NVIDIA Wants To Turn Your Home Wall Into A Mini Data Center By Packing 16 Blackwell GPUs And 4 server CPUs In A Liquid-Cooled, Wall-Mounted Box At $0 Upfront Cost - Wccftech

Skyler Monroe (0:10)

Hey, everyone, welcome back to the AI Hardware Show. I'm your host, Skyler Monroe, and if you're new here, this is the place where we break down everything happening in AI hardware, from chips and GPS to data centers and custom silicon. Big thanks to our sponsors keeping this show running. Ada, helping businesses integrate AI into their workflows.

Ago Consulting, that's Ago, your go-to for silicon development from AI to so, Su, and Zen Semiconductor, an umbrella company building the processors and AI fabric behind modern data centers with ventures like their Sierra RISC-V CPUs and Lume AI Fabric plus Ada for deploying AI at scale in any business.

Today, we've got a packed episode. Nvidia wants to mount a mini data center on your living room wall. Intel is finally giving its Crescent Island accelerator, a real name and some serious specs. We're talking about the unfortunate cancellation of the Grace Hopper Celebration, and we're diving into a fascinating piece of chip architecture called the GPN-PU. Let's get into it. All right. First up, and honestly, this one kind of blew my mind when I saw the headline, Nvidia wants to put 16 Blackwell GPS on your wall.

Like literally on your wall. We're talking about a wall-mounted liquid-cooled unit that packs in 16 Blackwell GPS and four server class CPUs, all in what sounds like a sleek, almost appliance style form factor. And here's the kicker, zero dollars upfront cost. So let's unpack what's actually going on here, because there's a lot to dig into. First, the hardware itself. 16 Blackwell GPS is not a small number. To put that in perspective, a single Blackwell GPU, we're talking about Nvidia's latest generation architecture, is an absolute beast. These are the chips powering frontier AI model training and inference in the biggest data centers on the planet. 16 of them together? That's serious compute. Add four server CPUs into the mix, and you're looking at a system that if you dropped it to a rack in a traditional data center, would be considered a meaningful cluster in its own right. Now the liquid cooling piece is really important here, and I don't want to gloss over it. Blackwell GPS run hot, really hot. Air cooling at that density just doesn't cut it. You'd need industrial grade airflow, and even then you'd be pushing the limits. Liquid cooling is what makes it possible to pack that much compute into a wall-mounted form factor without turning your home into a sauna. Think of it like the difference between cooling a server room with a box fan versus a proper HVay system. Liquid cooling is just a fundamentally more efficient way to move heat away from the silicon. So why would Nvidia do this? Well, the $0 upfront cost model is the real story here. This is almost certainly a compute as a service play. Nvidia isn't just selling you hardware anymore. They want to be in the business of selling you compute capacity, kind of like how a solar panel company might install panels on your roof for free in exchange for selling the electricity back to the grid. The idea would be that Nvidia or a partner installs this unit in your home or office, and you get access to some portion of that compute while the rest gets used for their cloud services or distributed inference workloads. This is a fascinating strategic move because it solves a few problems at once. Data centers are struggling with power and real estate constraints. There's only so much land and electrical capacity near major population centers. If you can distribute compute out to the edges, into homes and offices, you potentially get around some of those bottlenecks. It's also a way for Nvidia to get Blackwell Silicon deployed at a massive scale without waiting for hyperscalers to build more data centers. Whether this actually works logistically getting liquid cooling infrastructure into residential spaces, managing maintenance, dealing with network latency, those are all real challenges. But as a concept, it's bold and it signals where Nvidia sees the future of compute distribution heading. Okay, next story, Intel. And I have to say, Intel has been having a rough go of it lately on multiple fronts. But here's a piece of news that actually looks promising.

Their AI Inference Accelerator, which has been going by the code name Crescent Island in development circles, has now officially been named and specced out. Meet the Intel GPU D-Series. The headline numbers here are genuinely impressive. The flagship model, the D27C, is rated at 1.3 petaflops of dense FP8 compute.

For context, a petaflop is one quadrillion floating point operations per second.

An FP8 is a lower precision number format that's become the go-to for AI inference workloads because it lets you pack more math into the same silicon area without sacrificing too much accuracy for inference tasks. So 1.3 petaflops of FP8 is a meaningful number.