Future TechnologyNvidia built a laptop chip with an Arm CPU, Blackwell GPU, and 128GB of RAM
- RTX Spark pairs a 20-core Arm CPU co-designed with MediaTek and a Blackwell-class GPU with 6,144 CUDA cores
- Up to 128GB LPDDR5X unified memory at 300 GB/s bandwidth, enough to run a 70B parameter AI model locally
- The whole package runs at 80W and claims 1 petaflop of FP4 AI performance
- First machines ship autumn 2026 from Microsoft, ASUS, Dell, HP, Lenovo, and MSI
The spec sheet that actually matters
Twenty Arm CPU cores. A Blackwell-class GPU with 6,144 CUDA cores. Up to 128GB of LPDDR5X memory shared across both, connected over NVLink-C2C at 300 GB/s. The whole thing runs at 80W.
Nvidia and Microsoft announced RTX Spark at Computex on May 31, 2026, and first laptops are expected this autumn from Microsoft (Surface Laptop Ultra), ASUS, Dell, HP, Lenovo, and MSI.
What 128GB of unified memory actually means
The number that changes things is not the CUDA core count or the 1 petaflop FP4 claim. It is the 128GB of unified memory.
A 70 billion parameter AI model quantised to 4-bit precision needs roughly 35GB of RAM to load. Most laptops today top out at 32GB. RTX Spark at its maximum configuration gives you nearly four times that, in a single shared pool that both CPU and GPU can access without copying data back and forth.
That is the difference between running a capable local AI model and paying for a cloud subscription. For developers, researchers, and anyone handling data they cannot send to a third party, it matters.
The Arm question
RTX Spark runs on Arm, not x86. The CPU was co-designed with MediaTek, and it runs Windows on Arm. Qualcomm Snapdragon X Elite proved last year that Arm laptops can handle mainstream Windows workloads. Apple proved it even earlier with M-series silicon.
But Nvidia is betting that its GPU advantage makes RTX Spark different from every other Arm laptop chip. Qualcomm has a capable integrated GPU. Apple has a capable integrated GPU. Neither has 6,144 CUDA cores or NVLink connecting CPU to GPU at 300 GB/s.
The trade-off is software compatibility. Windows on Arm still has gaps, though they are shrinking steadily. Anything that depends on x86-specific instructions, like certain creative tools, legacy enterprise software, and older games, will need emulation or a rewrite.
Who this is for
Nvidia is not targeting the average laptop buyer. RTX Spark at 128GB will likely cost more than most workstation desktops. The target audience is AI developers who want to prototype locally, enterprise users running inference on sensitive data, and creative professionals whose workflows need both strong GPU compute and deep memory pools.
The 80W power envelope is the constraint worth watching. Desktop Blackwell GPUs pull 300W or more. Compressing that architecture into 80W means clock speeds and core utilisation will be lower. Whether RTX Spark delivers workstation-class output or just workstation-class specs on a laptop power budget is something only independent benchmarks will answer this autumn.
