Skip to main content
scalenta

Hopper architecture

NVIDIA H100

NVIDIA H100 is the workhorse Hopper accelerator: fourth-generation Tensor Cores, a Transformer Engine with FP8, NVLink at 900 GB/s and confidential computing — for LLM training, inference and HPC.

Request GPU ConsultationView specifications
GPU memory
80 GB HBM3
Memory bandwidth
3.35 TB/s
FP8 tensor
3,958 TFLOPS

Specifications as published by NVIDIA; Tensor Core ratings shown are sparsity-accelerated figures.

H100HOPPER ARCHITECTURE

Overview

Why H100 fits

Hopper architecture for LLMs and AI

H100 delivers the architecture most modern AI stacks were built and optimised around: it accelerated the jump to FP8 training and inference, and remains the reference platform for LLMs once memory requirements are understood. With 80 GB of HBM3 and 3.35 TB/s of bandwidth per SXM GPU, it handles training and high-throughput serving for a wide range of models.

We position H100 as the default starting point for serious GPU workloads unless the model's context or memory footprint argues for H200, or the workload is graphics/video-leaning and better served by Ada-class parts. Capacity is sourced through verified partners and every instance shape is confirmed per engagement.

Specifications

H100 technical specifications

Published figures for the accelerator, followed by the variant and configuration caveats that apply.

Architecture
NVIDIA Hopper
GPU memory
80 GB HBM3
Memory bandwidth
3.35 TB/s
FP8 Tensor Core
3,958 TFLOPS (with sparsity)
FP16 / BF16 Tensor Core
1,979 TFLOPS (with sparsity)
TF32 Tensor Core
989 TFLOPS (with sparsity)
FP64
34 TFLOPS
Interconnect
NVLink 900 GB/s · PCIe Gen5 128 GB/s
Form factor
SXM
Max TDP
Up to 700 W (configurable)
Multi-Instance GPU
Up to 7 partitions
Notable features
Transformer Engine, confidential computing

Variant & configuration note

Rows describe the H100 SXM form factor as published by NVIDIA. The PCIe-based H100 NVL variant differs (94 GB HBM3, 3.9 TB/s, NVLink bridge, 350–400 W). Host configuration is scoped per engagement.

Specifications describe the GPU itself as published by NVIDIA — not a fixed server configuration. Scalenta scopes complete instances (GPU count, interconnect, CPU, memory, storage) per engagement and states them in writing.

Use cases

Where H100 fits

Workload shapes this platform suits — the scoping conversation maps them to a concrete configuration.

  • LLM training and fine-tuning

    Foundation-model training and parameter-efficient fine-tuning with FP8 through the Transformer Engine, scaled across NVLink-connected GPUs.

  • High-throughput inference

    Serving chat, generation and RAG workloads where tokens per second and tail latency drive the architecture.

  • HPC and analytics

    Double-precision simulation, genomics and large-scale data analytics accelerated with CUDA and NVIDIA libraries.

Delivery model

Partner-sourced, scoped in writing

How GPU capacity is delivered at Scalenta — the same three commitments on every model page.

  • Partner-sourced capacity

    Accelerator capacity is sourced through verified data center and hardware partners — no facility-ownership claims, no stock promises.

  • Scoped per engagement

    Model, variant, GPU count, interconnect and host configuration are confirmed in writing before anything is committed.

  • Operated with 24/7 support

    Deployments are covered by 24/7 support and SLA-backed managed services scoped to your environment.

Related GPU models

Nearby options in the catalogue

H200HOPPER ARCHITECTUREHopper

NVIDIA H200

Hopper-class acceleration with the largest HBM3e buffer in the catalogue, for AI workloads that are constrained by memory.

GPU memory
141 GB HBM3e
Memory bandwidth
4.8 TB/s
FP8 tensor
3,958 TFLOPS
View specifications
A100AMPERE ARCHITECTUREAmpere

NVIDIA A100

A proven training-and-HPC accelerator with an 80 GB memory pool per GPU and mature tooling across the AI ecosystem.

GPU memory
80 GB HBM2e
Memory bandwidth
2,039 GB/s
FP16 tensor
624 TFLOPS
View specifications
L40SL40SADA LOVELACE ARCHITECTUREAda Lovelace

NVIDIA L40S

One card for generative AI, 3D graphics, rendering and video — the pragmatic choice where workloads refuse to stay in one lane.

GPU memory
48 GB GDDR6 ECC
Memory bandwidth
864 GB/s
FP8 tensor
1,466 TFLOPS
View specifications
Browse all GPU instances

FAQ

H100 — common questions

How many H100 GPUs does a typical project need?
It depends on model size, precision, batch profile and how much time you want the run to take — there is no honest one-number answer. We start from the workload: model, dataset, target latency or training window, then propose the smallest topology that meets it with documented headroom.
Can Scalenta supply H100 capacity in a specific region?
We work with partner data center and hardware providers and will confirm which regions can serve your requirement during scoping. Data handling, residency and latency requirements are reviewed before a region is committed in writing.
Is NVLink important for my workload?
For workloads that split a model or dataset across GPUs, yes — NVLink keeps communication off the PCIe path. For single-GPU inference it is less critical. The proposal will state whether the topology we recommend uses NVLink and why.

Request a GPU consultation

Talk to us about NVIDIA H100 — workload-first scoping, written findings and a proposal with clear scope. No obligation.

Request Infrastructure AssessmentTalk to our team