1.5 TB of VRAM on Mac Studio – RDMA over Thunderbolt 5

Apple’s $40k desk supercomputer splits the crowd: genius or cable spaghetti

TLDR: Apple tested a Mac Studio cluster that pools 1.5 TB memory over Thunderbolt to run big AI locally. The crowd split: fans love quiet, private power, while critics want server-grade networking, neural accelerators, and more RAM; skeptics warn not to trust AI blindly. It’s a glimpse at homegrown AI muscle.

Apple loaned a $40k stack of Mac Studios to demo a wild new trick: multiple Macs sharing one giant pool of memory (1.5 TB!) using RDMA—Remote Direct Memory Access—over Thunderbolt. Translation: they act like a single mega‑Mac, slashing lag from 300 microseconds to under 50. Cool tech, yes. But the comments turned it into a Thunderbolt soap opera. Some cheered the quiet, low‑power “home supercomputer,” while others dragged the cable spaghetti and the lack of proper network switches. One fan called the project “awesome” and thanked the creator for the good vibes, but the hardcore crowd wanted server‑grade features: behnamoh demanded DGX‑style high‑speed links and neural chips, not just fancy cables. rahimnathwani brought receipts, citing giant models like DeepSeek that need 700 GB just for the weights, asking if splitting across nodes truly helps. And delaminator channeled the “trust, but verify” mood: treat AI like Wikipedia—use it, don’t worship it. Meanwhile, newsclues’ video link fueled “this is evolving fast” hype. Jokes flew about Apple’s $70 cables, pressing the rear power button like rack‑mount yoga, and Thunderbolt “friendship bracelets” connecting every Mac to every other. The verdict: spectacular demo, spicy debate.

Key Points

  • Apple tested RDMA over Thunderbolt in macOS 26.2 on a four‑Mac Studio cluster totaling 1.5 TB unified memory.
  • RDMA reduced memory access latency from about 300 microseconds to under 50 microseconds, improving performance for large AI models.
  • The cluster comprised two 512 GB/32‑core Mac Studios ($11,699 each) and two with half the RAM ($8,099 each), costing just under $40,000.
  • Thunderbolt provided ~50–60 Gbps real‑world throughput; Thunderbolt 5 switches are unavailable, requiring direct cross‑connections up to four Macs.
  • The setup used a DeskPi TL1 mini rack; Mac Studios ran under ~250 W and remained quiet, suitable for creative and small‑scale scientific workloads.

Hottest takes

"Something like DGX QSFP link (200Gb/s, 400Gb/s) instead of TB5" — behnamoh
"don’t ever let AI replace your ability to think critically!" — delaminator
"weights take up 700GB RAM" — rahimnathwani
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.