NVIDIA GeForce RTX 5090 vs NVIDIA H20 NVL16 Comparison

FAQ

Q: What is the core architectural difference between the RTX 5090 and the H20 NVL16?

A: The RTX 5090 uses the GB202 chip built on Blackwell 2.0 architecture, while the H20 NVL16 uses the GH100 chip built on Hopper architecture. Both are fabricated by TSMC on a 5 nm process.

Q: How does memory capacity and bandwidth compare between the two cards?

A: The RTX 5090 has 32 GB of GDDR7 memory on a 512-bit bus, delivering 1.79 TB/s of bandwidth. The H20 NVL16 has 96 GB of HBM3 memory on a 6144-bit bus, delivering 4.03 TB/s of bandwidth, more than double the RTX 5090's bandwidth.

Q: Which card has higher raw FP32 compute performance?

A: The RTX 5090 delivers 104.8 TFLOPS of FP32 compute, while the H20 NVL16 delivers 39.54 TFLOPS. The RTX 5090 is roughly 2.65 times faster in this metric.

Q: What are the thermal design power (TDP) requirements for each card?

A: The RTX 5090 has a TDP of 575 W with a suggested PSU of 950 W, requiring a single 16-pin power connector. The H20 NVL16 has a TDP of 400 W with a suggested PSU of 800 W and comes as an SXM module with no discrete power connectors listed.

Q: Which card supports DirectX and Vulkan APIs?

A: The RTX 5090 supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. The H20 NVL16 lists N/A for DirectX, OpenGL, and Vulkan, reflecting its server-focused design with no display outputs.

Q: How do their benchmark scores compare?

A: The RTX 5090 has an average benchmark score of 79,842, placing it in the 92nd percentile of all GPUs. The H20 NVL16 has no recorded benchmark scores and sits at the 50th percentile with an average score of 0 in the database.

Where Each One Wins

The RTX 5090 dominates in every recorded benchmark category. Its 3DMark Steel Nomad DX12 score of 18,355, Geekbench OpenCL score of 334,370, and Geekbench Vulkan score of 376,728 all reflect substantial performance margins. In PassMark tests, the RTX 5090 scores 39,650 in G3D, 26,756 in GPU Compute, 1,413 in G2D, and shows strong legacy DirectX results: 395 in DX9, 341 in DX11, 226 in DX10, and 185 in DX12.

The H20 NVL16 has no benchmark entries in the database, so the data shows no recorded wins for this card. However, its specifications indicate a different purpose. The H20 NVL16 carries 96 GB of HBM3 memory with 4.03 TB/s of bandwidth, engineered for large model inference workloads that exceed the RTX 5090's 32 GB capacity. The RTX 5090 wins decisively on all measured performance tests, while the H20 NVL16 wins on memory capacity and memory bandwidth by a wide margin.

For gaming and client-side rendering workloads, the RTX 5090 is the clear choice based on its API support, display outputs, and benchmark scores. For server-side AI inference with very large models that require 96 GB of memory, the H20 NVL16 offers the capacity advantage, though without recorded benchmark data to quantify its compute performance in these tasks.

Architecture Differences

The RTX 5090 uses the GB202 chip on Blackwell 2.0 architecture, representing NVIDIA's latest client graphics design. It packs 92,200 million transistors on a 750 mm² die, achieving a transistor density of 122.9 million per mm². The H20 NVL16 uses the GH100 chip on Hopper architecture, which is NVIDIA's server-focused design. It contains 80,000 million transistors on a larger 814 mm² die, resulting in a lower density of 98.3 million per mm².

The RTX 5090 features 21,760 shading units, 680 texture mapping units, 176 ROPs, 170 ray tracing cores, and 680 tensor cores. The H20 NVL16 has 9,984 shading units, 312 TMUs, 24 ROPs, and 312 tensor cores, with no ray tracing cores listed. This disparity reflects their different design goals: the RTX 5090 prioritizes rasterization and ray tracing throughput for graphics workloads, while the H20 NVL16 prioritizes memory capacity and bandwidth for compute-heavy server tasks.

The RTX 5090 operates with a base clock of 2017 MHz and a boost clock of 2407 MHz, while the H20 NVL16 runs at a base clock of 1830 MHz and a boost clock of 1980 MHz. Both chips are fabricated on TSMC's 5 nm process, but their architectural priorities diverge significantly. The RTX 5090's pixel rate of 423.6 GPixel/s and texture rate of 1,636.8 GTexel/s vastly exceed the H20 NVL16's 47.52 GPixel/s and 617.8 GTexel/s, confirming the RTX 5090's graphics-first design.

The H20 NVL16's FP16 compute of 79.07 TFLOPS at a 2:1 ratio against its FP32 of 39.54 TFLOPS indicates a design optimized for mixed-precision AI workloads. The RTX 5090 delivers 104.8 TFLOPS for both FP16 and FP32 at a 1:1 ratio, showing a more balanced compute approach.

Specification Differences

The two cards differ across nearly every specification category. The RTX 5090 uses GDDR7 memory totaling 32 GB on a 512-bit bus, while the H20 NVL16 uses HBM3 memory totaling 96 GB on a 6144-bit bus. Memory bandwidth measures 1.79 TB/s for the RTX 5090 versus 4.03 TB/s for the H20 NVL16.

Clock speeds differ substantially: the RTX 5090 boosts to 2407 MHz versus 1980 MHz for the H20 NVL16. The RTX 5090's memory runs at 1750 MHz (28 Gbps effective), while the H20 NVL16's memory runs at 1313 MHz (5.3 Gbps effective). The RTX 5090 has a transistor count of 92,200 million on a 750 mm² die; the H20 NVL16 has 80,000 million transistors on an 814 mm² die.

The RTX 5090 is a dual-slot card measuring 304 mm in length, 137 mm in height, and 40 mm in width, with a 16-pin power connector, a 575 W TDP, and a 950 W suggested PSU. The H20 NVL16 is an SXM module with no listed dimensions, no power connectors, a 400 W TDP, and an 800 W suggested PSU. The RTX 5090 provides 1x HDMI 2.1b and 3x DisplayPort 2.1b outputs, while the H20 NVL16 has no display outputs.

API support also differs: the RTX 5090 supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, while the H20 NVL16 lists N/A for all three. The RTX 5090 was released on January 29, 2025, with a launch MSRP of 1,999 USD, while the H20 NVL16 was released on September 1, 2025. Both cards use PCIe 5.0 x16 interfaces and are currently marked as Active in production status.

Head-to-Head Benchmarks

The database contains no direct head-to-head benchmark comparisons between the RTX 5090 and the H20 NVL16. The RTX 5090 has ten recorded benchmark scores, while the H20 NVL16 has none.

The RTX 5090's performance profile is well-documented. Its 3DMark Steel Nomad DX12 score of 18,355 places it in the 92nd percentile of all GPUs. Geekbench results show 334,370 in OpenCL and 376,728 in Vulkan, indicating strong compute and graphics performance across different API paths. PassMark results show 39,650 in G3D and 26,756 in GPU Compute, with legacy DirectX scores of 395 in DX9, 341 in DX11, 226 in DX10, and 185 in DX12. The G2D score of 1,413 reflects 2D graphics throughput.

The nearest rivals for the RTX 5090 illustrate its standing. The NVIDIA Tesla P100 PCIe 16 GB scores an average of 79,605, which is 0.3% behind the RTX 5090's average of 79,842. The Tesla P100 PCIe 12 GB scores 79,396, 0.6% behind. The AMD Radeon RX 6850M XT scores 78,940, 1.1% behind. The AMD Radeon Pro Vega 64X scores 80,959, which is 1.4% ahead of the RTX 5090. These close margins indicate that the RTX 5090's average benchmark score sits within a tight competitive band, though the RTX 5090's individual test scores show it excels in modern workloads.

The H20 NVL16 has an average benchmark score of 0 and sits at the 50th percentile of all GPUs in the database, with no nearest rivals listed. This absence of data means the recorded measurements cannot quantify its real-world performance. The specification sheet indicates a memory-centric design, but benchmark results are unavailable for direct comparison.

The Verdict

The data presents two fundamentally different products. The RTX 5090 is a client graphics card with a 92nd percentile ranking, ten benchmark scores averaging 79,842, and support for DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4. The H20 NVL16 is a server accelerator with no recorded benchmarks, a 50th percentile ranking, and no API support or display outputs.

For anyone building a gaming or workstation PC with client graphics needs, the RTX 5090 is the only option with measurable performance in the database. Its 104.8 TFLOPS FP32, 1,636.8 GTexel/s texture rate, and 423.6 GPixel/s pixel rate provide concrete performance figures. Its 32 GB GDDR7 memory with 1.79 TB/s bandwidth supports demanding textures and datasets.

For server deployment requiring 96 GB of memory with 4.03 TB/s bandwidth, the H20 NVL16 offers specifications that the RTX 5090 cannot match. The 6144-bit memory bus and HBM3 memory type indicate a design focused on memory-bound AI inference. However, the database contains no benchmark scores to validate its compute performance. The H20 NVL16's FP32 of 39.54 TFLOPS and FP16 of 79.07 TFLOPS are lower than the RTX 5090's figures, but the memory advantage is substantial.

The RTX 5090's nearest rivals show it competes closely with older Tesla and Radeon Pro cards on average score, within 1.4% either direction. This context suggests that average benchmark scores alone do not capture the RTX 5090's strengths in modern DirectX 12 and Vulkan workloads, where its individual scores of 18,355 and 376,728 respectively demonstrate high performance.

The RTX 5090 has a launch MSRP of 1,999 USD, released January 29, 2025. The H20 NVL16 has no launch MSRP recorded, released September 1, 2025. Both use PCIe 5.0 x16 and 5 nm TSMC fabrication.

The choice depends on workload requirements. The RTX 5090 delivers measured graphics and compute performance across all recorded benchmarks, supports client display outputs, and fits a dual-slot form factor. The H20 NVL16 provides a memory capacity and bandwidth advantage with a 400 W TDP in an SXM module format, suited for server integration, but no benchmark data exists to assess its actual performance in the database.