AI & GPU Accelerators

France's First Exascale Supercomputer with AMD

€544 million for a supercomputer that might sip just 12 megawatts — or not. France's Alice Recoque promises exascale glory with AMD at the helm, but don't hold your breath for revolution.

Artist rendering of Alice Recoque exascale supercomputer racks with liquid cooling in GENCI Paris facility

Key Takeaways

  • France invests €544M in Alice Recoque, Europe's second exascale supercomputer powered primarily by AMD.
  • Power efficiency claims of 12 MW walked back by Eviden as 'preliminary' — expect higher real-world draw.
  • Modular design mixes AMD accelerators with SiPearl CPUs, targeting AI and traditional HPC, but timelines slip to 2026-2027.

€544 million. That’s what France is shelling out for its first exascale supercomputer.

And it’s not even built yet.

Meet Alice Recoque — named after some forgotten French computer pioneer, because why not slap a feel-good label on a rack of overheating chips? Eviden, Atos’s HPC arm, is in charge, with AMD supplying the Epyc CPUs and those shiny Instinct MI430X GPUs. Europe’s second exascale beast after Germany’s Jupiter, supposedly launching late 2026. Supposedly.

Look, Europe’s been playing catch-up in the supercomputing game forever. US has Frontier, China has its secrets, and now France wants in. But here’s the kicker: this modular monster at GENCI in Paris packs 94 racks of BullSequana XH3500, liquid-cooled because fans? So 2020. AMD FPGAs, BXI networking screaming at 800 Gbps — on paper, it sounds fierce.

Why Bet Big on AMD Venice Now?

AMD’s “Venice” chips haven’t shipped. Not one. Expected next year, they say. Optimistic? Sure. Delusional? Maybe.

France isn’t all-in on AMD, though. First partition: accelerators with those MI430X GPUs. Second: SiPearl’s Rhea2 CPUs, arriving 2027 — if they ever ship production silicon. SiPearl’s first Rhea1? Still vaporware for Jupiter’s expansion. Europe’s homegrown ARM dreams keep slipping, don’t they?

Eviden calls it an “AI Factory.” Cute. Nvidia’s version chews exaflops for training behemoths like GPT. Alice? More like a hybrid workhorse — simulations, data crunching, some AI on the side. Traditional HPC with an AI sticker slapped on. Because who isn’t doing that these days?

“This figure was part of a preliminary version and is not an exact science, rather an expectation that may vary depending on the production capacity.”

Eviden’s spokesperson, scrambling after we called out their 12-megawatt power claim. Yeah, they backpedaled faster than a politician at a scandal. “Not an exact science,” they say. Translation: We hyped it, now reality bites.

Twelve megawatts for exascale? Jupiter guzzles more like 15-20 under load. If Alice hits that — big if — it’s efficient. Liquid cooling helps, denser racks too. But averages? Pfft. Real workloads spike, and AI’s a power hog. Europe’s green dreams clash with physics every time.

But.

Here’s my unique hot take, absent from the press release fluff: This reeks of the 1990s supercomputer wars redux. Japan poured billions into Fujitsu’s Earth Simulator — topped Top500, wowed the world. Then? Crickets in AI leadership. Hardware flex doesn’t birth software empires. Europe risks the same: fancy flops, zero ChatGPT killers. AMD’s no Nvidia in CUDA lock-in. Will Alice train Europe’s Llama? Or just simulate fusion that never fuses?

Is Alice Recoque Europe’s Nvidia Killer?

Spoiler: No.

AMD’s pushing hard — MI300X already nipping at H100 heels, Venice promises more. Instinct line’s open ecosystem appeals to the “screw Nvidia” crowd. Europe hates vendor lock-in; hence SiPearl, BXI over InfiniBand. Smart geopolitics? Or penny-pinching?

€554 million from EuroHPC — that’s public cash for “societal challenges.” Simulations for climate? Drug discovery? Fine. But AI supremacy? France trails. Jupiter’s Nvidia Grace-Hopper core laps this in raw AI muscle today. Alice’s accelerators? Competitive, maybe. By 2026? AMD might close the gap — if yields hold, if power doesn’t balloon.

Eviden’s BullSequana: battle-tested, sure. Direct liquid cooling — fanless racks humming quietly. DDN storage, FPGA acceleration. BXI v3 at 800 Gbps? Blistering. But interconnects are the silent killer in exascale. One bottleneck, and your exaflops evaporate.

Delays loom. AMD chips: 2025 maybe. SiPearl: who knows. Site prep in Paris — GENCI’s no vast data center farm. Logistics nightmare.

And power. That walked-back 12 MW. Comparable systems hit 20+. AI workloads? Double it. Europe’s grids strain already; blackouts in France last winter weren’t cute. Green HPC? Noble lie.

Can Europe Catch the AI Train with €544M?

Bold prediction: Alice Recoque tops Top500. Briefly. Then fades like LUMI or MareNostrum upgrades. Hardware’s table stakes now. Software — that’s the moat. OpenAI didn’t win on flops; they won on data, talent, iteration.

France’s play signals intent. Dumping Nvidia — mostly — screams sovereignty. AMD’s US, yeah, but less China-tied. CHIPS Act vibes, Euro style.

Yet skepticism reigns. Eviden queried their own power spin post-publication. PR cleanup on aisle five.

Europe needs this horsepower. Climate models, quantum sims, drug folds — societal wins. AI Factory? Hype. But in a world where US hyperscalers hoard Nvidia, Alice levels the field. Slightly.

Short version: Impressive specs. Questionable timelines. Power claims? Laughable until proven. Europe flexes muscle — watch it atrophy without apps.


🧬 Related Insights

Frequently Asked Questions

What is Alice Recoque supercomputer?

France’s first exascale system, €544M project by Eviden using AMD Epyc Venice CPUs and MI430X GPUs, plus SiPearl Rhea2. Aims at AI, sims, data analysis. Launch ~2026.

When does France’s exascale supercomputer go live?

Accelerator partition late 2026, full CPU cluster 2027 — if chips arrive and no delays.

Why choose AMD over Nvidia for exascale?

Open ecosystem, less lock-in, Europe-friendly supply. But Nvidia dominates AI software; AMD’s catching up on hardware.

Aisha Patel
Written by

Former ML engineer turned writer. Covers computer vision and robotics with a practitioner perspective.

Frequently asked questions

What is Alice Recoque supercomputer?
France's first exascale system, €544M project by Eviden using AMD Epyc Venice CPUs and MI430X GPUs, plus SiPearl Rhea2. Aims at AI, sims, data analysis. Launch ~2026.
When does France's exascale supercomputer go live?
Accelerator partition late 2026, full CPU cluster 2027 — if chips arrive and no delays.
Why choose AMD over Nvidia for exascale?
Open ecosystem, less lock-in, Europe-friendly supply. But Nvidia dominates AI software; AMD's catching up on hardware.

Worth sharing?

Get the best Semiconductor stories of the week in your inbox — no noise, no spam.

Originally reported by The Register HPC

Stay in the loop

The week's most important stories from Chip Beat, delivered once a week.