#515 Feature request: Intel Arc B390 / Panther Lake iGPU support on Windows (64 GB shared RAM)

本文

## Feature request

Please consider support for the **Intel Arc B390 integrated GPU (Core Ultra X9 388H / Panther Lake)** on Windows, using a SYCL/Level Zero or Vulkan backend.

I found the earlier Intel Arc request in #299 and the ongoing SYCL port in #423. This request is specifically for the **B390 integrated GPU with shared system memory on Windows**, rather than the Arc Pro B70 discrete GPU on Ubuntu used to validate that port.

## Hardware and software

- Laptop: MSI Prestige 16 Flip AI+ C3MTG
- CPU: Intel Core Ultra X9 388H, 16 cores / 16 threads
- GPU: Intel Arc B390, integrated; no NVIDIA or AMD GPU detected
- RAM: 64 GB installed (63.5 GiB visible to Windows), reported at 8533 MT/s
- OS: Windows 11 Pro, version 10.0.26200
- Intel graphics driver: 32.0.101.8991
- Storage: Micron 2500 NVMe SSD
- Strata version reviewed: v0.1.36 / current main

## Current limitation and checks performed

The current setup only accepts `--backend cuda` or `--backend hip`, so this hardware has no supported installation path.

The local Vulkan stack successfully enumerates:

```text
deviceType = PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
deviceName = Intel(R) Arc(TM) B390 GPU
driverInfo = 101.8991
```

LM Studio's installed llama.cpp Vulkan runtime (2.29.1) also lists the B390 as an available device. These were **device-enumeration checks only**, not successful Strata inference tests; I have not installed or benchmarked Strata on this machine.

## Requested support

1. A native Windows installation/build path for Intel B390, potentially extending the SYCL work in #423.
2. Detection and allocation planning for integrated/shared-memory GPUs: CPU RAM and GPU memory share the same physical pool and should not be counted as independent capacities.
3. Conservative model/context recommendations for a 64 GB system, with clear errors when a requested allocation cannot fit.

An experimental text-only path would already be useful; image input and performance parity with discrete GPUs are not prerequisites.

Is B390 / Panther Lake support feasible through the existing SYCL port, and what additional device information or kernel checks would be needed to evaluate it?

Thank you for building Strata.