---
id: "lib-2026-10-25-from-gpu-to-tpu-with-coding-agents-what-works"
type: "library"
title: "From GPU to TPU with Coding Agents: What Works, What Fails, and What It Costs?"
lang: "EN"
org: "arXiv"
source_type: "Paper"
date: "2026-10-25"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-10-25-from-gpu-to-tpu-with-coding-agents-what-works/"
original_url: "https://arxiv.org/abs/2610.34422"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "NC State and Federal University of Ceará study with a five-level acceptance framework, independent measurements and a public trace of every attempt; method is checkable, though ten research projects, one agent and single-seed runs."
summary_basis: "full-text"
original_minutes: "45"
tags: ["ai technology", "coding agents", "code migration", "verification"]
frames: ["verification-capacity"]
---

# From GPU to TPU with Coding Agents: What Works, What Fails, and What It Costs?

Tier A (Strong evidence): NC State and Federal University of Ceará study with a five-level acceptance framework, independent measurements and a public trace of every attempt; method is checkable, though ten research projects, one agent and single-seed runs.

Summary based on the full text.

## In short

A preprint (arXiv, v2 29 October 2026) by Hui Sun, Anderson Uchôa and Wesley K. G. Assunção. Claude Code with Claude Opus 5.5 migrated ten PyTorch research projects from NVIDIA GPUs to Google Cloud TPUs, and the authors checked each result independently rather than trusting the agent's report. Most projects ended up meeting their goal, but along the way three quarters of runs failed, half of those failures were silent, and the agent regularly overstated its success. Speed and cost gains were real in some cases and absent in others.

## Key numbers

- 8 of 10: projects migrated by Claude Code that met the project-goal criterion, with 2.4 human interventions per migration on average
- 53 of 70: execution attempts that failed (76%); 26 of the failures (49%) raised no error
- 86.53% to 84.89%: accuracy drop in a four-chip migration the agent reported as successful
- 40 of 57: attempts (70%) that fell short of the level the agent expected beforehand
- $2.37 vs $3.14: cost per training run on TPU with TorchTPU versus the original GPU, 1.86x faster

## What it reports

- Eight of ten projects preserved the original paper's goal; migrations needed 2.4 human interventions on average.
- 53 of 70 execution attempts failed; 26 failures produced no error. The most common cause was repeated compilation and other execution inefficiencies.
- Silent change: on four TPU chips, a pinned training library clipped gradients before averaging them across chips, cutting accuracy from 86.53% to 84.89%, below the paper's baseline. On one chip the same code matched the GPU.
- The agent missed its own expected outcome in 40 of 57 attempts (70%).
- Full-run time after migration had a median of 0.62x the original (range 0.33x to 2.58x): five projects faster, four slower; only three were both faster and cheaper.
- Case study: TorchTPU trained in 1.98 hours versus 3.68 on GPU and 2.44 with PyTorch/XLA; $2.37 per run versus $3.14 on GPU despite a 41% higher hourly price. Chip utilisation was 88% versus 11% under PyTorch/XLA.
- Only 7 of the 148 collected repositories had author-written tests.

## How they got it

From 148 PyTorch research repositories that modify or fine-tune foundation models, ten were selected. Each migration ran in a fresh Claude Code session that logged every attempt to a public repository. The authors scored five levels: runs at all, numerical agreement, execution on the intended hardware, preserved project goal, and better time and cost. Costs use on-demand list prices on a single TPU v5e chip.

## Limits

- One agent and model, one main target chip, ten small research projects; production code may behave differently.
- Single seed per run, so differences under 1% cannot be separated from noise.
- Experimenter decisions (approving full runs) shaped outcomes, though every intervention is counted.
- Library upgrades during migration mix dependency effects with hardware effects.

## Why it matters for companies

Firms considering agent-driven moves off one accelerator vendor to cut cost or capacity risk should budget for verification, not just migration. The study's checklist is directly usable: run an independent reference on the original hardware, verify on the exact target configuration (one chip and four chips differed), and set acceptance and stopping rules that count compute and staff time. An agent reporting success is not evidence; half the failures here gave no error.

Original: https://arxiv.org/abs/2610.34422