NVIDIAHopper

NVIDIA Hopper H200 / H100

The industry-standard AI accelerators for enterprise training and inference. H100 delivers proven reliability with 80GB HBM3 and the Transformer Engine, while H200 extends memory capacity to 141GB HBM3e for larger models. Both feature 4th-gen Tensor Cores and 900GB/s NVLink for multi-GPU scaling.

H100
Hopper
80GB
PREMIUM
H200
Hopper + HBM3e
141GB
80-141GB HBM3/HBM3e Memory
4.8TB/s Max Bandwidth
528 4th Gen Tensor Cores
900GB/s NVLink 4.0

Why Choose Hopper

Proven architecture trusted by thousands of enterprise deployments worldwide

3X
Faster vs A100
AI training throughput
FP8
Transformer Engine
Automatic mixed precision
7
MIG Instances
Multi-tenant deployment
Mature
Software Stack
Full CUDA ecosystem

Hopper Configurations

Choose the right Hopper GPU for your workload requirements

PCIe H100 PCIe
Memory 80GB HBM3
Bandwidth 2.0TB/s
CUDA Cores 14,592
TDP 350W
Enterprise Price $20,000 - $25,000
Air cooling · Standard servers
SXM5 H100 SXM5
Memory 80GB HBM3
Bandwidth 3.35TB/s
CUDA Cores 16,896
TDP 700W
Enterprise Price $20,000 - $25,000
Liquid cooling · NVLink

Technical Specifications

Complete Hopper H100 and H200 technical details

Specification H100 PCIe H100 SXM5 H200 SXM
Architecture
GPU Architecture NVIDIA Hopper (GH100)
Process Node TSMC 4N (Custom 4nm)
Transistors 80 Billion
Compute
Streaming Multiprocessors 114 132 132
CUDA Cores 14,592 16,896 16,896
Tensor Cores 456 (4th Gen) 528 (4th Gen) 528 (4th Gen)
Transformer Engine Yes (FP8 support)
Memory
Memory Capacity 80GB HBM3 80GB HBM3 141GB HBM3e
Memory Bandwidth 2.0TB/s 3.35TB/s 4.8TB/s
Memory Bus Width 5120-bit 6144-bit
L2 Cache 50MB
ECC Memory Yes
Performance
FP64 26 TFLOPS 34 TFLOPS 34 TFLOPS
FP32 51 TFLOPS 67 TFLOPS 67 TFLOPS
TF32 Tensor 756 TFLOPS 989 TFLOPS 989 TFLOPS
FP8 Tensor 3,026 TFLOPS 3,958 TFLOPS 3,958 TFLOPS
LLM Inference Baseline Baseline Up to 2X
Connectivity
NVLink N/A 18 links (900GB/s)
PCIe PCIe 5.0 x16
Form Factor Dual-slot FHFL SXM5 Module
Power & Thermal
TDP 350W 700W
Cooling Air or Liquid Liquid Required
Max Operating Temp 83°C
Enterprise Features
MIG Support Up to 7 instances
Confidential Computing Yes
vGPU Support Yes

Ideal Use Cases

Hopper powers the most demanding enterprise AI workloads

LLM Training

Train large language models with the Transformer Engine and NVLink scaling to thousands of GPUs. Mature software ecosystem ensures reliable training runs.

AI Inference

Deploy models at scale with MIG partitioning. H200's 141GB enables serving larger models; H100 offers best price/performance for smaller models.

Large Models (H200)

Run Llama 2 70B, Mixtral 8x7B, and other large models that exceed 80GB. H200's 141GB eliminates memory constraints for memory-bound workloads.

Scientific HPC

34 TFLOPS FP64 performance for computational physics, chemistry, and climate modeling. ECC memory ensures data integrity.

Generative AI

Power text, image, video, and multimodal generation. Transformer Engine optimizes throughput for transformer-based architectures.

Multi-Tenant (H100 PCIe)

H100 PCIe's 350W TDP and air cooling enables deployment in standard servers for flexible, cost-effective inference infrastructure.

Architecture Comparison

How Hopper compares in NVIDIA's data center GPU lineup

H200 vs H100

  • Memory: 141GB vs 80GB (+76%)
  • Bandwidth: 4.8TB/s vs 3.35TB/s (+43%)
  • LLM Inference: Up to 2X faster
  • Compute: Same (67 TFLOPS FP32)
  • Architecture: Same Hopper
  • Best For: Large models vs General AI
Verdict: H200 for memory-bound workloads; H100 for best price/performance

Hopper vs Blackwell

  • Memory: 80-141GB vs 192-288GB
  • Training: Baseline vs 4X faster
  • Inference: Baseline vs 30X faster
  • Availability: Available now vs Newer
  • Software: Mature vs Emerging
  • Price: $20K-$30K vs $30K-$60K
Verdict: Hopper for proven reliability; Blackwell for maximum performance

Frequently Asked Questions

What is the difference between H100 and H200?

H200 is the memory-upgraded H100 with 141GB HBM3e (76% more than H100's 80GB HBM3) and 4.8TB/s bandwidth (43% faster). Both share the same Hopper architecture and compute capabilities. H200 excels at memory-bound workloads like LLM inference with up to 2X performance improvement for large models.

What is the difference between H100 SXM5 and PCIe?

H100 SXM5 delivers maximum performance with 700W TDP, 3.35TB/s bandwidth, 16,896 CUDA cores, and 18 NVLink connections (900GB/s). H100 PCIe offers easier deployment with 350W TDP, 2.0TB/s bandwidth, 14,592 CUDA cores, and standard PCIe 5.0. SXM5 requires liquid cooling; PCIe works in standard air-cooled servers.

How much does H100 and H200 cost?

H100 PCIe: $20,000-$25,000. H100 SXM5: $20,000-$25,000. H200 SXM: $24,000-$30,000. Complete 8-GPU DGX systems range from $250,000-$350,000. Contact SLYD for current pricing through our OEM partnerships.

What is MIG (Multi-Instance GPU)?

Multi-Instance GPU allows a single H100/H200 to be partitioned into up to 7 isolated GPU instances, each with dedicated compute, memory, and bandwidth. This enables multi-tenant deployments where different users or workloads share a single physical GPU securely.

Should I choose Hopper or Blackwell?

Choose Hopper (H100/H200) for proven reliability, mature software ecosystem, immediate availability, and competitive pricing. Choose Blackwell (B200/B300) if you need maximum performance with 4X faster training and 30X faster inference. Hopper is the safe choice; Blackwell is the performance leader.

Is H200 available in PCIe form factor?

No, H200 is only available in SXM form factor which requires specialized baseboards and liquid cooling. For PCIe deployments in standard servers, H100 PCIe remains the best option. Consider Blackwell B200 PCIe variants when they become available.

Deploy NVIDIA Hopper

Get enterprise H100 and H200 systems through our OEM partnerships with expert configuration and deployment support.

Available through:
Dell HPE Supermicro Lenovo Gigabyte
Reconnecting to the server...

Please wait while we restore your connection

An unhandled error has occurred. Reload 🗙