Uponor technical article hero

Technical article

From Raspberry Pi to RK3588J: How I Sourced a More Powerful Alternative for Our Edge AI Project

by Akiko Fujimori

How It Started

Last October, our engineering lead sent me a Slack message that said, verbatim: "We need a faster Raspberry Pi for the new edge AI thing. Can you source 20 units?"

I've been the office administrator at a 140-person software company for six years now. I manage all hardware and office supply ordering — roughly $380,000 annually across 14 vendors. I've negotiated Dell rates, compared SSD specs, and dealt with more invoice disputes than I care to count. But "faster Raspberry Pi" wasn't exactly a SKU I recognized.

I figured it'd be a 30-minute task. It took six weeks.

What follows is what I learned while hunting for a more powerful Raspberry Pi alternative — including stops at RK3588J modules, compute module 4 carrier board options, and a computer-on-module Atom kit that nearly won the bid. If you're a procurement person who suddenly got handed a hardware sourcing project outside your comfort zone, this is for you.

The Real Requirements

After a bit of back-and-forth with engineering, "faster" turned out to mean: "can run OpenAI embedding workloads locally without catching fire."

We were building an internal prototype that used OpenAI's embedding models to auto-classify support tickets. The Pi 4 we already had handled the networking side fine — but once an embedding model started crunching, everything slowed to a crawl. The prototype was basically unusable.

So my actual filter was:

  • More CPU and NPU throughput than a Pi 4
  • Decent thermals (our enclosures are small and sealed)
  • 12+ month availability commitment from the vendor
  • Not obscenely expensive

That still left dozens of options. I didn't know it yet, but "faster Raspberry Pi alternatives" covers everything from a $35 Pi 5 to $600 industrial boards. My ignorance was going to cost me time.

The RK3588J Rabbit Hole

The first board that kept coming up in every "Raspberry Pi replacement" thread was the RK3588J. I'd never heard of Rockchip before this project — I'm a procurement person, not an electrical engineer — but the spec sheet was genuinely impressive: 8 cores, 6 TOPS NPU, support for up to 32GB RAM.

It wasn't just faster. It was a different category of hardware entirely.

The problem was availability. One vendor quoted 16 weeks. Another said "TBD" and stopped responding to my emails. A third tried to sell me a dev board instead of the production units we actually needed. I started to feel like I was chasing a unicorn.

The Alternatives We Considered

While I was chasing RK3588J availability, our engineering team raised a second option: going the compute module 4 carrier board route. The appeal was understandable — we'd stay in the Pi ecosystem we already knew, just swap in a CM4 and a custom carrier board.

But this meant designing a board ourselves or paying someone to do it. For a 20-unit pilot, the NRE costs didn't pencil out. We'd be spending thousands just to save a few hundred on per-unit hardware.

Then there was the computer-on-module Atom option. Intel's low-power line doesn't get as much attention as the ARM crowd, but for our workload, it was genuinely competitive. The sticking point was power draw and heat. The Atom COM ran hotter than the RK3588J in our thermal chamber. For a sealed enclosure, that was a deal-breaker.

Here's what you need to know: every option had a trade-off. There was no free lunch.

The Moment I Almost Got It Wrong

I had been comparing spec sheets for about three weeks straight, and I was ready to recommend the RK3588J based on paper specs alone. My engineer pushed back. "Have you tested sustained performance?" he asked.

I had not. I'd been comparing peak numbers like an idiot.

Looking back, I should have ordered dev boards in week one instead of week four. At the time, I thought I was being efficient by narrowing down the list first. In reality, I was delaying the only test that actually mattered.

So we ordered three dev kits — one RK3588J, one CM4 carrier board setup, and one Atom COM eval kit. Set them up on a bench in our lab. Ran the actual embedding workload for 72 hours straight.

The RK3588J crushed the first 6 hours. Then it started throttling. The dev board's cooling was inadequate for sustained loads. We swapped in a better heatsink, and throttling dropped by about 80%. Not gone, but manageable.

The CM4 route held steady but slower. The Atom COM ran cool but lagged on throughput.

After 72 hours, the RK3588J with upgraded cooling was the clear winner for our specific workload.

What We Ended Up Doing

We ordered 18 production units of the RK3588J and installed them at our first deployment location. Five months in, no failures.

But I'm not going to pretend this was a smooth process. The lead time on those 18 units was still longer than I'd like. We ended up paying a premium to a distributor who actually had stock, because the direct vendor timeline was sliding again.

If you're a procurement person looking at faster Raspberry Pi alternatives, here's my advice:

  • Define "faster" before you shop. Faster at what? Under what constraints? The RK3588J wasn't universally "best" — it was best for our embedding workload and sealed enclosure. If we'd needed fanless operation, the Atom COM might have won.
  • Test with your actual workload. Spec sheets and benchmarks lie. Only your code on your hardware in your enclosure tells the truth.
  • Budget more time than you think. I planned for two weeks. It took six. Availability in this space is genuinely unpredictable.
  • Don't sleep on the boring option. The compute module 4 carrier board route felt "safe" because it kept us in the Pi ecosystem. But safe isn't cheap when NRE costs are involved for small volumes.

This worked for us, but our situation was specific — small pilot volume, sealed enclosures, and a workload that's NPU-heavy but not memory-bound. Your mileage may vary if you're dealing with higher volumes or different thermal constraints.

I can only speak to what we evaluated. There are probably other boards and vendors out there I never got to. If you're working with a bigger budget or a longer timeline, the calculus might be different.

Bottom line: don't buy on specs alone. Bring your workload to the evaluation. And start testing earlier than you think you need to.

Share LinkedIn X
Akiko Fujimori

Akiko Fujimori

Akiko Fujimori is a structural building-materials analyst covering sawn timber, structural composite lumber, plywood, OSB, masonry units, concrete products, reinforcing components, and metal building systems. She references ASTM D5456 and ASTM C90 within their material scopes while comparing bending strength, modulus of elasticity, compressive strength, density, moisture content, dimensional tolerance, connector capacity, and corrosion protection. Her evaluations help structural designers, fabricators, and procurement teams align declared properties, connection details, exposure conditions, and inspection evidence with project loads.

Previous EPS Sandwich Wall Panels vs. Aluminum Honeycomb Core Panels for Pharmaceutical Cleanrooms: A Quality Inspector's Comparison Next Price-First vs. Spec-Support: What to Look for in a Commercial Plumbing Supplier for Uponor PEX Projects