NVIDIA RTX PRO 6000 Blackwell Server

The NVIDIA RTX PRO 6000 Blackwell Server is a dual-slot accelerator built on the 5 nm GB202 chip, featuring 92,200 million transistors across a 750 mm² die. It is positioned within the Server Blackwell generation, succeeding the Server Hopper class and preceding the anticipated Server Rubin. The card is currently in active production, with a release date of March 17, 2025. This analysis draws strictly from the provided data, which does not include synthetic benchmark scores or rival comparisons, so the evaluation focuses on architectural specifications and derived performance indicators.

Benchmark Performance

The FACT PACK lists an average benchmark score of 0 and a percentile rank of 50 among all GPUs, indicating that no direct performance measurements are available for this specific database entry. Consequently, numerical comparisons against rivals cannot be made, as the nearestRivals field is empty. However, the raw compute specifications offer a clear picture of the card’s theoretical throughput. The FP32 performance is rated at 126.0 TFLOPS, which serves as the primary metric for general compute workloads. The FP16 performance is identical at 126.0 TFLOPS, leveraging a 1:1 ratio, meaning there is no dedicated half-precision boost; this is notable for workloads that typically accelerate FP16 operations, as the card does not differentiate between the two precisions.

The pixel rate is 502.5 GPixel/s, derived from 192 ROPs and a boost clock of 2617 MHz, while the texture rate stands at 1,968.0 GTexel/s, driven by 752 TMUs. These figures are substantial for a server part, suggesting strong rasterization throughput for visualization tasks. The shading units total 24,064, which is the core count for parallel processing. The presence of 188 RT cores and 752 tensor cores indicates dedicated hardware for ray tracing and AI inference, respectively, though the FACT PACK does not provide specific TFLOPS ratings for these units. Given the absence of benchmark scores, the data indicates that the RTX PRO 6000 Blackwell Server is architecturally capable of high-end compute, but its real-world performance relative to other GPUs remains unquantified in this entry. The clock behavior shows a base frequency of 1590 MHz and a boost of 2617 MHz, a 64.6% uplift, which is a significant dynamic range for sustained workloads.

Memory Subsystem

The memory configuration is a defining feature of this card. It comes with 96 GB of GDDR7 memory, a capacity that is substantial for server-class workloads requiring large datasets, such as AI model training or in-memory databases. The bus width is 512 bit, which is the maximum typically found in NVIDIA’s flagship parts, allowing for a wide data path. The memory clock is rated at 1750 MHz, with an effective data rate of 28 Gbps. This combination yields a total bandwidth of 1.79 TB/s. This bandwidth figure is critical for high-resolution rendering and large-scale compute tasks, as it dictates how quickly data can be fed to the 24,064 shading units. For high-resolution scenarios, such as 8K visualization or multi-display configurations, the 96 GB capacity mitigates the need for constant data swapping, while the 1.79 TB/s throughput ensures that texture fetches and buffer updates do not become bottlenecks. The GDDR7 type is a modern standard, offering higher density and efficiency compared to older GDDR6 variants, though the FACT PACK does not provide comparative bandwidth numbers for other memory types. The 512-bit bus width, when paired with 96 GB, suggests a balanced design aimed at capacity and bandwidth rather than latency optimization.

Who Should Consider It

Based on the available specifications, the RTX PRO 6000 Blackwell Server is intended for environments where memory capacity and compute throughput are paramount. The 96 GB VRAM and 126.0 TFLOPS FP32 performance position it for 4K and 8K rendering workloads, where texture-heavy scenes and large frame buffers exceed the capacity of consumer cards. For AI inference and training, the 752 tensor cores and 1.79 TB/s bandwidth support large model sizes, though the lack of FP16 acceleration means users should not expect a speedup when switching from FP32 to FP16 workloads. The dual-slot design and 600 W TDP suggest it is meant for rack-mounted server systems with adequate airflow, rather than compact workstations. The PCIe 5.0 x16 interface ensures compatibility with the latest server platforms, providing high host-to-device transfer rates, although the FACT PACK does not specify the actual throughput of this interface. Users with workloads that fit within 96 GB and require sustained FP32 performance are the primary audience; those with smaller memory needs may find the capacity excessive. The card’s 50th percentile rank, while lacking benchmark context, implies it sits at the median of all GPUs in the database, but this is a placeholder value given the zero benchmark score, so it should not be interpreted as a performance verdict.

How It Compares

The nearestRivals field is empty, and the benchmarks array contains no entries. Therefore, direct comparisons to specific rival products—such as AMD Instinct or Intel Data Center GPU offerings—cannot be made using the data. The predecessor is listed as “Server Hopper,” which indicates an architectural lineage, but no performance deltas are provided. The successor is “Server Rubin,” but again, no comparative metrics exist. Without rival scores or deltaPct values, any assessment of its position relative to competing hardware is purely speculative. The data only allows for a qualitative statement: the RTX PRO 6000 Blackwell Server is a high-specification accelerator, but its competitive standing is undefined in this database. The percentile rank of 50 is a statistical artifact of the zero score, not a measure of relative strength. Consequently, users should rely on the raw specifications—126.0 TFLOPS, 96 GB, 1.79 TB/s—to gauge its capability, but cannot benchmark it against named alternatives from the provided facts.

FAQ

Q: What is the memory capacity and type?

A: The card features 96 GB of GDDR7 memory.

Q: What is the peak FP32 performance?

A: The FP32 throughput is rated at 126.0 TFLOPS.

Q: Does the card support FP16 acceleration?

A: Yes, FP16 is supported at 126.0 TFLOPS, but it is a 1:1 ratio with FP32, meaning there is no separate boost for half-precision workloads.

Q: What is the total memory bandwidth?

A: The bandwidth is 1.79 TB/s, derived from a 512-bit bus and 28 Gbps effective memory clock.

Q: How many RT and tensor cores are present?

A: The card includes 188 RT cores and 752 tensor cores.

Q: What is the power consumption and PSU requirement?

A: The TDP is 600 W, and the suggested PSU is 1000 W.

Power and Cooling

The RTX PRO 6000 Blackwell Server has a thermal design power (TDP) of 600 W, which is a high figure that dictates the cooling solution. The card is dual-slot in width, measuring 267 mm in length, 111 mm in height, and 40 mm in width. This form factor is standard for server GPUs, allowing for passive or active cooling within a chassis. The power is delivered via a single 16-pin connector, which is the modern standard for high-wattage GPUs. The suggested PSU rating is 1000 W, which accounts for the card’s draw plus system overhead, though the FACT PACK does not specify the efficiency or headroom requirements. The boost clock of 2617 MHz under load will likely cause power draw to approach the 600 W limit, so robust cooling is essential. The 5 nm process node from TSMC helps mitigate heat generation relative to older nodes, but 600 W still requires a server-grade cooling solution, such as high-static-pressure fans or liquid cooling loops, though the FACT PACK does not provide specific cooling recommendations. The dual-slot design allows for a larger heatsink, but the 40 mm width is standard, not oversized. Users should ensure their server chassis can exhaust the heat from this card, as sustained workloads at 126.0 TFLOPS will generate significant thermal output. The single 16-pin connector simplifies cabling but requires a PSU that supports the 600 W draw on that single rail.