H100 PCIe: specs and benchmarks

Summary

NVIDIA started H100 PCIe sales 22 March 2022. This is a Hopper architecture desktop card based on 4 nm manufacturing process and primarily aimed at designers. 80 GB of HBM2e memory clocked at 1 GHz are supplied, and together with 5120 Bit memory interface this creates a bandwidth of 1,280 GB/s.

Compatibility-wise, this is a dual-slot graphics card attached via PCIe 5.0 x16 interface. Its manufacturer default version has a length of 267 mm. 8-pin EPS power connector is required, and power consumption is at 350 Watt.

Primary details

Some basic facts about H100 PCIe: architecture, market segment, release date etc.

Place in the ranking	not rated
Place by popularity	not in top-100
Architecture	Hopper (2022−2024)
GPU code name	GH100
Market segment	Workstation
Release date	22 March 2022 (3 years ago)

Detailed specifications

H100 PCIe's specs such as number of shaders, GPU base clock, manufacturing process, texturing and calculation speed. These parameters indirectly speak of H100 PCIe's performance, but for precise assessment you have to consider its benchmark and gaming test results.

Pipelines / CUDA cores	7296	of 24064 (RTX PRO 6000 Blackwell)
Core clock speed	1065 MHz	of 2670 MHz (Arc B580)
Boost clock speed	1650 MHz	of 3599 MHz (Radeon RX 7990 XTX)
Number of transistors	80,000 million	of 153,000 million (Radeon Instinct MI300)
Manufacturing process technology	4 nm	of 3 nm (GeForce RTX 5060 Laptop)
Power consumption (TDP)	350 Watt	of 2400 Watt (Data Center GPU Max Subsystem)
Texture fill rate	752.4	of 2,554 (Radeon Instinct MI300X)
Floating-point processing power	24.08 TFLOPS	of 126 (RTX PRO 6000 Blackwell)
ROPs	24	of 512 (Moore Threads MTT S4000)
TMUs	456	of 1280 (Data Center GPU Max NEXT)
Tensor Cores	456	of 1216 (Radeon Instinct MI300X)

Form factor & compatibility

This section provides details about the physical dimensions of H100 PCIe and its compatibility with other computer components. This information is useful when selecting a computer configuration or upgrading an existing one. For desktop graphics cards, it includes details about the interface and bus (for motherboard compatibility) and additional power connectors (for power supply compatibility).

Interface	PCIe 5.0 x16
Length	267 mm
Width	2-slot
Supplementary power connectors	8-pin EPS

VRAM capacity and type

Parameters of memory installed on H100 PCIe: its type, size, bus, clock and resulting bandwidth. Note that GPUs integrated into processors have no dedicated memory and use a shared part of system RAM instead.

Memory type	HBM2e
Maximum RAM amount	80 GB	of 288 GB (Radeon Instinct MI325X)
Memory bus width	5120 Bit	of 8192 Bit (Radeon Instinct MI250X)
Memory clock speed	1000 MHz	of 20000 (RTX 5000 Ada Generation Mobile)
Memory bandwidth	1,280 GB/s	of 5,171 GB/s (Radeon Instinct MI300X)
Resizable BAR	+

Connectivity and outputs

Types and number of video connectors present on H100 PCIe. As a rule, this section is relevant only for desktop reference graphics cards, since for notebook ones the availability of certain video outputs depends on the laptop model, while non-reference desktop models can (though not necessarily will) bear a different set of video ports.

Display Connectors

No outputs

API and SDK compatibility

APIs supported by H100 PCIe, sometimes including their particular versions.

DirectX	N/A
Shader Model	N/A
OpenGL	N/A	of 4.6 (GeForce RTX 5090 D)
OpenCL	3.0
Vulkan	N/A
CUDA	9.0
DLSS	+

Benchmark performance

Synthetic benchmark performance of H100 PCIe. The combined score is measured on a 0-100 point scale.

---

---