Skip to main content

NVIDIA H100 NVL

Overview

NVIDIA H100 NVL is the LLM-optimized variant of the H100, featuring a dual-die package (MCM-like), delivering 94GB HBM3 (two 48GB stacks) or 188GB HBM3e per card. It is purpose-built for ultra-large LLM inference (GPT-4-class 1.8T models) and can load such models on a single card.

The PCIe Gen5 interface (vs. the original H100 SXM's NVLink) enables easy deployment in PCIe servers.

Core Specifications

ItemSpec
ArchitectureHopper (GH100, dual-die)
ProcessTSMC 4N
CUDA Cores14,592 × 2 = 29,184 (dual-die)
Tensor Cores456 × 2 = 912 (4th Gen)
Base Clock1.08 GHz
Boost Clock1.78 GHz
Memory94 GB HBM3 (two 48GB stacks) / 188 GB HBM3e
Memory Bandwidth3.35 TB/s (HBM3) / 7.0 TB/s (HBM3e)
FP16 Tensor1979 TFLOPS (sparse)
FP8 Tensor3958 TFLOPS (sparse)
TDP350-400 W (PCIe) / 700 W (SXM-NVL)
InterfacePCIe Gen5 ×16
Launch2024-Q1
Price$30,000-$40,000 (OEM)

Comparison with H100 SXM

MetricH100 SXM 80GBH100 NVL 94GBDifference
Memory Capacity80GB HBM394GB HBM3+17.5%
Memory Bandwidth3.35 TB/s3.35 TB/sSame
NVLink900 GB/s600 GB/s (PCIe)-33%
Interconnect BandwidthNVLink 18-linkPCIe Gen5Flexible
TDP700W350-400W-43%
Use CaseTraining/HPCLLM inferenceOptimized

LLM Inference Scenarios

Single-Card Loading for Ultra-Large Models

  • Llama 3 70B FP16 (~140GB) → Requires 2 cards
  • Llama 3 70B INT4 (~40GB)Single card sufficient
  • Mixtral 8x22B INT4 (~80GB) → Single card can load
  • Llama 3 405B INT4 (~230GB) → 3 NVL cards

Multi-Card NVL Bridge

  • 2× H100 NVL connected via NVLink Bridge → 188GB unified memory
  • 2× H100 NVL for 405B model inference (quantized ~230GB)

Real-World Deployments

  • NVIDIA AI Enterprise certified servers (Dell, HPE, Supermicro)
  • CoreWeave / Lambda Labs cloud NVL instances
  • Hugging Face Inference Endpoints select H100 instances

Vendor Information

ItemDetail
VendorNVIDIA
Product Pagehttps://www.nvidia.com/en-us/data-center/h100/
OEM Price$30,000-$40,000 (OEM-dependent)
Target MarketUltra-large LLM inference, enterprise GenAI

Use Cases

  • Ultra-large LLM inference (70B+ single card)
  • ✅ Long-context LLM (94GB fits 200K context)
  • ✅ PCIe server deployment
  • ✅ Dual-card NVL Bridge for 405B models
  • ❌ Training (H100 SXM is better suited, higher NVLink bandwidth)
  • ❌ HPC (use H100 SXM/H200)