AMD Radeon Pro W6900X vs NVIDIA GeForce RTX 5090 Comparison

Head-to-Head Benchmarks

The recorded data shows a decisive NVIDIA GeForce RTX 5090 victory in the two directly comparable benchmark tests. In Geekbench OpenCL, the RTX 5090 scores 334,370 against the AMD Radeon Pro W6900X's 130,035, a difference of 61.1 percent in favor of NVIDIA. The Vulkan result follows the same pattern: the RTX 5090 reaches 376,728 while the W6900X manages 148,865, giving NVIDIA a 60.5 percent lead. These are not marginal advantages; they represent a fundamental performance gap in raw compute workloads.

The W6900X has no head-to-head wins in the database. Its strongest recorded results include a Geekbench Metal score of 226,821, an OpenCL score of 130,035, and a Vulkan score of 148,865. The RTX 5090's OpenCL score alone is more than 2.5 times the W6900X's entire average benchmark score. The average benchmark score for the W6900X sits at 168,574, while the RTX 5090's average is 79,842, though this comparison is skewed because the RTX 5090 includes several low-scoring Passmark legacy tests in its average. Looking at the actual compute-relevant results, the RTX 5090's OpenCL and Vulkan numbers dwarf the W6900X's corresponding scores.

Context from the nearest rivals helps interpret these figures. The W6900X sits in the 97th percentile of all GPUs, with its nearest competitor, the NVIDIA RTX 4500 Ada Generation, scoring 166,094, a mere 1.5 percent behind. The NVIDIA RTX A5500 trails by 2 percent at 165,217, and the AMD Radeon PRO W7800 comes in 2.2 percent lower at 164,894. The NVIDIA A100 PCIe 40 GB scores 162,504, 3.7 percent behind the W6900X. This places the W6900X in a solid but not exceptional tier among workstation-class accelerators.

The RTX 5090's percentile ranking is 92, lower than the W6900X's 97, yet its raw compute results are vastly superior. This discrepancy arises because the RTX 5090's average score includes Passmark tests like DirectX 9 at 395, DirectX 10 at 226, DirectX 11 at 341, DirectX 12 at 185, G2D at 1,413, G3D at 39,650, and GPU Compute at 26,756. These legacy benchmarks pull down its average despite the Geekbench scores being much higher. The nearest rivals for the RTX 5090 include the NVIDIA Tesla P100 PCIe 16 GB at 79,605 (0.3 percent behind), the Tesla P100 PCIe 12 GB at 79,396 (0.6 percent behind), and the AMD Radeon RX 6850M XT at 78,940 (1.1 percent behind). The AMD Radeon Pro Vega 64X actually scores higher at 80,959, showing a negative delta of 1.4 percent. These rival figures illustrate that the RTX 5090's average is dragged down by the Passmark suite, not that its compute performance is mediocre.

Architecture Differences

The two cards represent fundamentally different design philosophies. The AMD Radeon Pro W6900X uses the Navi 21 chip built on RDNA 2.0 architecture, fabricated on a 7 nm process at TSMC. The RTX 5090 employs the GB202 chip on Blackwell 2.0 architecture, using a 5 nm TSMC process. The transistor counts reveal the scale of this difference: the W6900X contains 26,800 million transistors on a 520 mm² die, giving a density of 51.5 million transistors per square millimeter. The RTX 5090 packs 92,200 million transistors on a 750 mm² die, achieving 122.9 million per square millimeter, more than double the density.

Memory configurations diverge sharply. The W6900X ships with 32 GB of GDDR6 on a 256-bit bus, delivering 512.0 GB/s of bandwidth at 16 Gbps effective. The RTX 5090 also has 32 GB, but it is GDDR7 on a 512-bit bus, producing 1.79 TB/s of bandwidth at 28 Gbps effective. The RTX 5090's memory bandwidth is roughly 3.5 times higher, a massive advantage for data-intensive workloads.

Compute resources differ by an even wider margin. The W6900X has 5,120 shading units, 320 texture mapping units, and 128 render output units, with 80 ray tracing cores and no tensor cores. The RTX 5090 features 21,760 shading units, 680 TMUs, and 176 ROPs, alongside 170 ray tracing cores and 680 tensor cores. The shading unit count is more than four times higher on the RTX 5090. Clock speeds also favor NVIDIA: the W6900X runs at 1,825 MHz base and 2,171 MHz boost, while the RTX 5090 operates at 2,017 MHz base and 2,407 MHz boost.

The resulting throughput numbers tell the story. The W6900X achieves 277.9 GPixel/s pixel rate, 694.7 GTexel/s texture rate, and 22.23 TFLOPS FP32 with 44.46 TFLOPS FP16 at a 2:1 ratio. The RTX 5090 reaches 423.6 GPixel/s, 1,636.8 GTexel/s, and 104.8 TFLOPS FP32 with 104.8 TFLOPS FP16 at a 1:1 ratio. The FP32 compute is nearly five times higher on the RTX 5090, and the FP16 rate matches FP32 rather than halving it.

Power requirements reflect these specifications. The W6900X has a TDP of 300 W with a suggested 700 W PSU. The RTX 5090 draws 575 W and recommends a 950 W PSU. Physical dimensions differ as well: the W6900X measures 267 mm long and 120 mm tall with an Apple MPX bus interface, while the RTX 5090 is 304 mm long, 137 mm tall, and 40 mm wide, using a dual-slot design with a 16-pin power connector and PCIe 5.0 x16 interface. Display outputs also vary: the W6900X offers 1x HDMI 2.1 and 4x Thunderbolt, while the RTX 5090 provides 1x HDMI 2.1b and 3x DisplayPort 2.1b. Both support DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4.

The Verdict

The data paints a clear picture. The RTX 5090 is the superior performer in raw compute across both OpenCL and Vulkan, with leads of 61.1 percent and 60.5 percent respectively. Its memory bandwidth of 1.79 TB/s, FP32 throughput of 104.8 TFLOPS, and 21,760 shading units place it in an entirely different performance class. The W6900X, despite its 97th percentile ranking and respectable average score of 168,574, cannot compete with the RTX 5090's Geekbench results.

The RTX 5090 is aimed at users who need maximum compute throughput, ray tracing capability with 170 RT cores, and tensor core acceleration for AI workloads. Its 680 tensor cores and 1:1 FP16 ratio make it suitable for machine learning inference and training tasks. The 32 GB GDDR7 memory on a 512-bit bus provides the bandwidth needed for large datasets and high-resolution textures.

The W6900X is an end-of-life product from 2021, designed for Apple MPX systems with Thunderbolt outputs. Its 32 GB GDDR6 memory and 512.0 GB/s bandwidth are adequate for professional visualization, but the 22.23 TFLOPS FP32 performance and 80 ray tracing cores place it below the RTX 5090 in every measurable compute category. The launch MSRP for the W6900X was 5,999 USD, while the RTX 5090 launched at 1,999 USD, though the database records no current pricing comparisons.

For the specific workload of Geekbench OpenCL and Vulkan compute, the RTX 5090 wins decisively. The W6900X retains its position among workstation cards like the RTX 4500 Ada Generation and RTX A5500, but the RTX 5090 is in a separate tier entirely.

FAQ

Q: Which card has better OpenCL performance?

A: The RTX 5090 scores 334,370 in Geekbench OpenCL compared to the W6900X's 130,035, a 61.1 percent advantage for NVIDIA.

Q: What is the memory bandwidth difference?

A: The W6900X delivers 512.0 GB/s from 32 GB of GDDR6 on a 256-bit bus. The RTX 5090 provides 1.79 TB/s from 32 GB of GDDR7 on a 512-bit bus, roughly 3.5 times higher.

Q: Do both cards support the same APIs?

A: Yes, both support DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4.

Q: How do the transistor counts compare?

A: The W6900X has 26,800 million transistors on a 520 mm² die, while the RTX 5090 has 92,200 million on a 750 mm² die, with densities of 51.5M per mm² and 122.9M per mm² respectively.

Q: What are the power requirements?

A: The W6900X has a 300 W TDP and suggests a 700 W PSU. The RTX 5090 has a 575 W TDP and recommends a 950 W PSU.

Q: Which card has tensor cores?

A: Only the RTX 5090 has tensor cores, with 680 of them. The W6900X has no tensor core count recorded.

Where Each One Wins

The RTX 5090 wins in every head-to-head benchmark recorded in the database. Its OpenCL score of 334,370 and Vulkan score of 376,728 are unmatched by the W6900X's 130,035 and 148,865. The RTX 5090 also holds advantages in FP32 compute at 104.8 TFLOPS versus 22.23 TFLOPS, texture rate at 1,636.8 GTexel/s versus 694.7 GTexel/s, and pixel rate at 423.6 GPixel/s versus 277.9 GPixel/s.

The W6900X's niche appears in its form factor and interface. It uses the Apple MPX bus with 4x Thunderbolt outputs, making it the only option for certain Apple Mac Pro configurations. Its 300 W TDP is lower than the RTX 5090's 575 W, and its 267 mm length fits in more compact chassis compared to the RTX 5090's 304 mm. The W6900X also supports a 2:1 FP16 ratio, which could benefit workloads that do not require full FP16 throughput.

For users constrained by power supply capacity, the W6900X's 700 W suggested PSU versus the RTX 5090's 950 W recommendation is a practical consideration. The W6900X's 32 GB GDDR6 memory matches the RTX 5090's 32 GB capacity, so memory size is not a differentiator, though bandwidth heavily favors NVIDIA.

Specification Differences

| Specification | AMD Radeon Pro W6900X | NVIDIA GeForce RTX 5090 |

|---|---|---|

| Architecture | RDNA 2.0 | Blackwell 2.0 |

| Process Node | 7 nm | 5 nm |

| Transistors | 26,800 million | 92,200 million |

| Die Size | 520 mm² | 750 mm² |

| Transistor Density | 51.5M / mm² | 122.9M / mm² |

| Base Clock | 1825 MHz | 2017 MHz |

| Boost Clock | 2171 MHz | 2407 MHz |

| Memory Type | GDDR6 | GDDR7 |

| Memory Bus Width | 256 bit | 512 bit |

| Memory Bandwidth | 512.0 GB/s | 1.79 TB/s |

| Shading Units | 5120 | 21760 |

| TMUs | 320 | 680 |

| ROPs | 128 | 176 |

| RT Cores | 80 | 170 |

| Tensor Cores | None recorded | 680 |

| Pixel Rate | 277.9 GPixel/s | 423.6 GPixel/s |

| Texture Rate | 694.7 GTexel/s | 1,636.8 GTexel/s |

| FP32 | 22.23 TFLOPS | 104.8 TFLOPS |

| FP16 | 44.46 TFLOPS (2:1) | 104.8 TFLOPS (1:1) |

| TDP | 300 W | 575 W |

| Suggested PSU | 700 W | 950 W |

| Bus Interface | Apple MPX | PCIe 5.0 x16 |

| Display Outputs | 1x HDMI 2.1, 4x Thunderbolt | 1x HDMI 2.1b, 3x DisplayPort 2.1b |

| Dimensions | 267 mm x 120 mm | 304 mm x 137 mm x 40 mm |

| Production Status | End-of-life | Active |

| Release Date | 2021-08-02 | 2025-01-29 |