MIG/MPS notes in GPU software stack.

ECC, vGPU, Confidential Computing¶

ECC: professional only. RTX PRO 6000, RTX 6000 Ada, L40S have ECC GDDR; GeForce 5090/4090 do not. For long-running training or any correctness-sensitive workload, no-ECC consumer memory is a liability.
vGPU: professional only. RTX PRO 6000, RTX 6000 Ada, L40S support NVIDIA vGPU (vWS / vApps profiles); GeForce does not. No vGPU means no licensed VDI/MIG-backed virtualization on consumer cards.
Confidential Computing: these are not CC parts in the way Hopper/Blackwell datacenter GPUs are (TEE / TEE-I/O). Treat CC as a datacenter-SXM feature; verify per-SKU on the datasheet before claiming it.

Power connectors and cooling¶

Connectors: RTX 4090 uses 12VHPWR (the original 16-pin). RTX 5090 uses 12V-2x6, the PCI-SIG-revised 16-pin intended to fix 12VHPWR's contact issues. RTX PRO 6000 uses a single CEM5 16-pin. SXM datacenter parts use a busbar, not a cable; see Blackwell platform.
12VHPWR melting history: from the RTX 4090 launch (Oct 2022), melted connectors were reported within two weeks. NVIDIA acknowledged "about 50 cases globally" (~0.04–0.05% of early units); independent repair shops reported higher local counts. Root causes centred on partially seated connectors and weak early adapters. The 12V-2x6 revision improved seating margins, but melting has still been reported on the 575 W RTX 5090, especially with old 12VHPWR cables/adapters; full insertion and a quality native PSU cable are mandatory. See thermal/cooling emergency.
Cooling and form factor: GeForce cards are open-air triple-slot coolers that exhaust into the chassis; they are designed for a desktop with case airflow, not for front-to-back dense-server racking. The RTX PRO 6000 Server Edition is the only part here built for a server (passive dual-slot air, FHFL, or single-slot liquid, relying on chassis airflow). Packing open-air GeForce cards into a dense server starves them of airflow and causes thermal throttling or shutdown. See datacentre readiness.

Install & setup¶

Reference template; commands are unexecuted and version-pinned for illustration. Not hardware-tested. Verify against the current NVIDIA driver matrix (GPU software stack) before running.

Pick the driver family first: it is the operational fork¶

Three different driver products serve the cards on this page, and the choice is contractual as much as technical:

GeForce RTX 5090 / 4090 → the GeForce Game Ready / Studio driver (or the open kernel-module build for Turing-and-newer silicon). Datacenter-EULA caveat: the NVIDIA Driver License Agreement (the GeForce/consumer "License For Customer Use of NVIDIA Software") restricts GeForce/Titan software to GeForce/Titan hardware and prohibits datacenter deployment (a narrow blockchain-processing carve-out aside). The silicon runs; running the GeForce driver in a datacenter breaches the licence. Use a professional/datacenter part there instead.
RTX PRO 6000 Blackwell / RTX 6000 Ada → the RTX Enterprise (Production Branch) driver: a rebrand of the Quadro Optimal Driver for Enterprise, with ISV certification, long-lifecycle support, and regular security updates, and no datacenter restriction. Note the support-lag gotcha: brand-new Blackwell parts may land first in the New Feature Branch (NFB) before the Production Branch picks them up; check the NVIDIA Unix driver page for the branch that lists your exact SKU.
L40S → the datacenter (Tesla) driver, licensed for datacenter use. On Ubuntu the apt path is the CUDA network repo keyring plus the driver metapackage; on arm64 swap the x86_64 repo path for sbsa:

# Datacenter driver on Ubuntu (L40S; open kernel modules are the 560+ default flavour)
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb && sudo apt update
sudo apt install -y nvidia-open          # open kernel modules; use `cuda-drivers` for the proprietary flavour

Open kernel modules are supported on Turing and newer and have been the default/suggested flavour since the 560 driver series. For Blackwell (RTX 5090, RTX PRO 6000 Blackwell) you must use the open kernel modules; proprietary modules are unsupported on Blackwell; the RTX Enterprise .run installer takes -m=kernel-open, and the apt path is nvidia-open (not cuda-drivers). See GPU software stack for the per-tier driver and CUDA matrix.

Identify the card, set persistence, confirm features¶

# Identify the card and confirm driver branch (GeForce vs RTX Enterprise / datacenter)
nvidia-smi --query-gpu=name,driver_version,ecc.mode.current,vbios_version --format=csv

# Persistence: prefer the daemon (the legacy nvidia-smi flag is slated for deprecation
# in favour of the NVIDIA Persistence Daemon).
sudo systemctl enable --now nvidia-persistenced
# Fallback if the daemon package is absent (transient; lost on driver reload):
sudo nvidia-smi -pm 1
# Note: ECC, MIG, vGPU, NVLink, GPUDirect RDMA are NOT available on GeForce.

MIG: RTX PRO 6000 Blackwell only¶

Among the cards here, only the RTX PRO 6000 Blackwell partitions. Prerequisites that bite on a workstation board: Display Mode must be off (MIG needs compute-only firmware, so flip it with displaymodeselector and reboot), Resizable BAR (ReBAR) enabled in firmware, and a recent stack (driver 580.95+ with CUDA 13.0 per NVIDIA/partner guidance). Then:

# RTX PRO 6000 Blackwell — enable MIG, then carve instances
sudo nvidia-smi -mig 1                          # enable MIG (GPU idle; reboot/reset may be required)
sudo nvidia-smi mig -lgip                       # list GPU-instance profiles; 1g.24gb is profile ID 14
sudo nvidia-smi mig -cgi 14 -C                  # create 1g.24gb GI + compute instance (by profile ID)
sudo nvidia-smi -L                              # list MIG UUIDs to hand to CUDA_VISIBLE_DEVICES

Four 1g.24gb instances is the documented partitioning of the 96 GB board; verify the exact profile set against the MIG supported-profiles guide for your driver. GeForce, RTX 6000 Ada, and L40S have no MIG.

Confirm there is no NVLink and no Fabric Manager¶

nvidia-smi topo -m                              # links read PIX/PXB/PHB/SYS, never NV#
nvidia-smi nvlink --status                      # expect "inactive" / no NVLink on every part here

There is no Fabric Manager to install on any of these cards (no NVSwitch). For multi-GPU NCCL, expect PCIe/host transports; tune NCCL_P2P_LEVEL / NCCL_P2P_DISABLE / NCCL_SHM_DISABLE rather than NVLink envs. See GPU software stack and NCCL hang runbook.

When to use it¶

Consumer (RTX 5090 / 4090) makes sense when:

Single-GPU development boxes: local model dev, debugging, small fine-tunes, dataset work.
Single-GPU or low-fan-out inference: the RTX 5090's 32 GB GDDR7 at 1,792 GB/s serves many open models well at low cost per token.
Cost-driven workstations where datacenter compliance and multi-GPU scaling are not requirements.

Consumer does NOT make sense when:

Multi-GPU training at scale: no NVLink, no GPUDirect RDMA, no NVSwitch; collectives are PCIe/host-bound.
Multi-tenant serving: no MIG, no vGPU, no ECC, no hard isolation.
Datacenter deployment: the GeForce driver licence (§2.8) prohibits it.

Professional (RTX PRO 6000 / RTX 6000 Ada / L40S) makes sense when you need ECC, vGPU, datacenter-licensed drivers, GPUDirect RDMA, large per-GPU memory (96 GB on the PRO 6000), or MIG partitioning on a single GPU (PRO 6000 Blackwell only), but still do not need an NVLink-coupled multi-GPU domain. For NVLink/NVSwitch scale-up, move to SXM/HGX (Blackwell platform, DGX systems).

Networking¶

No NVLink anywhere on this page → PCIe peer-to-peer only. There is no high-bandwidth scale-up fabric between GPUs; the only inter-GPU path is PCIe P2P (same root complex) or staging through host memory. There is no NVSwitch, no rail-aligned NVLink domain, and no Fabric Manager. P2P works only when both GPUs sit under the same PCIe root complex with ACS not blocking it; GPUs on separate root complexes (SYS in the topology matrix) fall back to host staging. Plan multi-GPU topology around PCIe lanes and root-complex boundaries. See networking fabric and Kubernetes for GPUs for device exposure and topology-aware scheduling.
No GPUDirect RDMA on GeForce means multi-node GeForce traffic stages through host memory (the nvidia-peermem NIC-to-GPU zero-copy path is unavailable); throughput and latency suffer versus an RDMA-capable datacenter part. The professional parts (RTX PRO 6000, RTX 6000 Ada, L40S) do support GPUDirect RDMA over InfiniBand/RoCE.

Benchmark PCIe P2P bandwidth (the only inter-GPU link you have)¶

Two NVIDIA tools measure the GPU-to-GPU PCIe path directly:

# nvbandwidth (CUDA 11.x+; needs cmake >= 3.20, a C++17 compiler)
git clone https://github.com/NVIDIA/nvbandwidth && cd nvbandwidth
cmake . && make
./nvbandwidth -l                                  # list test cases
./nvbandwidth -t device_to_device_memcpy_read_ce  # GPU<->GPU copy-engine read bandwidth (P2P path)
./nvbandwidth -t device_to_device_memcpy_write_ce # and the write direction

# p2pBandwidthLatencyTest (from NVIDIA/cuda-samples) — the classic P2P matrix
# build: Samples/5_Domain_Specific/p2pBandwidthLatencyTest, then run the binary.
./p2pBandwidthLatencyTest                          # prints uni/bi-directional BW with P2P=Enabled vs Disabled

Read the P2P=Enabled figure against the PCIe-link ceiling (Gen5 x16 ~ 64 GB/s per direction; Gen4 x16 ~ 32 GB/s), not against an NVLink number. Cross-root-complex pairs will report P2P disabled and a much lower host-staged figure.

For the application-level proof and the multi-node path, do not duplicate the command set here; see the shared keystone Fabric bring-up, validation & benchmarking for nccl-tests over PCIe and over RoCE. Since these parts have no IB-class GPUDirect RDMA (and GeForce has no RDMA at all), the realistic multi-node fabric is RoCE on the professional parts; GeForce multi-node collectives are host-staged and PCIe/NIC-bound. NCCL falls back to PCIe / SHM transports automatically; benchmark before trusting a tensor-parallel layout (train tensor parallel).

Gotchas & failure modes¶

GeForce in a datacenter: a licence violation (NVIDIA Driver License Agreement §2.8), plus open-air cooling unsuited to dense racking and no ECC/MIG/RDMA. Avoid for production cluster use.
Expecting NVLink: nvidia-smi nvlink --status reports inactive/none; NCCL silently uses PCIe/host and tensor-parallel jobs become communication-bound. Confirm the topology with nvidia-smi topo -m (no NV# links).
12VHPWR / 12V-2x6 connector: under-seated connectors melt, especially at 450–575 W. Seat fully, prefer a native PSU cable over old adapters, and re-check after transport. See thermal/cooling emergency.
No ECC on GeForce: silent bit-flips are uncorrected; unsuitable for long training or correctness-critical inference.
MIG assumptions: only the RTX PRO 6000 Blackwell partitions; GeForce, RTX 6000 Ada, and L40S do not. Do not design a multi-tenant plan around MIG on those.
Open-air thermals in servers: GeForce coolers exhaust into the case and throttle when racked without front-to-back airflow; use the RTX PRO 6000 Server Edition for server chassis. See datacentre readiness.
Driver-branch mismatch: do not mix a GeForce driver and an RTX Enterprise driver expectation across a fleet; pin the correct branch per card class (GPU software stack).

References¶

NVIDIA GeForce RTX 5090 product page: https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/
NVIDIA GeForce RTX 4090 product page: https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/
NVIDIA RTX PRO 6000 Blackwell Workstation Edition: https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000/
NVIDIA RTX PRO 6000 Blackwell Workstation Edition datasheet (PDF): https://www.nvidia.com/content/dam/en-zz/Solutions/data-center/rtx-pro-6000-blackwell-workstation-edition/workstation-blackwell-rtx-pro-6000-workstation-edition-nvidia-us-3519208-web.pdf
NVIDIA RTX PRO 6000 Blackwell Server Edition: https://www.nvidia.com/en-us/data-center/rtx-pro-6000-blackwell-server-edition/
NVIDIA RTX 6000 Ada Generation: https://www.nvidia.com/en-us/design-visualization/rtx-6000/
NVIDIA L40S product page: https://www.nvidia.com/en-us/data-center/l40s/
NVIDIA GeForce driver licence (NVIDIA Driver License Agreement, §2.8 datacenter restriction): https://www.nvidia.com/en-us/drivers/geforce-license/
NVIDIA Blackwell architecture: https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/
NVIDIA Ada Lovelace architecture: https://www.nvidia.com/en-us/geforce/ada-lovelace-architecture/
NVIDIA Multi-Instance GPU (MIG) supported profiles: https://docs.nvidia.com/datacenter/tesla/mig-user-guide/supported-mig-profiles.html
NVIDIA MIG user guide — getting started (enable MIG, create instances): https://docs.nvidia.com/datacenter/tesla/mig-user-guide/getting-started-with-mig.html
NVIDIA RTX Enterprise / Production Branch driver downloads: https://www.nvidia.com/en-us/drivers/
NVIDIA datacenter (Tesla) driver installation guide — Ubuntu (cuda-keyring, nvidia-open vs cuda-drivers, sbsa repo): https://docs.nvidia.com/datacenter/tesla/driver-installation-guide/latest/ubuntu.html
NVIDIA driver installation guide — kernel modules (open modules Turing+, 560-series default): https://docs.nvidia.com/datacenter/tesla/driver-installation-guide/kernel-modules.html
NVIDIA Driver Persistence (legacy persistence mode deprecation; nvidia-persistenced daemon): https://docs.nvidia.com/deploy/driver-persistence/index.html
NVIDIA nvbandwidth — GPU bandwidth measurement tool (PCIe P2P device-to-device tests): https://github.com/NVIDIA/nvbandwidth
NVIDIA cuda-samples p2pBandwidthLatencyTest: https://github.com/NVIDIA/cuda-samples/blob/master/Samples/5_Domain_Specific/p2pBandwidthLatencyTest/README.md
Enabling MIG on Blackwell RTX PRO 5000/6000 (Display Mode off, ReBAR, 580.95+/CUDA 13), secondary: https://support.exxactcorp.com/hc/en-us/articles/35746564117271-NVIDIA-MIG-Support-How-To-Enable-on-NVIDIA-Blackwell-RTX-Pro-5000-6000-GPUs
Ada Lovelace microarchitecture (NVLink removal, AD102), secondary: https://en.wikipedia.org/wiki/Ada_Lovelace_(microarchitecture)