Intel Arc B580 runs 125B AI model nearly 3x faster with XeStrata in new test

A graphics card with 12GB of VRAM might not seem an obvious choice for running a 125-billion-parameter AI model. However, a new test involving Intel's Arc B580 suggests that the software handling the workload can make a substantial difference to performance.

In testing published on October 26, Japanese technology publication PC Watch recorded 31.1 tokens per second using the open-source XeStrata inference engine. The same system achieved 10.9 tokens per second with Strata, the project from which XeStrata was derived.

That works out to approximately 2.85 times the original generation speed, or an improvement of around 185%, without changing the graphics card.

The 12GB GPU is only part of the equation

The test used an ASRock Intel Arc B580 Challenger 12GB OC alongside an Intel Core Ultra 5 250KF Plus processor and 64GB of DDR5 memory. Ubuntu 26.04 LTS was the operating system.

The model was Qwen3.8-Flash-Next, configured with IQ3_S quantisation, a 64K-token context and an 8-bit KV cache. These details are significant because quantisation, context length and memory allocation can all affect inference performance.

Despite the model's 125-billion-parameter designation, the test does not demonstrate that the entire model can fit inside 12GB of VRAM. Strata and XeStrata are designed to make use of system memory alongside GPU acceleration, allowing substantially larger models to run on hardware with comparatively limited graphics memory.

The particular quantisation selected for this test requires approximately 50GB of system RAM according to Strata's setup information. Consequently, the 64GB memory configuration is an essential part of the result, rather than an incidental detail.

For someone considering local AI inference on a budget, that distinction is arguably more useful than the GPU specification alone.

What changed between the two results?

The comparison involved Strata 0.2.1 and XeStrata, a derivative incorporating additional optimisations for supported GPU architectures.

PC Watch installed XeStrata using the 0.1.43-2 Ubuntu package, corresponding to the project's xe0.1.43.1 release. Both engines were tested using the same model and configuration.

XeStrata's release documentation describes improvements to Intel GPU execution paths, including the use of XMX matrix hardware for certain operations. However, the published PC Watch comparison does not establish how much of the measured speed difference comes from any individual optimisation.

The result should therefore be understood as a comparison between two software configurations on the same computer, not proof that one specific technique is responsible for the entire gain.

It also concerns text-generation throughput. It does not establish equivalent improvements in prompt processing, response latency, model accuracy or other AI workloads.

Faster inference does not eliminate hardware limits

At 31.1 tokens per second, the XeStrata configuration produced text at nearly three times the rate observed with Strata. That is a meaningful difference in the reported test, particularly because both results came from the same 12GB graphics card.

Nevertheless, running a large model remains demanding. The test required 64GB of system memory, and the reviewer considered the Arc B580 insufficient for heavier agent-based workloads, where greater GPU performance, more VRAM and longer context windows may be preferable.

There is also an important software-version distinction. The XeStrata package tested by PC Watch is no longer the latest available. The project's October 24 release introduced additional changes and documented Intel driver-related performance issues affecting certain Linux configurations.

Accordingly, the 31.1-token result should not be treated as a benchmark for every XeStrata release or driver combination.

The most interesting takeaway is not that a 12GB GPU has somehow overcome the memory requirements of a 125B model. It is that two inference engines running the same model on identical hardware produced markedly different generation speeds.

For users experimenting with large language models locally, that makes software selection an important consideration alongside VRAM, system memory and the graphics card itself.