NVIDIA GeForce RTX 5090 D vs NVIDIA H200 NVL Comparison

Where Each One Wins

The recorded data splits these two NVIDIA parts into completely different usage arenas. The NVIDIA H200 NVL wins the only directly comparable compute benchmark, Geekbench OpenCL, by 7.8% over the GeForce RTX 5090 D. That is the single head-to-head measurement available, and it favors the server accelerator. The H200 NVL sits at the 100th percentile among all GPUs in the database, meaning no other recorded GPU scores higher in aggregate. Its nearest rivals are all data center parts: the B200 is 3.1% faster, the B300 SXM6 AC is 9.4% faster, while the Instinct MI300X trails by 5.3% and the L40S by 13.2%.

The GeForce RTX 5090 D, by contrast, occupies the 92nd percentile and its average benchmark score is dragged down by a wide spread of tests that include DirectX 9, 10, 11, and 12 workloads, Vulkan, and Passmark compute. Its strongest recorded result is Geekbench Vulkan at 376,915, which is higher than its OpenCL score of 310,674. The H200 NVL has no Vulkan or DirectX entries in the database, so the client card wins every graphics API test by default of having any recorded result at all. The 5090 D also posts a Passmark G3D score of 44,065 and a GPU compute score of 28,396, while the H200 NVL has no Passmark entries. If the workload is graphics rendering, gaming, or any DirectX/Vulkan path, the 5090 D is the only one of the two with data; if the workload is raw OpenCL compute, the H200 NVL leads.

Architecture Differences

The H200 NVL is built on the Hopper architecture, specifically the GH100 chip, while the RTX 5090 D uses Blackwell 2.0 on the GB202 die. Both are fabricated on a 5 nm process at TSMC, but the transistor counts diverge: the H200 NVL packs 80,000 million transistors on an 814 mm² die, while the 5090 D carries 92,200 million transistors on a smaller 750 mm² die. That gives the 5090 D a higher transistor density of 122.9M per mm² versus 98.3M per mm² for the H200 NVL.

Memory is the largest architectural gap. The H200 NVL uses 141 GB of HBM3e on a 6144-bit bus, delivering 4.89 TB/s of bandwidth. The 5090 D uses 32 GB of GDDR7 on a 512-bit bus, delivering 1.79 TB/s. That is nearly 2.7 times more bandwidth for the H200 NVL, and more than four times the memory capacity. The H200 NVL also has a much lower pixel rate at 42.84 GPixel/s versus 423.6 GPixel/s for the 5090 D, which reflects its role as a compute accelerator with only 24 ROPs compared to 176 on the 5090 D.

Shader and tensor configurations differ sharply. The 5090 D has 21,760 shading units, 680 TMUs, 170 RT cores, and 680 tensor cores. The H200 NVL has 16,896 shading units, 528 TMUs, and 528 tensor cores, with no RT core count listed. FP32 throughput favors the 5090 D at 104.8 TFLOPS versus 60.32 TFLOPS for the H200 NVL. FP16 tells a different story: the H200 NVL reaches 120.6 TFLOPS with a 2:1 ratio, while the 5090 D delivers 104.8 TFLOPS at 1:1. Clock speeds also favor the client card, with a 2017 MHz base and 2407 MHz boost versus 1365 MHz base and 1785 MHz boost on the H200 NVL.

Power and physical design are close but not identical. The H200 NVL draws 600 W with an 8-pin EPS connector and a suggested 1000 W PSU; the 5090 D draws 575 W with a single 16-pin connector and a suggested 950 W PSU. Both are dual-slot, but the 5090 D is longer at 304 mm versus 267 mm, taller at 137 mm versus 111 mm, and has a listed 48 mm width while the H200 NVL has no width recorded. The H200 NVL has no display outputs and no DirectX, OpenGL, or Vulkan API support listed; the 5090 D offers 1x HDMI 2.1b and 3x DisplayPort 2.1b, with DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4.

Head-to-Head Benchmarks

The only direct comparison in the database is Geekbench OpenCL. The H200 NVL scores 334,891 against 310,674 for the 5090 D, a 7.8% advantage. That margin is meaningful but not overwhelming. It places the H200 NVL ahead of the 5090 D in raw OpenCL compute, but the 5090 D still posts a respectable score that is within striking distance. For context, the H200 NVL's nearest rival, the B200, beats it by 3.1%, while the Instinct MI300X trails by 5.3%. The 5090 D's OpenCL result would sit below all of those server parts, but the database does not list the 5090 D against those rivals directly.

The 5090 D has a richer benchmark portfolio. Its Geekbench Vulkan score of 376,915 exceeds its own OpenCL score by 21.3%, which suggests that the client card handles Vulkan compute or graphics workloads more efficiently than OpenCL. Its Passmark G3D score of 44,065 is the highest single graphics-oriented result in its file, while its Passmark GPU compute score of 28,396 is lower than its OpenCL figure. The DirectX legacy tests are all low three-digit scores: 434 in DirectX 9, 371 in DirectX 11, 231 in DirectX 10, and 219 in DirectX 12. Those numbers are not comparable to modern GPUs in a meaningful way, but they are the only DirectX data points in the entire comparison.

The H200 NVL has no other benchmark entries beyond OpenCL, so its average benchmark score equals that single result at 334,891. The 5090 D's average of 77,712 is pulled down by the low Passmark DirectX scores and the G2D result of 1,487. That average is misleading for a GPU with a Vulkan score above 376,000, so the database percentile of 92 for the 5090 D versus 100 for the H200 NVL reflects the breadth of testing rather than pure compute capability. The H200 NVL wins the only shared test, but the 5090 D demonstrates strength in Vulkan and graphics-specific workloads that the H200 NVL cannot contest due to lack of recorded data.

The Verdict

Pick the H200 NVL if the task is large-scale compute, AI inference, or any workload that needs massive memory capacity and bandwidth. The 141 GB HBM3e pool and 4.89 TB/s bandwidth dwarf the 5090 D's 32 GB and 1.79 TB/s, and the 7.8% OpenCL lead confirms its compute edge. It also carries the 100th percentile ranking, meaning no other GPU in the database has a higher aggregate score. The absence of display outputs and graphics API support makes it unsuitable for any interactive or rendering role, but for server-side compute it is the clear choice.

Pick the RTX 5090 D if the workload touches graphics, gaming, DirectX, Vulkan, or any client-side rendering pipeline. It has the only recorded Vulkan score at 376,915, the only DirectX scores, the only Passmark G3D result, and it supports HDMI 2.1b and DisplayPort 2.1b outputs. Its 104.8 TFLOPS FP32 and 423.6 GPixel/s pixel rate are far ahead of the H200 NVL in those metrics, and its 170 RT cores provide hardware ray tracing that the H200 NVL does not list. The 5090 D also runs at higher clocks and has a smaller transistor density advantage on a smaller die.

The H200 NVL leads in FP16 throughput at 120.6 TFLOPS versus 104.8 TFLOPS, which matters for AI training and inference workloads that use mixed precision. The 5090 D counters with a 1:1 FP16 ratio, meaning its FP16 and FP32 are equal, while the H200 NVL's 2:1 ratio indicates its FP16 is double its FP32. Neither card is a universal winner; the database clearly separates them by role. The H200 NVL is a server compute accelerator, the 5090 D is a client graphics and compute card.

FAQ

Q: Which card has higher raw FP32 performance?

A: The GeForce RTX 5090 D, at 104.8 TFLOPS versus 60.32 TFLOPS for the H200 NVL.

Q: How much memory bandwidth does each card have?

A: The H200 NVL has 4.89 TB/s from 141 GB of HBM3e on a 6144-bit bus. The RTX 5090 D has 1.79 TB/s from 32 GB of GDDR7 on a 512-bit bus.

Q: Which card wins in Geekbench OpenCL?

A: The H200 NVL scores 334,891 versus 310,674 for the RTX 5090 D, a 7.8% margin.

Q: Does the RTX 5090 D support ray tracing?

A: Yes, it lists 170 RT cores. The H200 NVL has no RT core count in the database.

Q: What display outputs does the H200 NVL have?

A: None. It has no display outputs, while the RTX 5090 D has 1x HDMI 2.1b and 3x DisplayPort 2.1b.

Q: What is the power draw and PSU requirement for each?

A: The H200 NVL draws 600 W with a suggested 1000 W PSU and an 8-pin EPS connector. The RTX 5090 D draws 575 W with a suggested 950 W PSU and a single 16-pin connector.

Specification Differences

| Field | NVIDIA H200 NVL | NVIDIA GeForce RTX 5090 D |

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

| Chip | GH100 | GB202 |

| Architecture | Hopper | Blackwell 2.0 |

| Generation | Server Hopper (Hxx) | GeForce 50 |

| Transistors | 80,000 million | 92,200 million |

| Die Size | 814 mm² | 750 mm² |

| Transistor Density | 98.3M / mm² | 122.9M / mm² |

| Base Clock | 1365 MHz | 2017 MHz |

| Boost Clock | 1785 MHz | 2407 MHz |

| Memory Clock | 1593 MHz 6.4 Gbps effective | 1750 MHz 28 Gbps effective |

| Memory Size | 141 GB | 32 GB |

| Memory Type | HBM3e | GDDR7 |

| Memory Bus Width | 6144 bit | 512 bit |

| Memory Bandwidth | 4.89 TB/s | 1.79 TB/s |

| Shading Units | 16896 | 21760 |

| TMUs | 528 | 680 |

| ROPs | 24 | 176 |

| RT Cores | Not listed | 170 |

| Tensor Cores | 528 | 680 |

| Pixel Rate | 42.84 GPixel/s | 423.6 GPixel/s |

| Texture Rate | 942.5 GTexel/s | 1,636.8 GTexel/s |

| FP32 | 60.32 TFLOPS | 104.8 TFLOPS |

| FP16 | 120.6 TFLOPS (2:1) | 104.8 TFLOPS (1:1) |

| TDP | 600 W | 575 W |

| Power Connectors | 8-pin EPS | 1x 16-pin |

| Suggested PSU | 1000 W | 950 W |

| Display Outputs | No outputs | 1x HDMI 2.1b, 3x DisplayPort 2.1b |

| DirectX | N/A | 12 Ultimate (12_2) |

| OpenGL | N/A | 4.6 |

| Vulkan | N/A | 1.4 |

| Length | 267 mm 10.5 inches | 304 mm 12 inches |

| Height | 111 mm 4.4 inches | 137 mm 5.4 inches |

| Width | Not listed | 48 mm 1.9 inches |

| Release Date | 2024-11-17 | 2025-01-29 |

| Predecessor | Server Ada | GeForce 40 |

| Successor | Server Blackwell | GeForce 60 |

| Launch MSRP | Not listed | 2,299 USD |