everyone's ignoring the $300 GPU that almost matches a $900 RTX 5070 — here's the catch

12GB of VRAM for around $300 sounds like the easy answer to Nvidia's price spike. It mostly is — you just have to fight the drivers a little.

Aliteq
Ravi Malhotra · Hardware Editor

The short version

Arc B580 launched at $249; August 2026 street price is ~$290-$310 — up roughly 20%, versus 36-39% price spikes across Nvidia's RTX 50 lineup this month.

The short version

12GB GDDR6 at 456 GB/s bandwidth, built on Intel's Xe2 (Battlemage) architecture with dedicated XMX tensor cores for INT8/FP16 math.

The short version

Standard Ollama does not run on Arc GPUs out of the box — you need Intel's patched IPEX-LLM fork, which redirects to the oneAPI/SYCL backend.

The short version

Counterintuitively, plain llama.cpp with the Vulkan backend often outperforms Intel's own IPEX-LLM SYCL path on the B580.

The short version

Flash Attention support is patchy — partial on SYCL, unsupported on Vulkan — which caps performance on long-context workloads.

Intel Arc B580 for local AI

A genuinely strong 12GB card at a price that hasn't moved much while Nvidia's lineup spiked — but budget an evening for driver and backend setup, and stick to 7-9B models where its XMX engines…

Aliteq

Read the full story

everyone's ignoring the $300 GPU that almost matches a $900 RTX 5070 — here's the catch

Read the full story on Aliteq