NVIDIA H100 SXM5 80GB – Hopper Tensor Core Data Center GPU
NVIDIA H100 SXM5 80GB – Hopper Tensor Core Data Center GPU - Image 2
NVIDIA H100 SXM5 80GB – Hopper Tensor Core Data Center GPU - Image 3
NVIDIA H100 SXM5 80GB – Hopper Tensor Core Data Center GPU - Image 4
Home / Memory / NVIDIA H100 SXM5 80GB – Hopper Tensor Core Data Center GPU

NVIDIA H100 SXM5 80GB – Hopper Tensor Core Data Center GPU

The NVIDIA H100 SXM5 80GB is a flagship Hopper-generation data-center GPU accelerator designed for large-scale artificial intelligence, machine learning, generative AI, high-performance computing (HPC), and scientific workloads. It combines 80GB HBM3 memory, 3.35 TB/s memory bandwidth, 16,896 CUDA cores, fourth-generation Tensor Cores, and 900 GB/s NVIDIA NVLink connectivity for highly scalable multi-GPU computing.

Category:

Description

The NVIDIA H100 SXM5 80GB is a high-performance GPU accelerator based on NVIDIA’s Hopper architecture and designed specifically for data-center and accelerated-computing environments.

Unlike the H100 PCIe 80GB, which uses HBM2e and a PCIe card interface, the H100 SXM5 uses 80GB of HBM3 memory and an SXM5 module form factor. NVIDIA specifies memory bandwidth of approximately 3.35 TB/s, providing extremely high data throughput for AI models, large datasets, simulations, and other memory-intensive workloads.

The SXM5 GPU incorporates 132 Streaming Multiprocessors (SMs) and 16,896 FP32 CUDA cores, together with 528 fourth-generation Tensor Cores. These resources are optimized for deep-learning and HPC workloads across FP64, FP32, TF32, FP16, BF16, FP8, and INT8 precision formats.

For AI workloads, NVIDIA specifies up to 989 TFLOPS TF32 Tensor Core performance, 1,979 TFLOPS BF16/FP16, and 3,958 TFLOPS FP8/INT8 performance with sparsity. The architecture also incorporates NVIDIA’s Transformer Engine to accelerate modern transformer-based AI models.

One of the major advantages of the SXM5 platform is its high-speed GPU-to-GPU connectivity. The H100 SXM5 provides 900 GB/s NVIDIA NVLink bandwidth, enabling efficient communication between GPUs in HGX and DGX systems and helping scale large AI and HPC workloads across multiple accelerators.

The H100 SXM5 supports Multi-Instance GPU (MIG) technology, allowing the accelerator to be partitioned into multiple isolated GPU instances. NVIDIA documents configurations of up to seven MIG instances, including a full 80GB GPU instance.

The H100 SXM5 is a server-integrated accelerator module, not a conventional PCIe graphics card. It is intended for purpose-built systems such as NVIDIA HGX H100 platforms and NVIDIA DGX H100 systems and requires the appropriate SXM5 carrier, power delivery, cooling, and system architecture.

3. Complete Technical Specification

Specification Details
Manufacturer NVIDIA
Product H100 Tensor Core GPU
Variant H100 SXM5 80GB
Architecture NVIDIA Hopper
GPU Architecture GH100
Form Factor SXM5
GPU Memory 80GB
Memory Type HBM3
Memory Stacks 5 HBM3 stacks
Memory Interface 5,120-bit
Memory Bandwidth 3.35 TB/s
Streaming Multiprocessors 132 SMs
CUDA Cores 16,896 FP32 CUDA cores
Tensor Cores 528 fourth-generation Tensor Cores
FP64 Performance Up to 34 TFLOPS
FP64 Tensor Core Up to 67 TFLOPS
FP32 Performance Up to 67 TFLOPS
TF32 Tensor Core Up to 989 TFLOPS*
BF16 Tensor Core Up to 1,979 TFLOPS*
FP16 Tensor Core Up to 1,979 TFLOPS*
FP8 Tensor Core Up to 3,958 TFLOPS*
INT8 Tensor Core Up to 3,958 TOPS*
L2 Cache 50MB
GPU Interconnect NVIDIA NVLink
NVLink Bandwidth Up to 900 GB/s
Host/System Interface System-dependent SXM5 platform
TDP Up to 700W, configurable
MIG Supported
Maximum MIG Instances Up to 7
MIG Full-GPU Profile 7g.80gb
Video Decoders 7 NVDEC
JPEG Decoders 7
Primary Deployment Data center / accelerated computing
Supported Systems NVIDIA HGX H100, NVIDIA DGX H100 and qualified partner platforms
Primary Workloads AI, ML, generative AI, HPC, scientific computing, analytics

* NVIDIA’s published Tensor Core figures marked with an asterisk are measured with sparsity.

4. Applications

The H100 SXM5 80GB is designed for demanding accelerated-computing workloads, including:

  • Artificial intelligence and machine learning
  • Large Language Model (LLM) training
  • Generative AI
  • Large-scale AI inference
  • Deep learning
  • Transformer-based model training
  • Natural language processing
  • Computer vision
  • High-performance computing (HPC)
  • Scientific and engineering simulation
  • Computational fluid dynamics
  • Molecular and pharmaceutical research
  • Seismic and geophysical analysis
  • Financial modeling
  • Large-scale data analytics
  • GPU-accelerated databases
  • Digital twins
  • Multi-GPU AI clusters
  • Enterprise AI infrastructure

5. Compatibility

The H100 SXM5 80GB is intended for systems specifically designed to accommodate NVIDIA SXM5 GPU modules.

NVIDIA identifies the H100 SXM platform for:

  • NVIDIA HGX H100 systems
  • NVIDIA DGX H100 systems
  • NVIDIA-Certified partner systems designed around H100 SXM GPUs

NVIDIA’s current product specifications identify HGX H100 configurations with 4 or 8 GPUs, while the DGX H100 platform uses 8 H100 GPUs.

Important Compatibility Note

The H100 SXM5 cannot be installed into a standard PCIe GPU slot. It is fundamentally different from the H100 PCIe 80GB.

A compatible host system must provide:

  • SXM5 GPU carrier/interface
  • Appropriate 700W-class GPU power delivery
  • Dedicated high-performance GPU cooling
  • Compatible system firmware
  • H100-supported NVIDIA drivers/software stack
  • Appropriate NVLink infrastructure where required
  • Qualified motherboard/baseboard architecture

The H100 SXM5 should therefore be sold as a server-integrated accelerator module, rather than as a conventional PCIe expansion card.

6. Key Benefits

  • 80GB HBM3 high-bandwidth GPU memory
  • Exceptional 3.35 TB/s memory bandwidth
  • 16,896 CUDA cores for massively parallel computing
  • 528 fourth-generation Tensor Cores
  • Hopper architecture optimized for AI and HPC
  • Up to 3,958 TFLOPS FP8 with sparsity
  • Up to 3,958 TOPS INT8 with sparsity
  • 900 GB/s NVLink GPU-to-GPU interconnect
  • Up to 700W configurable TDP for maximum compute performance
  • Support for up to seven MIG instances
  • Designed for large-scale multi-GPU systems
  • Optimized for LLMs and generative AI
  • Excellent memory bandwidth for large AI models
  • Purpose-built for enterprise data centers and HPC environments

7. SEO Title

NVIDIA H100 SXM5 80GB HBM3 Tensor Core GPU – 3.35TB/s AI & HPC Accelerator

Related products