NVIDIA H100 CNX

The NVIDIA H100 CNX is a server accelerator built on the Hopper architecture, fabricated on a 5 nm process at TSMC. The data sheet lists an 80,000 million transistor count on an 814 mm² die, with a base clock of 690 MHz and a boost clock of 1845 MHz. Its benchmark profile is unusual: the average benchmark score is 0, while the percentile versus all GPUs sits at 50. This combination suggests a part positioned exactly at the median of the database's performance distribution, yet with no recorded benchmark runs to substantiate a raw score. The release date is 2023-03-20T17:00:00.000Z, and production status is Active.

How It Compares

The nearestRivals field in the FACT PACK is empty, meaning no direct competitor scores are available for this entry. Consequently, the analysis must rely on the aggregate percentile and the raw specifications. A 50th percentile placement indicates that, in the absence of benchmark data, the H100 CNX sits exactly at the midpoint of all GPUs in the database. This is a neutral position, neither dominant nor lagging, but it is a provisional ranking pending actual workload results.

The database lists its predecessor as Server Ada and its successor as Server Blackwell, though no scores are provided for either, so a generational comparison cannot be quantified. Without rival scores, the H100 CNX's position is defined solely by its own specifications. The FP32 throughput of 53.84 TFLOPS and FP16 throughput of 215.4 TFLOPS (4:1) are the primary quantitative markers of its compute capability. These figures indicate a substantial compute pipeline, but the zero benchmark score means that no real-world test has confirmed these theoretical limits.

Given the lack of nearest rivals, the percentile of 50 is the only comparative metric. It implies that half of the GPUs in the database are above this point and half are below, based on the aggregate scoring system. However, the zero average benchmark score means that no actual benchmark has contributed to this percentile, so the 50 is a default or provisional ranking. This places the H100 CNX in a state of analytical limbo—its specifications are known, but its measured performance is unverified.

Memory Subsystem

The memory subsystem is a defining characteristic of the H100 CNX. It features 80 GB of HBM2e memory on a 5120-bit bus, delivering a bandwidth of 2.04 TB/s. The memory clock is listed at 1593 MHz, with an effective data rate of 3.2 Gbps. This combination of capacity and bandwidth is substantial for server workloads. For high-resolution or large-scale data processing, the 2.04 TB/s bandwidth allows rapid movement of data to and from the compute cores. The 80 GB capacity can hold large models or datasets that would otherwise require multiple smaller GPUs.

In the context of high-resolution rendering or scientific simulation, the memory bandwidth is often the bottleneck. The H100 CNX's 2.04 TB/s figure is a key specification that mitigates this. The 5120-bit bus width is exceptionally wide, which contributes to the high bandwidth. This is not a consumer graphics card; it lacks display outputs, so the memory is used exclusively for compute. The effective data rate of 3.2 Gbps is the transfer rate per pin, and the aggregate bandwidth is the product of the bus width and the data rate. The 80 GB capacity is particularly notable, as it allows entire large neural networks to reside in memory without partitioning.

Ray Tracing and Feature Set

The ray tracing capabilities of the H100 CNX are not quantified in the data sheet. The RT cores field is null, meaning no dedicated ray tracing hardware is listed. This is consistent with a server accelerator that prioritizes compute over graphics rendering. The tensor core count, however, is explicitly listed at 456. These tensor cores are designed for matrix operations, which are fundamental to AI and deep learning workloads. The FP16 performance of 215.4 TFLOPS (4:1) indicates a high throughput for half-precision operations, which are common in AI models.

The API support is entirely absent: DirectX, OpenGL, and Vulkan are all null. This indicates that the H100 CNX does not expose a graphics API interface. Furthermore, the display outputs field is 'No outputs', confirming that this is a compute-only device. The feature set is therefore focused on FP32, FP16, and tensor operations. The pixel rate is 44.28 GPixel/s and the texture rate is 841.3 GTexel/s, but these are likely theoretical figures for the compute pipeline, as there are no rasterization outputs. The 456 tensor cores are a significant asset for machine learning inference and training, while the absence of RT cores means that any ray tracing would have to be performed via compute shaders, which is inefficient but outside the intended use case.

FAQ

Q: What is the memory capacity and type of the NVIDIA H100 CNX?

A: It has 80 GB of HBM2e memory.

Q: What is the TDP of the H100 CNX?

A: The TDP is 350 W.

Q: What is the boost clock speed?

A: The boost clock is 1845 MHz.

Q: How many tensor cores does it have?

A: It has 456 tensor cores.

Q: What is the FP32 performance in TFLOPS?

A: The FP32 performance is 53.84 TFLOPS.

Q: Does it have display outputs?

A: No, it has no display outputs.

Who Should Consider It

The H100 CNX is intended for server environments where display output is unnecessary. The data shows a compute-focused design with 80 GB of HBM2e memory and a 2.04 TB/s bandwidth. This makes it suitable for workloads that require large memory footprints, such as training large language models or processing scientific datasets. The FP16 throughput of 215.4 TFLOPS (4:1) is particularly relevant for AI applications that use mixed-precision training. Given the 50th percentile placement and the zero benchmark score, there is no empirical performance data to recommend specific resolution or settings. However, the specifications suggest that it can handle high-resolution compute tasks without memory swapping.

The 14592 shading units and 456 TMUs provide a substantial compute pipeline, though the 24 ROPs are low, indicating that pixel output is not a priority. For users who need a compute accelerator for data center deployment, the H100 CNX offers a large memory pool and high bandwidth. It is not suitable for gaming or desktop graphics, as it lacks display outputs and graphics API support. The absence of RT cores and display outputs means that it is strictly for compute. The 5 nm process and 80,000 million transistors on an 814 mm² die indicate a high transistor density of 98.3M / mm². This density supports the high compute throughput. The suggested PSU is 750 W, which is the recommended power supply for a system incorporating this card.

Power and Cooling

The H100 CNX has a TDP of 350 W. The suggested power supply is 750 W, which provides headroom for the rest of the system. The power connector is an 8-pin EPS, which is a server-grade connector. The card is dual-slot in width, with a length of 267 mm (10.5 inches) and a height of 111 mm (4.4 inches). These dimensions are standard for a dual-slot server accelerator. Cooling is handled by the dual-slot design, which typically allows for a passive heatsink with airflow from server chassis fans. The 350 W TDP requires adequate cooling to maintain the boost clock of 1845 MHz. The base clock is 690 MHz, so the boost clock is significantly higher, indicating a wide dynamic range under load.

The power connector is 8-pin EPS, which differs from the common 8-pin PCIe connectors found on consumer cards, so compatibility with server power supplies is expected. The lack of display outputs simplifies the power and cooling design, as there is no need for video output circuitry. The 750 W PSU recommendation is a system-level figure, accounting for the CPU and other components. The dimensions and slot width are important for chassis compatibility. The PCIe 5.0 x16 bus interface ensures high data transfer rates to the host system, which is critical for compute workloads that stream data to and from the GPU.