Sharon AI Launches Cloud Media Tech with On-Demand GPU Service
NVIDIA H100
Tensor Core GPU
The NVIDIA H100 Tensor Core GPU, built on the NVIDIA Hopper™ architecture, delivers industry-leading performance, scalability, and security for AI and high-performance computing (HPC) workloads. With innovations such as the Transformer Engine and fourth-generation Tensor Cores, it accelerates large language models (LLMs), securely handles enterprise workloads, and supercharges AI training with up to 4X performance improvement over previous models.
Basic Product Information
Product Name
NVIDIA H100 Tensor Core GPU
Architecture
NVIDIA Hopper™
Memory
80GB HBM3 (SXM), 94GB HBM3 (NVL)
Compute Power
Up to 3,958 TFLOPS FP8 performance
Release Year
2022
Use Cases
AI training, large language models, high-performance computing (HPC), scientific computing
Key Advantages
Up to 4X Faster Training
Enhanced AI model training with FP8 precision.
Transformer Engine
60 TFLOPS FP64 Performance
Accelerates scientific and high-performance computing tasks.
Enterprise-Ready
Includes NVIDIA AI Enterprise for streamlined AI deployment.
7X Performance for HPC Applications
Ideal for tasks like genome sequencing and 3D FFT.
Specifications
Performance Specifications
LLM Inference Performance
Up to 5X over NVIDIA A100 systems for LLMs up to 70 billion parameters
FP64
30 teraFLOPS
FP64
Tensor Core: 60 teraFLOPS
FP32
60 teraFLOPS
TF32 Tensor Core
835 teraFLOPS (with sparsity)
BFLOAT16 Tensor Core
1,671 teraFLOPS (with sparsity)
FP16 Tensor Core
1,671 teraFLOPS (with sparsity)
FP8 Tensor Core
3,341 teraFLOPS (with sparsity)
INT8 Tensor Core
3,341 TOPS (with sparsity)
Decoders
7 NVDEC, 7 JPEG
Confidential Computing
Supported
Multi-Instance GPUs
Up to 7 MIGs @12GB each
Memory and Bandwidth
GPU Memory
80GB HBM3 (SXM) / 94GB HBM3 (NVL)
Memory Bandwidth
3.35TB/s (SXM) / 3.9TB/s (NVL)
Memory Clock
2619 MHz
Thermal and Power
Power
Configurable between 350-400W
Server Options
Compatible with partner and NVIDIA-Certified Systems that can accommodate 1 to 8 GPUs
Board Specifications
Form Factor
SXM or PCIe dual-slot
Interconnect
NVIDIA NVLink: Offers 600GB/s bidirectional bandwidth PCIe Gen5: Supports up to 128GB/s NVLink Bridge: Can be connected with another H100 NVL using 2- or 4-way NVLink bridges for increased bandwidth
Supported Technologies
Multi-Instance GPU (MIG)
Up to 7 MIGs per GPU (10GB each for SXM, 12GB each for NVL)
Confidential Computing
Provides hardware-based security for data in use.
AI Enterprise Software
NVIDIA AI Enterprise included for secure and scalable AI deployment.
Server Compatibility
NVL
Compatible with NVIDIA DGX H100 and NVIDIA-Certified Systems with 1-8 GPUs.
Additional Features
01Transformer Engine
Uses mixed FP8 and FP16 precision to dramatically accelerate training and inference for large AI models.