Skip to content
#explainable ai Open access

Approaching the Memory-Bandwidth Ceiling: Three As-Deployed LLM Inference Stacks on a 128 GB Unified-Memory APU

Oct 2026 · Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques

Abstract

Single-node measurement note. Two-day campaign measuring three as-deployed LLM inference stacks (model + runtime + quantization: Qwen3.8-Flash-Next on halogen, GLM-5.3 on ExLlamaV3, DeepSeek-V4-Flash on ember) on one 128 GB unified-memory AMD Ryzen AI Max+ 395 node, under one measurement contract and one frozen corpus. Measured memory-read ceiling: 240.3 GB/s; the three stacks decode at 85-90 % of it, so the differences are primarily explained by effective bytes per token rather than by under-use of the hardware. Includes the paired per-item bootstrap of the quality differences (95 % CIs exclude zero for all three pairs, within this corpus) and four measured negative results (speculative decoding: no gain; --num-draft R=4: VRAM OOM; factory sampler: ~5 %, not the 13-21 % quoted in forums; aggregate concurrency: ~0 %). Limits are declared: one node, no energy/cost measurement, n=3 in short cells, quantization differs between stacks (declared confounder), and ~10-15 % of the ceiling is unexplained and NOT attributed to any cause. Personal work; not peer reviewed. Resumen en castellano. Nota técnica de una campaña de dos días en un solo nodo con memoria unificada de 128 GB (AMD Ryzen AI Max+ 395). Se midió el techo real de lectura de memoria: 240,3 GB/s, frente a los 256 GB/s teóricos de ficha y los ~215 GB/s que reporta la comunidad. Los tres stacks de inferencia evaluados, tal y como estaban desplegados (modelo + motor + cuantización), trabajan entre el 85 % y el 90 % de ese techo: las diferencias se explican sobre todo por los bytes efectivos por token, no por desaprovechamiento del hardware. En decode, la velocidad tiene como límite superior el ancho de banda dividido entre los bytes leídos por token; en un MoE cuentan los parámetros activos, no los totales. Incluye el bootstrap pareado de las diferencias de calidad (los intervalos del 95 % excluyen el cero en los tres pares) y cuatro resultados negativos medidos: la decodificación especulativa no aporta ganancia, el sampler de fábrica cuesta ~5 % (no el 13-21 % que circula) y la concurrencia agregada es ~0 %. Límites declarados: un nodo, un corpus, sin medir energía ni coste, y con cuantización distinta entre stacks. El texto canónico del informe es el inglés.

View source

Similar papers

#artificial intelligence Conference Open access Apr 2020

ECCOLA - a Method for Implementing Ethically Aligned AI Systems

The method, ECCOLA, is presented, which aims at making the high-level AI ethics principles more practical, making it possible for developers to more easily implement them in practice.

Ville Vakkuri, Kai-Kristian Kemell, P. Abrahamsson · 64 citations · ⚡6
#computer vision Review Apr 2024

AI-powered Code Review with LLMs: Early Results

The goal is to not only refine the accuracy of the LLM-based tool but also to underscore its potential in streamlining the software development lifecycle through proactive code improvement and education.

Z. Rasheed, Malik Abdul Sami, Muhammad Waseem et al. · 62 citations · ⚡3
#computer vision Open access Mar 2024

LLM-based agents for automating the enhancement of user story quality: An early report

The use of large language models to automatically improve the user story quality in Austrian Post Group IT agile teams is explored, with a reference model for an Autonomous LLM-based Agent System developed and implemented at the company.

Zheying Zhang, M. Rayhan, Tomas Herda et al. · 48 citations · ⚡4
#computer vision Review Mar 2024

System for systematic literature review using multiple AI agents: Concept and an empirical evaluation

This paper introduces a novel multi-AI-agent system designed to fully automate SLRs, and demonstrates how it substantially reduces the time and effort traditionally required for SLRs while maintaining comprehensiveness and precision.

Abdul Malik Sami, Z. Rasheed, Kai-Kristian Kemell et al. · 44 citations · ⚡2
#computer vision Feb 2024

Can Large Language Models Serve as Data Analysts? A Multi-Agent Assisted Approach for Qualitative Data Analysis

The proposed LLM-based multi-agent system automates qualitative data analysis process, creating opportunities for researchers and practitioners, and future improvements focus on enhancing multilingual performance and integrating continuous expert feedback.

Z. Rasheed, Muhammad Waseem, Aakash Ahmad et al. · 41 citations
#artificial intelligence Conference Open access Jun 2018

The Key Concepts of Ethics of Artificial Intelligence

It is suggested that the focus on finding keywords is the first step in guiding and providing direction for future research in the AI ethics field.

Ville Vakkuri, P. Abrahamsson · 39 citations · ⚡2

Related blog posts

MIT News · Artificial Intelligence Oct 15, 2026

AI model decodes the language cells use to communicate

A new computational approach reveals how cells communicate to assemble into tissues. This could accelerate stem cell engineering, improve disease models, and inform therapies.

GPT-Lab Oct 14, 2026

Building trustworthy AI: The role of AI harnesses, intent contracts and evals

AI is moving beyond impressive demos. After World Summit AI 2026 in Amsterdam, GPT-Lab's Muhammad Waseem reflects on what organizations need to make AI agents reliable, responsible and genuinely useful in everyday work. The post Building trustworthy AI: The role of AI harnesses, intent contracts and evals appeared first on GPT-Lab.

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.