First impressions: Microsoft Surface RTX Spark Dev Box, the screenless Surface that wants your cloud...
Microsoft is charging $5,999.99 for a Surface without a screen. The Surface RTX Spark Dev Box is basically a flat aluminum slab that stays on a developer’s desk and runs AI models on its own hardware. I didn’t expect Surface to go in this direction.
Related: Surface Laptop Ultra vs. MacBook Pro with M5 Pro, and why CUDA ends up deciding the whole fight
My first reaction was a shrug. GB already shipped the same chip in last year’s DGX Spark. The pitch is the part I keep chewing on. Microsoft wants your team to stop renting GPUs by the hour and just own one.
The parts I like
Microsoft Surface RTX Spark Dev Box
Microsoft put 1,000 vents in the lid, one for each teraflop of AI compute it claims. They did it on purpose and made sure everyone noticed, and I’m fine with a little showing off. The aluminum case also works as the heatsink for a 100 W thermal envelope, and from above it looks a lot like an Xbox Series X.
The 128 GB of unified memory matters most to me. Both the CPU and the GPU pull from it, and that’s what lets Microsoft say the box runs models over 120 billion parameters. Plenty of gaming PCs top out at 16 GB of video memory on something like an RTX 5080, a fraction of what the Dev Box has.
I’m more inclined to believe Microsoft’s claim that GitHub Copilot can shift up to 20% of its work to local models than the grander promises surrounding AI PCs. The Windows 11 Pro build also comes with VS Code and WS ready for coding.
What bugs me
NVIDIA called RTX Spark “the most efficient PC chip ever built” back in June. It didn’t show a single chart. Months later, Microsoft’s spec page still doesn’t list memory bandwidth, and with local AI models that figure has a lot to do with how quickly you get answers back.
The petaflop number has a catch as well. Microsoft’s own footnote says it’s measured at FP4 with sparsity turned on. That’s the most generous way to measure it.
So I’m stuck imagining a developer who loads a 120B model on the first day, types a question and waits, and I couldn’t tell you whether the reply would take two seconds or 20, because neither Microsoft nor NVIDIA has put out a single tokens-per-second figure for the chip. I’d want that number first.
Linux worries me a bit too. NVIDIA hasn’t committed to Linux drivers, and plenty of AI tools still assume you’re on Linux. Older x86 Windows apps also need emulation on an Arm chip. Preorders are only on Microsoft.com in the US, and it ships in November.
Who should care
I can see the $5,999.99 price making more sense for a Windows team that keeps paying to test models in the cloud. Local processing gives them a chance to cut some of that expense. I’d struggle to justify the same purchase for a chatbot sitting on my desk.
My take for now
I’d hold off on the Surface RTX Spark Dev Box until someone tests a 120B model and publishes the tokens-per-second results. The hardware looks promising, but the price has climbed above NVIDIA’s original ask for nearly the same chip. Teams paying hefty cloud bills have more reason to take a chance on it than I do right now.
