NVIDIA RTX PRO 6000 Blackwell
Last reviewed
Sources
10 citations
Review status
Source-backed
Revision
v2 · 1,965 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
10 citations
Review status
Source-backed
Revision
v2 · 1,965 words
Add missing citations, update stale details, or suggest a clearer explanation.
The NVIDIA RTX PRO 6000 Blackwell is a professional workstation and server graphics processing unit (GPU) built on NVIDIA's Blackwell architecture, pairing the large GB202 die with 96 GB of GDDR7 memory and 24,064 CUDA cores. Announced at the GPU Technology Conference (GTC) on March 18, 2025, it is the flagship of the RTX PRO Blackwell family of professional graphics products and the successor to the RTX 6000 Ada Generation. NVIDIA describes the Workstation Edition as "the most powerful desktop GPU ever created" and positions the lineup for agentic AI, generative AI, large-model inference, rendering, scientific simulation, and professional design and visualization.[1][5]
The RTX PRO 6000 Blackwell delivers a generational leap in memory capacity and AI throughput over its predecessor, doubling memory from 48 GB to 96 GB and adding support for low-precision FP4 inference. It is offered in three editions targeting desktop workstations, thermally constrained workstations, and rack-mounted servers. A modified, lower-bandwidth derivative reported as the RTX PRO 6000D was subsequently created for the Chinese market to comply with United States export controls.
The RTX PRO 6000 Blackwell is NVIDIA's top professional graphics card of the Blackwell generation, sold for desktop workstations, dense multi-GPU workstations, and enterprise servers. Unlike NVIDIA's pure data-center accelerators, it is a unified AI-and-graphics part: it carries fourth-generation RT Cores and DisplayPort outputs for ray-traced rendering and visualization while also providing fifth-generation Tensor Cores and 96 GB of memory for on-device and on-premises AI inference. The naming dropped the long-running "Quadro" branding and the prior generation's "RTX A" and "RTX Ada" schemes in favor of the "RTX PRO" family name, with the "6000" denoting the flagship tier.[1][5]
NVIDIA unveiled the RTX PRO Blackwell lineup on March 18, 2025, during GTC 2025 in San Jose, California. The announcement spanned desktop, laptop, and server form factors, with the RTX PRO 6000 Blackwell sitting at the top of the desktop and server stacks. Lower desktop tiers announced alongside it included the RTX PRO 5000, RTX PRO 4500, RTX PRO 4000, and RTX PRO 2000 Blackwell, while a separate laptop series ranged from the RTX PRO 5000 down to the RTX PRO 500 Blackwell.[1][2]
NVIDIA framed the launch around enabling professionals to "build and collaborate with agentic AI," emphasizing on-device and on-premises AI workloads, neural rendering, ray tracing, physical AI, simulation, and content creation. Bob Pette, NVIDIA's vice president of enterprise platforms, said: "Bringing NVIDIA Blackwell to workstations and servers will take productivity, performance and speed to new heights, accelerating AI inference serving, data science, visualization and content creation."[2]
The company named a broad ecosystem of system and cloud partners for the new GPUs, including Dell Technologies, Hewlett Packard Enterprise, Lenovo, Supermicro, Cisco, ASUS, GIGABYTE, BOXX, HP Inc., and PNY, along with cloud providers Amazon Web Services, Google Cloud, Microsoft Azure, CoreWeave, and Lambda.[2]
The RTX PRO 6000 Blackwell is sold in three distinct editions that share the same GPU silicon and 96 GB of memory but differ in cooling, form factor, and power envelope.
| Edition | Target | Cooling | Form factor | Max power |
|---|---|---|---|---|
| RTX PRO 6000 Blackwell Workstation Edition | Desktop workstations | Active, double flow-through | 5.4 in H x 12.0 in L, dual slot | 600 W |
| RTX PRO 6000 Blackwell Max-Q Workstation Edition | Multi-GPU / thermally constrained workstations | Active | 4.4 in H x 10.5 in L, dual slot | 300 W |
| RTX PRO 6000 Blackwell Server Edition | Enterprise servers and data centers | Passive (requires chassis airflow) | 4.4 in H x 10.5 in L, dual slot | 400 W to 600 W (configurable) |
The Workstation Edition is the full-power desktop card. The Max-Q Workstation Edition trades peak performance for a 300 W envelope, allowing denser multi-GPU configurations in a single workstation. The Server Edition is a passively cooled, data-center card that relies on host-chassis airflow, supports up to eight GPUs per server, and is offered in both air-cooled and liquid-cooled variants.[1][3][4]
All three editions are built on the GB202 die and expose the same core configuration of 24,064 CUDA cores, 752 fifth-generation Tensor Cores, and 188 fourth-generation RT Cores, paired with 96 GB of GDDR7 memory with error-correcting code (ECC) on a 512-bit interface. The fifth-generation Tensor Cores add support for the FP4 data format, and the Server Edition supports Multi-Instance GPU (MIG) partitioning into up to four fully isolated 24 GB instances.[3][5]
| Specification | Workstation / Max-Q | Server Edition |
|---|---|---|
| Architecture | Blackwell (GB202) | Blackwell (GB202) |
| CUDA cores | 24,064 | 24,064 |
| Tensor Cores | 752 (5th gen) | 752 (5th gen) |
| RT Cores | 188 (4th gen) | 188 (4th gen) |
| Memory | 96 GB GDDR7 with ECC | 96 GB GDDR7 with ECC |
| Memory interface | 512-bit | 512-bit |
| Memory bandwidth | 1,792 GB/s | 1,597 GB/s |
| FP4 AI performance | 4,000 AI TOPS (with sparsity) | 4 PFLOPS |
| FP32 performance | 125 TFLOPS | 120 TFLOPS |
| RT Core performance | 380 TFLOPS | 355 TFLOPS |
| Media engines | 4x NVENC (9th gen), 4x NVDEC (6th gen) | 4x NVENC (9th gen), 4x NVDEC (6th gen) |
| Interface | PCIe Gen 5 x16 | PCIe Gen 5 x16 |
| Display outputs | 4x DisplayPort 2.1 | 4x DisplayPort 2.1 |
| MIG | Not applicable | Up to 4 instances of 24 GB |
| Max power | 600 W (WS) / 300 W (Max-Q) | 400 W to 600 W (configurable) |
The Server Edition runs its memory at a slightly lower effective bandwidth (1,597 GB/s versus 1,792 GB/s on the Workstation Edition) to suit data-center thermal and power constraints. The ninth-generation NVENC engines add 4:2:2 H.264 and HEVC encoding support, and the sixth-generation NVDEC engines provide up to double the H.264 decoding throughput of the prior generation. The Workstation Edition launched at an approximate price of 8,565 US dollars.[5][6]
The RTX PRO 6000 Blackwell succeeds the RTX 6000 Ada Generation, which launched in late 2022 on the Ada Lovelace architecture using the AD102 die. The naming convention dropped the "Quadro" branding entirely and introduced the "RTX PRO" family name. The generational gains are substantial across every axis.
| Specification | RTX 6000 Ada Generation | RTX PRO 6000 Blackwell (Workstation) |
|---|---|---|
| Architecture | Ada Lovelace (AD102) | Blackwell (GB202) |
| CUDA cores | 18,176 | 24,064 |
| Tensor Cores | 568 (4th gen) | 752 (5th gen) |
| RT Cores | 142 (3rd gen) | 188 (4th gen) |
| Memory | 48 GB GDDR6 with ECC | 96 GB GDDR7 with ECC |
| Memory interface | 384-bit | 512-bit |
| Memory bandwidth | 960 GB/s | 1,792 GB/s |
| FP32 performance | 91 TFLOPS | 125 TFLOPS |
| Max power | 300 W | 600 W |
| Interface | PCIe Gen 4 x16 | PCIe Gen 5 x16 |
The doubling of memory capacity to 96 GB and the addition of FP4 precision position the Blackwell card for large-language-model workloads that the Ada-generation part could not host on a single GPU. The trade-off is a doubling of the maximum board power to 600 W on the Workstation Edition, alongside an upgrade from a 384-bit GDDR6 bus to a 512-bit GDDR7 bus that nearly doubles memory bandwidth from 960 GB/s to 1,792 GB/s.[5]
NVIDIA positions the RTX PRO 6000 Blackwell as a unified platform for AI and graphics workloads that previously required separate hardware. The 96 GB of memory allows a single card to hold large generative models, supporting on-premises AI inference and fine-tuning without offloading to external accelerators. The fifth-generation Tensor Cores and FP4 support target high-throughput inference for large language models and generative AI, an area otherwise served by data-center accelerators such as the L40S.
Beyond AI, the fourth-generation RT Cores and high memory capacity accelerate ray-traced and neural rendering for media and entertainment, real-time 3D design, architecture, engineering and construction (AEC) workflows, manufacturing prototyping, scientific and engineering simulation, and physical AI development. In the Server Edition, MIG partitioning and multi-GPU scaling extend these capabilities to shared enterprise and cloud environments, where NVIDIA promoted RTX PRO Servers as a path to running agentic AI alongside visualization on mainstream enterprise systems.[2][3]
United States export controls progressively restricted the sale of NVIDIA's most capable accelerators to China, beginning with the A100 and H100 class and continuing through China-specific parts such as the H20. To remain in the Chinese market under these rules, NVIDIA developed a modified version of the RTX PRO 6000 reported as the RTX PRO 6000D, associated in reporting with the internal codename "B40."
According to reporting in mid-2025, the China variant retains the Blackwell architecture and GDDR7 memory but constrains memory bandwidth to roughly 1,398 GB/s, kept deliberately below the United States threshold of about 1.4 TB/s that governs which chips may be exported. Early reports, citing supply-chain outlet DigiTimes, described a card delivering around 1,100 GB/s of bidirectional bandwidth, fabricated on TSMC's N4 process, with shipments targeted for the third quarter of 2025. The RTX PRO 6000D notably lacks NVLink, relying on PCIe or external network interface cards such as ConnectX for multi-GPU communication, which sharply limits its usefulness for scaling large-model inference across many GPUs. Its reported retail price in China was around 50,000 yuan (roughly 7,000 US dollars).[7][8]
The variant met a difficult reception. In September 2025, reporting by the Financial Times and others, including Bloomberg, indicated that the Cyberspace Administration of China directed major technology companies such as Alibaba and ByteDance to halt testing and cancel orders for the RTX PRO 6000D. The directive was described as stronger than earlier guidance that had targeted the H20, and it pushed Chinese buyers toward domestic alternatives from suppliers such as Huawei, Cambricon, and Biren. Industry coverage noted that the chip's value proposition was weak relative to grey-market consumer cards: banned GeForce RTX 5090 boards reportedly circulated for around 3,500 US dollars while delivering competitive inference performance, undercutting the more expensive, bandwidth-limited 6000D.[9][10]
The RTX PRO 6000 Blackwell marks the point at which NVIDIA's professional desktop line absorbed data-center-class memory capacity, with 96 GB on a single workstation card enabling local development and serving of large AI models. By spanning workstation, Max-Q, and server editions from one silicon design, it lets organizations standardize on a single GPU across desktops, dense multi-GPU workstations, and enterprise servers. At the same time, the RTX PRO 6000D episode illustrates how export controls reshape product strategy: a deliberately bandwidth-limited derivative, stripped of NVLink to satisfy regulatory thresholds, found little demand in a market that was simultaneously being steered by its own government toward domestic accelerators.