NVIDIA H100 CNX vs NVIDIA Rubin GPU Comparison

Head-to-Head Benchmarks

The database records show no direct benchmark comparisons between the NVIDIA H100 CNX and the NVIDIA Rubin GPU. Both entries carry an average benchmark score of zero, and the head-to-head benchmark list is empty. The percentile versus all GPUs is identical for both at 50, indicating neither part has accumulated measurable performance data in the database at this time.

Given the absence of recorded benchmark results, the comparison must rest on the architectural specifications that determine theoretical peak performance. The FP32 compute figures provide the clearest starting point. The NVIDIA Rubin GPU delivers 130.0 TFLOPS of FP32 throughput, while the H100 CNX delivers 53.84 TFLOPS. The Rubin GPU leads by a factor of approximately 2.4 in raw single-precision compute. In FP16, the Rubin GPU records 260.0 TFLOPS (2:1), while the H100 CNX records 215.4 TFLOPS (4:1). The Rubin GPU leads by roughly 21% in FP16, though the H100 CNX achieves its FP16 figure at a 4:1 ratio, meaning that the H100 CNX's FP16 rate is derived from its tensor cores at a higher reduction factor.

Texture throughput follows a similar pattern. The Rubin GPU posts 2,031.2 GTexel/s, which is more than double the H100 CNX's 841.3 GTexel/s. Pixel rate favors the Rubin GPU less decisively: 54.41 GPixel/s versus 44.28 GPixel/s, a lead of roughly 23% for the Rubin part. The shading unit count explains the compute gap: the Rubin GPU packs 28,672 shading units against 14,592 on the H100 CNX. Tensor core counts also double: 896 on Rubin versus 456 on H100 CNX.

Memory bandwidth presents the largest proportional difference. The Rubin GPU offers 22.1 TB/s of bandwidth, which is over ten times the H100 CNX's 2.04 TB/s. The Rubin GPU's memory subsystem uses 288 GB of HBM4 across a 16,384-bit bus, while the H100 CNX uses 80 GB of HBM2e across a 5,120-bit bus. The Rubin GPU's memory clock runs at 2,695 MHz with 10.8 Gbps effective data rate, while the H100 CNX runs at 1,593 MHz with 3.2 Gbps effective.

Clock speeds show the Rubin GPU operating at a higher boost frequency: 2,267 MHz versus 1,845 MHz. Base clocks are closer, with the Rubin GPU at 700 MHz and the H100 CNX at 690 MHz. The Rubin GPU's higher boost clock contributes to its compute lead, but the doubling of execution resources is the dominant factor.

The transistor and process node data reinforce the generational gap. The Rubin GPU uses a 3 nm TSMC process and contains 336,000 million transistors on a 1,456 mm² die. The H100 CNX uses a 5 nm TSMC process and contains 80,000 million transistors on an 814 mm² die. Transistor density on the Rubin GPU is 230.8M per mm², versus 98.3M per mm² on the H100 CNX. The Rubin GPU carries more than four times the transistor count of the H100 CNX.

Power draw scales accordingly. The Rubin GPU has a TDP of 2,300 W, while the H100 CNX is rated at 350 W. The suggested power supply for the Rubin GPU is 2,700 W, while the H100 CNX suggests 750 W. The Rubin GPU is a SXM module, while the H100 CNX is a dual-slot card with an 8-pin EPS power connector. Neither part has display outputs.

The Verdict

The data indicates a clear performance hierarchy. The NVIDIA Rubin GPU is the more capable compute device on every measurable specification in the database. Its FP32 throughput is 2.4 times higher, its FP16 throughput is 21% higher, its texture rate is 2.4 times higher, its pixel rate is 23% higher, and its memory bandwidth is over ten times higher. The Rubin GPU also carries more than double the shading units, tensor cores, and texture mapping units, and it has 3.6 times the memory capacity.

The H100 CNX is not without its strengths. It draws far less power: 350 W versus 2,300 W. It uses a conventional dual-slot PCIe 5.0 x16 form factor, whereas the Rubin GPU is an SXM module with PCIe 6.0 x16. The H100 CNX has a longer recorded physical footprint at 267 mm, but the Rubin GPU has no recorded dimensions. The H100 CNX also has a lower transistor count and a larger process node, which may be relevant in contexts where thermal density or manufacturing constraints matter.

The verdict from the database is unambiguous: for raw compute and memory throughput, the Rubin GPU is the superior part. For power-constrained deployments or systems requiring a standard PCIe card form factor, the H100 CNX holds an advantage. The H100 CNX's lower TDP and slot design make it more adaptable to existing server infrastructure, while the Rubin GPU demands a dedicated high-power SXM environment.

Architecture Differences

The two GPUs come from different architectural generations. The H100 CNX uses the Hopper architecture with the GH100 chip, while the Rubin GPU uses the Rubin architecture with the GR100 chip. The H100 CNX belongs to the Server Hopper generation (Hxx), and the Rubin GPU belongs to the Server Rubin generation (Rxx). The predecessor and successor relationships in the database place the Rubin GPU after the H100 CNX: the H100 CNX's successor is listed as Server Blackwell, and the Rubin GPU's predecessor is Server Blackwell.

Manufacturing processes differ significantly. The H100 CNX is built on a 5 nm TSMC process, while the Rubin GPU is built on a 3 nm TSMC process. Transistor counts diverge sharply: 80,000 million on the H100 CNX versus 336,000 million on the Rubin GPU. Die sizes also differ, with the H100 CNX at 814 mm² and the Rubin GPU at 1,456 mm². The transistor density on the Rubin GPU is more than double that of the H100 CNX.

Memory technology is a major architectural split. The H100 CNX uses HBM2e memory at 80 GB, while the Rubin GPU uses HBM4 memory at 288 GB. The memory bus on the Rubin GPU is 16,384 bits, three times wider than the H100 CNX's 5,120 bits. The memory clock on the Rubin GPU runs at 2,695 MHz with 10.8 Gbps effective data rate, while the H100 CNX runs at 1,593 MHz with 3.2 Gbps effective. The resulting bandwidth gap is substantial.

Compute resources scale with architecture. The Rubin GPU has 28,672 shading units, 896 texture mapping units, and 896 tensor cores. The H100 CNX has 14,592 shading units, 456 texture mapping units, and 456 tensor cores. The Rubin GPU operates at a higher boost clock of 2,267 MHz versus 1,845 MHz. The FP16 ratio also differs: the H100 CNX uses a 4:1 ratio, while the Rubin GPU uses a 2:1 ratio, meaning the Rubin GPU's FP16 throughput is achieved with less reduction overhead.

Bus interface and form factor differ. The H100 CNX uses PCIe 5.0 x16, while the Rubin GPU uses PCIe 6.0 x16. The H100 CNX is a dual-slot card with an 8-pin EPS power connector and a suggested 750 W power supply. The Rubin GPU is an SXM module with no recorded power connector and a suggested 2,700 W power supply. The H100 CNX has recorded dimensions of 267 mm length and 111 mm height, while the Rubin GPU has no recorded dimensions.

API support is also distinct. The Rubin GPU lists DirectX, OpenGL, and Vulkan as N/A, while the H100 CNX lists no API data at all. Both GPUs have no display outputs, confirming their server-oriented design.

FAQ

Q: Which GPU has higher FP32 performance?

A: The NVIDIA Rubin GPU delivers 130.0 TFLOPS of FP32 compute, while the NVIDIA H100 CNX delivers 53.84 TFLOPS. The Rubin GPU leads by a factor of approximately 2.4.

Q: How do the memory bandwidth figures compare?

A: The Rubin GPU offers 22.1 TB/s of bandwidth across a 16,384-bit HBM4 bus, while the H100 CNX offers 2.04 TB/s across a 5,120-bit HBM2e bus. The Rubin GPU's bandwidth is over ten times higher.

Q: What is the transistor count difference?

A: The Rubin GPU contains 336,000 million transistors on a 1,456 mm² die, while the H100 CNX contains 80,000 million transistors on an 814 mm² die. The Rubin GPU has more than four times the transistor count.

Q: Are these GPUs suitable for display output?

A: No. Both the H100 CNX and the Rubin GPU have no display outputs, indicating they are designed for compute workloads rather than graphics rendering.

Q: What power supply is recommended for each?

A: The H100 CNX suggests a 750 W power supply, while the Rubin GPU suggests a 2,700 W power supply. The H100 CNX has a TDP of 350 W, and the Rubin GPU has a TDP of 2,300 W.

Q: Which GPU uses the newer process node?

A: The Rubin GPU uses a 3 nm TSMC process, while the H100 CNX uses a 5 nm TSMC process. The Rubin GPU also achieves a higher transistor density of 230.8M per mm² versus 98.3M per mm².

Where Each One Wins

The NVIDIA Rubin GPU wins in every compute-oriented category in the database. Its FP32 throughput of 130.0 TFLOPS is 2.4 times the H100 CNX's 53.84 TFLOPS. Its FP16 throughput of 260.0 TFLOPS is 21% higher than the H100 CNX's 215.4 TFLOPS. Texture rate favors the Rubin GPU at 2,031.2 GTexel/s versus 841.3 GTexel/s. Pixel rate favors the Rubin GPU at 54.41 GPixel/s versus 44.28 GPixel/s. Memory bandwidth favors the Rubin GPU at 22.1 TB/s versus 2.04 TB/s. Memory capacity favors the Rubin GPU at 288 GB versus 80 GB. The Rubin GPU also has double the shading units, tensor cores, and texture mapping units, and it runs at a higher boost clock of 2,267 MHz versus 1,845 MHz.

The NVIDIA H100 CNX wins in power efficiency and form factor flexibility. Its TDP of 350 W is drastically lower than the Rubin GPU's 2,300 W. Its suggested power supply of 750 W is far below the Rubin GPU's 2,700 W. The H100 CNX is a dual-slot PCIe card with an 8-pin EPS connector, which fits into standard server chassis, while the Rubin GPU is an SXM module that requires a specialized high-power socket. The H100 CNX also has recorded physical dimensions of 267 mm length and 111 mm height, while the Rubin GPU has no recorded dimensions, suggesting the H100 CNX is designed for more conventional installation.

The H100 CNX also holds a transistor density advantage in one narrow sense: while the Rubin GPU has higher density overall, the H100 CNX's lower absolute transistor count may be easier to manufacture or yield at scale, though the database does not record yield data. The H100 CNX's base clock of 690 MHz is close to the Rubin GPU's 700 MHz, but the Rubin GPU's boost clock advantage is substantial.

Specification Differences

The two GPUs differ in nearly every recorded specification field. The chip names differ: GH100 for the H100 CNX versus GR100 for the Rubin GPU. The architecture differs: Hopper versus Rubin. The generation differs: Server Hopper (Hxx) versus Server Rubin (Rxx). The process node differs: 5 nm versus 3 nm. The foundry is the same, TSMC, for both.

Transistor count differs: 80,000 million versus 336,000 million. Die size differs: 814 mm² versus 1,456 mm². Transistor density differs: 98.3M per mm² versus 230.8M per mm². Base clock differs: 690 MHz versus 700 MHz. Boost clock differs: 1,845 MHz versus 2,267 MHz. Memory clock differs: 1,593 MHz with 3.2 Gbps effective versus 2,695 MHz with 10.8 Gbps effective.

Memory size differs: 80 GB versus 288 GB. Memory type differs: HBM2e versus HBM4. Memory bus width differs: 5,120 bit versus 16,384 bit. Memory bandwidth differs: 2.04 TB/s versus 22.1 TB/s. Shading units differ: 14,592 versus 28,672. Texture mapping units differ: 456 versus 896. Raster operations pipelines are the same at 24. Tensor cores differ: 456 versus 896. Pixel rate differs: 44.28 GPixel/s versus 54.41 GPixel/s. Texture rate differs: 841.3 GTexel/s versus 2,031.2 GTexel/s. FP32 differs: 53.84 TFLOPS versus 130.0 TFLOPS. FP16 differs: 215.4 TFLOPS (4:1) versus 260.0 TFLOPS (2:1).

TDP differs: 350 W versus 2,300 W. Slot width differs: dual-slot versus SXM Module. Power connectors differ: 8-pin EPS versus none recorded. Suggested power supply differs: 750 W versus 2,700 W. Bus interface differs: PCIe 5.0 x16 versus PCIe 6.0 x16. Display outputs are the same: none. APIs differ: no data for the H100 CNX versus N/A for DirectX, OpenGL, and Vulkan on the Rubin GPU. Dimensions differ: the H100 CNX has recorded length of 267 mm and height of 111 mm, while the Rubin GPU has no recorded dimensions. Release dates differ: 2023-03-20 for the H100 CNX versus 2025-12-31 for the Rubin GPU. Predecessors differ: Server Ada for the H100 CNX versus Server Blackwell for the Rubin GPU. Successors differ: Server Blackwell for the H100 CNX versus none recorded for the Rubin GPU.