NVIDIA H100 SXM5 vs NVIDIA GeForce RTX 3090 Ti. Specifications, performance, tests

Comparison of basic technical data of graphics cards NVIDIA H100 SXM5 and NVIDIA GeForce RTX 3090 Ti, chip, information processing units.

• Architecture
  Hopper Ampere
• Execution Units
  8448 -
• GPU Chip
  GH100 GA102-350-A1
• L1 Cache
  192 KB 128 KB
• L2 Cache
  50 MB 6 MB
• Render Output Units
  24 112
• Texture Units
  528 336
• Based on
  - GA102
• CUDA Cores
  - 10752
• Shader
  - 10752

Video memory capacity, type, bus width and bandwidth describe different properties. These values alone do not determine graphics performance.

• Memory Bandwidth
  1,920 GB/s 1,008 GB/s
• Memory Bus
  5120 bit 384 bit
• Memory Clock
  1500 MHz -
• Memory Data Rate
  3 Gbps -
• Memory Size
  80 GB 24 GB
• Memory Type
  HBM3 GDDR6X
• Memory Speed (legacy frequency label)
  - 1300 MHz

Published GPU clock frequencies describe different operating modes. Clock speed alone does not determine graphics performance.

• Base Clock
  1065 MHz 1560 MHz
• Boost Clock
  1780 MHz 1860 MHz
• Overclocking
  - No

Graphics card power limits and recommended system power supply describe different specifications.

• Board Number
  PG520 -
• System Power Supply
  1100 W (PSU recommendation) 850 W (System power requirement)
• TDP
  700 W 450 W
• Graphics Card Power
  - 450 W
• Maximum GPU Temperature
  - 92 °C

Connectivity and connections.

• Max. Displays
  1 4
• PCIe Interface
  PCIe 5.0 x16 PCIe 4.0 x16
• Power Connectors
  8-pin EPS 3x PCIe 8-pin cables (adapter in box) OR 450W or greater PCIe Gen 5 cable
• DisplayPort Ports
  - 3 × DisplayPort 1.4a
• HDCP Version
  - 2.3
• HDMI Ports
  - 1 × HDMI 2.1
• Other Display Outputs (unparsed)
  - HDMI

Technical data that is used to its full potential in computer games.

• CUDA
  9.0 -
• DirectX
  No 12_2
• Max. resolution
  3840x2160@60Hz 7680 × 4320
• OpenCL
  3.0 3.0
• OpenGL
  No 4.6
• Shader Model
  No 6.6
• Vulkan
  No 1.3

Comparison of interfaces and release dates for GPU NVIDIA GeForce RTX 3090 Ti and NVIDIA H100 SXM5.

• Process Technology
  4 nm 8 nm
• Release date
  Q2/2022 -
• Manufacturer specification scope
  - NVIDIA Founders Edition/reference design. Clock specifications apply to gaming use. AIB shipping specifications and complete-system power needs may differ.

• FP16 (theoretical)
  120.3 TFLOPS 40 TFLOPS
• FP32 (theoretical)
  30.07 TFLOPS 40 TFLOPS
• FP64 (theoretical)
  15.04 TFLOPS 0.625 TFLOPS
• Pixel Rate
  42.72 GPixel/s 208.3 GPixel/s
• Texture Rate
  939.8 GTexel/s 625.0 GTexel/s