Skip to content
AQUATRON atom markAQUATRONDATA CENTER LLC
AQUATRON and Los Alamos AI benchmark project visualization
All projects
02TECHNICAL BENCHMARK

Los Alamos AI Benchmark

Advanced AI Compute Validation — NVIDIA H200 vs Intel Gaudi 3

Comparative AI-inference benchmark on Llama 3.1 405B, published with Los Alamos National Laboratory.

Location
United States
Date
29 May 2025
Status
Published
Classification
TECHNICAL BENCHMARK
Project Overview
LOCATION
United States
MILESTONE DATE
29 May 2025
CLASSIFICATION
TECHNICAL BENCHMARK
STATUS
Published
CLAIM STATUS
Technical Benchmark

Executive summary

Published on 29 May 2025, this benchmark was conducted by AQUATRON in partnership with Los Alamos National Laboratory. Brand-new Supermicro systems configured with 8 × NVIDIA H200 accelerators were compared against 8 × Intel Gaudi 3 accelerators, running 16-bit precision Llama 3.1 405B across three scenarios: AI research, history and coding.

AQUATRON role

Benchmark partner, working with Los Alamos National Laboratory on the comparative AI-inference validation and its publication.

Participants

01

Los Alamos National Laboratory

Research partner

  • Benchmark partnership
  • Research scenario context

02

AQUATRON

Benchmark partner

  • System configuration
  • Benchmark execution and publication

03

Supermicro systems

System platform

  • Brand-new servers hosting both accelerator configurations

Technology

The stack behind the work.

  • 018 × NVIDIA H200
  • 028 × Intel Gaudi 3
  • 03Supermicro systems
  • 04Llama 3.1 405B · 16-bit precision
  • 05AI research scenario
  • 06History scenario
  • 07Coding scenario

Published metrics

NVIDIA H200 throughput
≈ 25–26 tokens/s
PUBLISHED BENCHMARK
Source: AQUATRON × LANL benchmark publication
Intel Gaudi 3 throughput
≈ 2.7 tokens/s
PUBLISHED BENCHMARK
Source: AQUATRON × LANL benchmark publication
Reported H200 advantage
≈ 9–9.5×
PUBLISHED BENCHMARK
Source: AQUATRON × LANL benchmark publication
In tested scenarios only.

Benchmark visualization

Llama 3.1 405B · 16-bit precision · tokens per second

≈ 9–9.5×

Reported H200 advantage in tested scenarios

NVIDIA H200 8 × H200 · Supermicro

≈ 25–26 tok/s

Intel Gaudi 3 8 × Gaudi 3 · Supermicro

≈ 2.7 tok/s

AI research scenario

History scenario

Coding scenario

Accuracy note. Benchmark results apply to the reported test configuration and workloads and should not automatically be generalized into universal power, TCO, cost or ROI claims.

Timeline

  1. 29 May 2025

    Benchmark publication

    Results published for H200 vs Gaudi 3 on Llama 3.1 405B.

Evidence classification

Publicly Documented

Technical Benchmark

Benchmark results apply to the reported test configuration and workloads and should not automatically be generalized into universal power, TCO, cost or ROI claims.

SOURCE
AQUATRON × Los Alamos National Laboratory benchmark publication, 29 May 2025

Related projects

07PRODUCT PLATFORM

AQUATRON POD

Modular · Deployable globally · Product platform

Company-described modular, plug-and-play AI infrastructure platform with integrated power-to-compute architecture.

Company ReportedExplore AQUATRON POD

Collaborate

Discuss this project with AQUATRON.

Infrastructure collaboration, technology partnership and sovereign-AI discussions are welcome.