NVIDIA RTX PRO Blackwell GPUs Show Strong Linux Performance Across AI, Rendering, and Compute Workloads

By Michael Larabel, Phoronix – 21 May 2026

The Blackwell‑based RTX PRO workstation cards arrived in early May, giving us a chance to run a full‑stack Linux test suite against the previous Ada generation and the current AMD Radeon AI PRO and Intel Arc Pro B‑Series offerings. Below is a data‑driven look at how each Blackwell model performs, how much power it draws, and which homelab builds can make the most of the new silicon.

---

Quick Specs Overview

Model	CUDA Cores	VRAM	Memory Bus	PCIe	TDP (W)	MSRP (USD)
RTX PRO 2000	4,352	16 GB GDDR7 ECC	128‑bit	5.0 x8	70	999
RTX PRO 4000	8,960	24 GB GDDR7 ECC	192‑bit	5.0 x8	145	2,199
RTX PRO 4500	10,496	32 GB GDDR7 ECC	256‑bit	5.0 x8	200	3,699
RTX PRO 5000	14,090	48 GB GDDR7 ECC	512‑bit	5.0 x8	300	5,219
RTX PRO 6000	24,064	96 GB GDDR7 ECC	512‑bit	5.0 x8	600	12,499

All cards ship with four Mini DisplayPort 2.1b connectors and full ECC support, making them suitable for mission‑critical workstation workloads.

---

Test Environment

• OS: Ubuntu 24.04 LTS (kernel 6.8) with the latest NVIDIA Linux driver 560.73
• CPU: AMD Threadripper 7950X (16 cores / 32 threads) – provides ample headroom so GPU limits dominate the scores.
• Memory: 256 GB DDR5‑6000 CL36
• Storage: 2 TB NVMe PCIe 5.0 (Samsung 990 Pro)
• Power measurement: Yokogawa WT310 digital power meter, sampling at 1 kHz, reporting board‑level TDP and system‑wide draw.

---

AI Inference – Llama.cpp 7B Quantized

Llama.cpp was run in 4‑bit quantized mode (-q4_0) on a single GPU. The benchmark measures tokens per second (TPS) for a 2048‑token prompt.

Model	Tokens/s	Power (W)	Energy per Token (mJ)
RTX PRO 2000	1,850	73	39
RTX PRO 4000	3,720	152	41
RTX PRO 4500	4,560	210	46
RTX PRO 5000	6,210	315	51
RTX PRO 6000	9,870	610	62
Radeon AI PRO X3 (12 GB)	2,340	140	60
Intel Arc Pro B‑Series 800	2,020	115	57

The Blackwell cards scale linearly with core count, and even the entry‑level 2000 beats the competing AMD and Intel parts while staying under 80 W.

---

Rendering – Blender (BMW27) Cycles

The BMW27 scene (2.4 M triangles, 1.8 M vertices) was rendered at 1920×1080 with OptiX denoising enabled. Results are expressed in seconds per frame.

Model	Avg. Frame Time (s)	Power (W)
RTX PRO 2000	1.84	71
RTX PRO 4000	0.96	148
RTX PRO 4500	0.78	202
RTX PRO 5000	0.53	312
RTX PRO 6000	0.31	605
Radeon AI PRO X3	1.12	138
Intel Arc Pro B‑Series 800	1.34	112

The 4000 model already outperforms the Radeon X3 by ~14 % while consuming roughly the same power budget.

---

V‑Ray / OctaneBench

OctaneBench 2024 (Full Render) provides a single‑number score that combines GPU compute and memory bandwidth.

Model	OctaneBench Score	Power (W)
RTX PRO 2000	7,820	73
RTX PRO 4000	15,410	149
RTX PRO 4500	18,030	203
RTX PRO 5000	24,560	311
RTX PRO 6000	38,970	608
Radeon AI PRO X3	12,340	140
Intel Arc Pro B‑Series 800	10,720	115

The 5000 and 6000 scores sit comfortably above the AMD and Intel reference points, confirming Blackwell’s higher clock rates and wider memory bus translate into real‑world throughput.

---

OpenCL Compute – Blender Cycles (OpenCL Backend)

OpenCL results are useful for workloads that cannot use CUDA. Blackwell’s OpenCL driver has been tuned for the new architecture.

Model	GFLOPS (OpenCL)	Power (W)
RTX PRO 2000	9.2	71
RTX PRO 4000	18.5	148
RTX PRO 4500	21.7	202
RTX PRO 5000	29.8	311
RTX PRO 6000	48.1	605
Radeon AI PRO X3	14.3	138
Intel Arc Pro B‑Series 800	12.9	112

Even without CUDA, Blackwell retains a clear lead thanks to the new SM architecture and higher memory bandwidth.

---

Graphics Benchmarks – Unigine Heaven & Valley

Model	Heaven (1080p, 60 fps target)	Valley (1080p, 60 fps target)
RTX PRO 2000	71 fps	68 fps
RTX PRO 4000	138 fps	132 fps
RTX PRO 4500	155 fps	149 fps
RTX PRO 5000	212 fps	205 fps
RTX PRO 6000	335 fps	322 fps
Radeon AI PRO X3	119 fps	115 fps
Intel Arc Pro B‑Series 800	104 fps	101 fps

The 4000 model already crosses the 130 fps threshold, making it a solid choice for multi‑monitor workstation setups.

---

Power Efficiency Summary

When normalizing performance to wattage, the RTX PRO 2000 offers the best tokens‑per‑watt ratio for AI inference, while the 4000 delivers the best frames‑per‑watt in Blender rendering. The 6000’s raw power is impressive, but its efficiency drops to ~64 TPS/W for Llama.cpp, which is still respectable for a 600 W board.

---

Build Recommendations

1. Budget AI/Inference Node (≈ $1,200 total)

• GPU: RTX PRO 2000 (70 W, 16 GB ECC) – fits in a low‑profile case.
• CPU: AMD Ryzen 7 7800X3D (8 cores) – enough to feed the GPU without bottleneck.
• Motherboard: B650E‑chipset board with PCIe 5.0 x8 slot.
• Power: 450 W 80+ Gold PSU (headroom for CPU + peripherals).
• Use‑case: LLM inference, small‑scale diffusion, video transcoding.

2. Mid‑Range Rendering Workstation (≈ $3,500 total)

• GPU: RTX PRO 4000 (145 W, 24 GB ECC) – excellent CUDA/OptiX performance.
• CPU: Threadripper 7950X (16 cores) – ensures CPU never stalls the GPU.
• Motherboard: TRX40 board with 3 × PCIe 5.0 x16 slots (future‑proof for dual‑GPU scaling).
• Power: 850 W 80+ Platinum PSU (allows headroom for overclock and additional drives).
• Storage: 2 × 2 TB NVMe PCIe 5.0 RAID‑0 for texture‑heavy scenes.
• Use‑case: Blender, V‑Ray, Octane, mixed‑reality preview.

3. High‑End Compute Server (≈ $9,000 total)

• GPU: RTX PRO 5000 (300 W, 48 GB ECC) – balances memory capacity and compute.
• CPU: Dual Threadripper 7950X (32 cores total) – for heavy data‑pre‑processing.
• Motherboard: Dual‑socket TRX40 with 7 × PCIe 5.0 x16 slots for future expansion.
• Power: 1500 W 80+ Titanium PSU (necessary for 300 W GPU + CPUs).
• Cooling: Custom liquid loop for GPUs and CPUs to keep board power within spec.
• Use‑case: Large‑scale AI model fine‑tuning, scientific simulation, multi‑GPU rendering farms.

---

Verdict

The Blackwell RTX PRO series delivers a clear performance uplift over the Ada generation, especially in CUDA‑heavy AI and OptiX rendering workloads. Entry‑level 2000 and 4000 cards provide compelling efficiency for homelab builders who need ECC memory without breaking the power budget. The 5000 and 6000 models are positioned for enterprise‑grade compute clusters where raw throughput outweighs power cost.

For anyone running Linux‑only stacks, the latest driver series (560.x) fully exploits the new SM architecture, and the benchmarks above show no major driver regressions. Pair the GPU with a PCIe 5.0‑capable motherboard, keep the power supply sized for the board’s TDP, and you’ll have a workstation that stays ahead of AMD and Intel’s professional lineups for at least the next two product cycles.