Microsoft updated Windows AI roadmap, provided clear indication for local AI PC hardware requirements. This is very important because they are putting hard-coded memory requirement for local AI so we know exactly how much memory is required to do sophisticated AI things locally, not on cloud. In this way, whoever buying a future proof PC, he/she knows exactly what type of PC he needs to buy.
Roadmap establishes memory floors for local AI processing across standard, Copilot+, and Builder PCs
Here are 3 tiers base on memory capacity instead of clock speed. Tier 0 is for normal Windows PC. Minimum memory is 8GB.
Spec comparison
| Spec | Tier 0 | Tier 1 | Tier 2 |
|---|---|---|---|
| Memory Requirement | 8GB | 16GB, 32GB, 64GB | 128GB or more |
| Target Use Case | Standard Windows PC | Copilot+ PC | Builder PC |
| Example Processors | Current Gen | AMD Ryzen AI 500, Intel Core Ultra Series 4 | Snapdragon Next, NVIDIA RTX Spark |
Tier 1 is Copilot+ PC. Minimum memory is 16GB, 32GB or 64GB. Tier 2 is for Builder PC.
Minimum memory is 128GB or more. These high end machines are meant to be used in high work load environment.
Microsoft named some upcoming processor family for the mentioned tiers: – AMD Ryzen AI 500. Zen6 – Intel Core Ultra Series 4. Nova Lake – Intel Core Ultra Next. (generic term for future Gen)
Also they mentioned high memory for specific hardware. Such as – Snapdragon Next. – NVIDIA RTX Spark. – Intel Crescent Island. (Intel's officially announced inference GPU architecture.) They confirmed up to 480GB LPDDR5X for this gpu.
AMD or Intel hasn't officially announced the commercial name of this next Gen client processor yet. This is a roadmap for industry. It's not final product launch schedule. Microsoft is using it to show the world what is their plan about hybrid intelligence for different hardware capabilities.
128GB for a builder PC is insane. My RAM budget just cried. 💀
Tier 2 is basically a workstation masquerading as a consumer device.
Does the 128GB requirement imply unified memory or discrete VRAM? The article mentions LPDDR5X for Intel but is ambiguous on the CPU side.
The roadmap lists Intel Crescent Island with 480GB LPDDR5X. For the client CPUs like Ryzen AI 500 it likely relies on unified system memory since discrete GPU memory allocation isn’t specified for Tier 2 client targets.
assuming the 16gb tier actually runs anything useful tbh
Memory capacity does not equal inference speed. Without independent benchmarks on token generation rates this is just marketing fluff.
Linux compatibility for these local AI stacks will depend heavily on driver support for the Crescent Island architecture. Proton is not a magic wand for proprietary kernel modules.
The driver overhead for managing 128GB of local model weights in user space is going to be a nightmare for Windows memory management.
I just need 8GB to survive my current workflow. This roadmap is for people with money to burn.