GEFORCE

NVIDIA H100 PCIe 96 GB

Name: NVIDIA H100 PCIe 96 GB
Brand: NVIDIA

NVIDIA graphics card specifications and benchmark scores

96 GB

VRAM

1837

MHz Boost

700W

TDP

5120

Bus Width

Tensor Cores

At a Glance

NVIDIA

VRAM 96 GB

Boost Clock 1,837 MHz

Shaders 16,896

Bus Width 5120-bit

TDP 700W

Memory Type HBM3

Architecture Hopper

Process 5 nm

Released Mar 2023

NVIDIA H100 PCIe 96 GB Specifications

H100 PCIe 96 GB GPU Core

Shader units and compute resources

The NVIDIA H100 PCIe 96 GB GPU core specifications define its raw processing power for graphics and compute workloads. Shading units (also called CUDA cores, stream processors, or execution units depending on manufacturer) handle the parallel calculations required for rendering. TMUs (Texture Mapping Units) process texture data, while ROPs (Render Output Units) handle final pixel output. Higher shader counts generally translate to better GPU benchmark performance, especially in demanding games and 3D applications.

Shading Units

16,896

Shaders

16,896

TMUs

528

ROPs

SM Count

132

H100 PCIe 96 GB Clock Speeds

GPU and memory frequencies

Clock speeds directly impact the H100 PCIe 96 GB's performance in GPU benchmarks and real-world gaming. The base clock represents the minimum guaranteed frequency, while the boost clock indicates peak performance under optimal thermal conditions. Memory clock speed affects texture loading and frame buffer operations. The H100 PCIe 96 GB by NVIDIA dynamically adjusts frequencies based on workload, temperature, and power limits to maximize performance while maintaining stability.

Base Clock

1665 MHz

Base Clock

1,665 MHz

Boost Clock

1837 MHz

Boost Clock

1,837 MHz

Memory Clock

1313 MHz 5.3 Gbps effective

NVIDIA's H100 PCIe 96 GB Memory

VRAM capacity and bandwidth

VRAM (Video RAM) is dedicated memory for storing textures, frame buffers, and shader data. The H100 PCIe 96 GB's memory capacity determines how well it handles high-resolution textures and multiple displays. Memory bandwidth, measured in GB/s, affects how quickly data moves between the GPU and VRAM. Higher bandwidth improves performance in memory-intensive scenarios like 4K gaming. The memory bus width and type (GDDR6, GDDR6X, HBM) significantly influence overall GPU benchmark scores.

Memory Size

96 GB

VRAM

98,304 MB

Memory Type

HBM3

VRAM Type

HBM3

Memory Bus

5120 bit

Bus Width

5120-bit

Bandwidth

3.36 TB/s

H100 PCIe 96 GB by NVIDIA Cache

On-chip cache hierarchy

On-chip cache provides ultra-fast data access for the H100 PCIe 96 GB, reducing the need to fetch data from slower VRAM. L1 and L2 caches store frequently accessed data close to the compute units. AMD's Infinity Cache (L3) dramatically increases effective bandwidth, improving GPU benchmark performance without requiring wider memory buses. Larger cache sizes help maintain high frame rates in memory-bound scenarios and reduce power consumption by minimizing VRAM accesses.

L1 Cache

256 KB (per SM)

L2 Cache

50 MB

H100 PCIe 96 GB Theoretical Performance

Compute and fill rates

Theoretical performance metrics provide a baseline for comparing the NVIDIA H100 PCIe 96 GB against other graphics cards. FP32 (single-precision) performance, measured in TFLOPS, indicates compute capability for gaming and general GPU workloads. FP64 (double-precision) matters for scientific computing. Pixel and texture fill rates determine how quickly the GPU can render complex scenes. While real-world GPU benchmark results depend on many factors, these specifications help predict relative performance levels.

FP32 (Float)

62.08 TFLOPS

FP64 (Double)

31.04 TFLOPS (1:2)

FP16 (Half)

248.3 TFLOPS (4:1)

Pixel Rate

44.09 GPixel/s

Texture Rate

969.9 GTexel/s

H100 PCIe 96 GB Ray Tracing & AI

Hardware acceleration features

The NVIDIA H100 PCIe 96 GB includes dedicated hardware for ray tracing and AI acceleration. RT cores handle real-time ray tracing calculations for realistic lighting, reflections, and shadows in supported games. Tensor cores (NVIDIA) or XMX cores (Intel) accelerate AI workloads including DLSS, FSR, and XeSS upscaling technologies. These features enable higher visual quality without proportional performance costs, making the H100 PCIe 96 GB capable of delivering both stunning graphics and smooth frame rates in modern titles.

Tensor Cores

528

Hopper Architecture & Process

Manufacturing and design details

The NVIDIA H100 PCIe 96 GB is built on NVIDIA's Hopper architecture, which defines how the GPU processes graphics and compute workloads. The manufacturing process node affects power efficiency, thermal characteristics, and maximum clock speeds. Smaller process nodes pack more transistors into the same die area, enabling higher performance per watt. Understanding the architecture helps predict how the H100 PCIe 96 GB will perform in GPU benchmarks compared to previous generations.

Architecture

Hopper

GPU Name

GH100

Process Node

5 nm

Foundry

TSMC

Transistors

80,000 million

Die Size

814 mm²

Density

98.3M / mm²

NVIDIA's H100 PCIe 96 GB Power & Thermal

TDP and power requirements

Power specifications for the NVIDIA H100 PCIe 96 GB determine PSU requirements and thermal management needs. TDP (Thermal Design Power) indicates the heat output under typical loads, guiding cooler selection. Power connector requirements ensure adequate power delivery for stable operation during demanding GPU benchmarks. The suggested PSU wattage accounts for the entire system, not just the graphics card. Efficient power delivery enables the H100 PCIe 96 GB to maintain boost clocks without throttling.

TDP

700 W

TDP

700W

Power Connectors

8-pin EPS

Suggested PSU

1100 W

H100 PCIe 96 GB by NVIDIA Physical & Connectivity

Dimensions and outputs

Physical dimensions of the NVIDIA H100 PCIe 96 GB are critical for case compatibility. Card length, height, and slot width determine whether it fits in your chassis. The PCIe interface version affects bandwidth for communication with the CPU. Display outputs define monitor connectivity options, with modern cards supporting multiple high-resolution displays simultaneously. Verify these specifications against your case and motherboard before purchasing to ensure a proper fit.

Slot Width

Dual-slot

Length

268 mm 10.6 inches

Height

111 mm 4.4 inches

Bus Interface

PCIe 5.0 x16

Display Outputs

No outputs

Display Outputs

No outputs

NVIDIA API Support

Graphics and compute APIs

API support determines which games and applications can fully utilize the NVIDIA H100 PCIe 96 GB. DirectX 12 Ultimate enables advanced features like ray tracing and variable rate shading. Vulkan provides cross-platform graphics capabilities with low-level hardware access. OpenGL remains important for professional applications and older games. CUDA (NVIDIA) and OpenCL enable GPU compute for video editing, 3D rendering, and scientific applications. Higher API versions unlock newer graphical features in GPU benchmarks and games.

OpenCL

3.0

CUDA

9.0

H100 PCIe 96 GB Product Information

Release and pricing details

The NVIDIA H100 PCIe 96 GB is manufactured by NVIDIA as part of their graphics card lineup. Release date and launch pricing provide context for comparing GPU benchmark results with competing products from the same era. Understanding the product lifecycle helps evaluate whether the H100 PCIe 96 GB by NVIDIA represents good value at current market prices. Predecessor and successor information aids in tracking generational improvements and planning future upgrades.

Manufacturer

NVIDIA

Release Date

Mar 2023

Production

Active

Predecessor

Server Ada

Successor

Server Blackwell

H100 PCIe 96 GB Benchmark Scores

No benchmark data available for this GPU.

About NVIDIA H100 PCIe 96 GB

The NVIDIA H100 PCIe 96 GB is a server-grade accelerator built on the Hopper architecture and the GH100 chip, fabricated by TSMC on a 5 nm process. It carries 80,000 million transistors on an 814 mm² die, giving a transistor density of 98.3 million per square millimeter. The card ships with 96 GB of HBM3 memory across a 5120-bit bus, and its memory clock is listed at 1313 MHz, or 5.3 Gbps effective, yielding 3.36 TB/s of bandwidth. Power draw is rated at 700 W with an 1100 W suggested PSU, and the board uses a dual-slot layout with an 8-pin EPS power connector. It connects via PCIe 5.0 x16, has no display outputs, and was released on 2023-03-20. The production status is Active, with the predecessor listed as Server Ada and the successor as Server Blackwell.

Benchmark Performance

The fact pack contains no benchmark scores for the H100 PCIe 96 GB. The benchmarks array is empty, and the average benchmark score is recorded as 0. The percentile versus all GPUs is 50, which places this part at the median of the database's ranking distribution, but without any actual measured results, that percentile carries no comparative weight. In other words, the database has no performance entries to analyze, and no nearest rivals are listed either. The nearestRivals field is empty, so there are no delta percentages to report against competing accelerators.

What the fact pack does provide is raw compute throughput. The FP32 rate is 62.08 TFLOPS, and the FP16 rate is 248.3 TFLOPS at a 4:1 ratio. These figures are derived from the shading units and tensor cores rather than from application benchmarks. The shading unit count is 16896, with 528 TMUs and 24 ROPs. The texture rate is 969.9 GTexel/s and the pixel rate is 44.09 GPixel/s. These are hardware specifications, not measured scores, so they describe the theoretical ceiling of the chip rather than its real-world standing.

Because there are no benchmark entries, any statement about how this card performs relative to other products would be unsupported by the data. The percentile of 50 suggests a median placement in the database, but the zero average score indicates that no results have been aggregated. This is a common situation for server accelerators that are not subjected to the same test suites as consumer cards. The absence of data is itself a finding: the database does not currently rank this part against any rivals, and no performance deltas can be computed.

Memory Subsystem

The memory subsystem is the most distinctive part of this specification. The H100 PCIe 96 GB carries 96 GB of HBM3 memory, which is a high-bandwidth stack design rather than a conventional layout. The bus width is 5120 bit, which is substantially wider than typical memory buses, and the memory clock is 1313 MHz with a 5.3 Gbps effective data rate. The resulting bandwidth is 3.36 TB/s.

For high-resolution workloads, bandwidth of this magnitude matters because the data throughput scales with the number of pixels and the complexity of the scene. A 5120-bit bus moves data in very wide bursts, which reduces the number of memory transactions needed to feed the compute units. The 96 GB capacity is large enough to hold substantial datasets in memory without spilling to system RAM. In server contexts, this capacity is often more relevant than in gaming, where smaller capacities are common. The HBM3 type also has different thermal and power characteristics than other memory types, though the fact pack does not provide a separate memory power figure.

The memory clock of 1313 MHz is the base clock for the memory, and the effective rate of 5.3 Gbps accounts for the double data rate behavior. The bandwidth calculation is the product of the bus width and the effective data rate. With 5120 bits and 5.3 Gbps effective, the 3.36 TB/s figure follows. This is the kind of bandwidth that supports large framebuffers, high-resolution textures, and compute kernels that stream through memory repeatedly. Without benchmark data, the practical impact on resolution scaling cannot be quantified, but the specifications indicate that memory bandwidth is not a bottleneck in the traditional sense.

Ray Tracing and Feature Set

The fact pack lists the tensor core count as 528, but the RT core count is null. This means the database does not record a dedicated ray tracing core count for this part. The architecture is Hopper, which is NVIDIA's server-focused design, and the chip is GH100. The API fields for DirectX, OpenGL, and Vulkan are all null, indicating that no API support is recorded in the fact pack. The card has no display outputs, so it is not intended to drive a monitor directly.

The tensor cores are the notable compute feature. With 528 tensor cores, the card is oriented toward matrix operations, which are the foundation of deep learning training and inference. The FP16 throughput of 248.3 TFLOPS at a 4:1 ratio reflects the tensor core path, while the FP32 rate of 62.08 TFLOPS reflects the standard shading path. The 4:1 ratio means that FP16 operations run four times faster than FP32 on the same hardware, which is a common design for AI workloads that tolerate reduced precision.

Ray tracing, in the traditional gaming sense, is not represented in the data. The null RT core count and the null API entries mean there is no information about hardware ray tracing acceleration or graphics API compatibility. This is consistent with a server accelerator that has no display outputs and is designed for compute rather than rendering. The feature set, as recorded, is therefore compute-focused: tensor cores for AI, a wide memory bus for data movement, and no graphics output path.

Who Should Consider It

Based on the specifications, this card is aimed at server and datacenter environments rather than desktop gaming. The 700 W TDP and the 1100 W suggested PSU are high power requirements that typically demand server chassis and power delivery systems. The dual-slot design and the 268 mm length (10.6 inches) with a 111 mm height (4.4 inches) are physical constraints that matter for server rack fit. The 8-pin EPS power connector is a server-style connector, not the connector common on consumer cards.

The absence of display outputs is decisive: this card cannot drive a monitor, so it is not for interactive use. The compute throughput — 62.08 TFLOPS FP32 and 248.3 TFLOPS FP16 — along with the 96 GB HBM3 memory, points to workloads that need large memory capacity and high bandwidth, such as large-scale model training, scientific simulation, or data processing. The fact pack does not include resolution-based performance data, so no recommendation can be made about gaming at specific resolutions. Instead, the card should be considered by users who need a server accelerator with a specific memory footprint and compute profile, and who have the power and cooling infrastructure to support a 700 W board.

The release date of 2023-03-20 places it in the Hopper generation, with Server Ada as its predecessor and Server Blackwell as its successor. The production status is Active, so it is still a current product. Users considering this card should verify that their software stack supports the Hopper architecture and that their power delivery meets the 700 W requirement with an 1100 W PSU recommendation.

How It Compares

The nearestRivals field in the fact pack is empty. There are no rival names, no scores, and no delta percentages to report. Consequently, this section cannot compare the H100 PCIe 96 GB to any specific competing product using database values. The only positional information is the percentile versus all GPUs, which is 50, and the average benchmark score of 0. The percentile indicates a median placement in the overall database, but with no benchmark entries, that placement is not based on measured performance.

The predecessor is listed as Server Ada and the successor as Server Blackwell. These are generation-level labels, not rival products with scores. The fact pack does not provide any benchmark data for either the predecessor or the successor, so no performance trajectory can be established. The card sits between these two server generations in the product line, but the database records no comparative numbers.

In the absence of rival data, the only honest statement is that the database currently holds no comparison points for this accelerator. Any attempt to rank it against other accelerators would require figures that are not present in the fact pack. The percentile of 50 is a placeholder derived from an empty benchmark set, so it should not be interpreted as a meaningful ranking.