12 -- AI Inference & LLMs

ℹ️ NOTE

The BC-250 is capable of running local AI inference, including large language models (LLMs) and image generation (Stable Diffusion), using the Vulkan compute backend. Its 16 GB unified memory and 24 RDNA 2 CUs make it a viable budget option for local AI workloads.

Overview

What Works

Tool	Status	Notes
llama.cpp (Vulkan)	Confirmed (elektricM radv.md)	Primary path. Pre-built Vulkan binaries work. Use `-ngl 999`.
llama.cpp (ROCm/HIP)	Partial (Discord community)	hammercoral achieved 2.5-3x speedup over Vulkan but no guide published. As of May 2026, neoney got partial results.
llama.cpp (RPC/distributed)	Confirmed (xseol, Discord)	Multi-board via Ethernet. Performance limited by 1GbE.
Ollama (Vulkan)	Confirmed (Discord community)	Works but lags behind standalone llama.cpp (~56% slower due to vendored old version akandr/bc250).
KoboldCPP	Confirmed (Discord community)	Uses stablediffusion.cpp / GGML-Vulkan.
LM Studio	Confirmed, suboptimal (Discord community)	VRAM/system RAM balancing issues on UMA.
Stable Diffusion (stablediffusion.cpp + Vulkan)	Confirmed (Discord community)	2x performance vs RX 6600 in one user's test.
vLLM	Does not work (Discord community)	Not supported. Use llama.cpp RPC for multi-board.
PyTorch / ComfyUI	Partial — working via ROCm/HIP (gabriwar, Aug 2026)	Requires custom kernel patches + rocBLAS gfx1013 kernels. See GabriWar/bc250-rocm-working ↗ for complete setup guide. Caveats: kernel warm-up required before generating; VRAM OOM on larger batches (gabriwar, Aug 2026).

Why Vulkan and Not ROCm

The BC-250 GPU is identified as gfx1013 (Cyan Skillfish), a hybrid RDNA 1.5 architecture. ROCm does not officially support gfx1013 -- the GPU is absent from AMD's compatibility matrix (ROCm docs). The open-source TheRock project tracks gfx1010/gfx1011/gfx1012 but not gfx1013 (ROCm/TheRock). This means rocBLAS lacks pre-compiled kernels for this architecture, requiring manual compilation or workarounds (elektricM radv.md).

For most users, the Vulkan backend via RADV is the reliable path. Recent llama.cpp Vulkan improvements (Wave32 flash attention PR #19625, graphics queue PR #20551) have closed much of the performance gap between Vulkan and HIP backends on AMD hardware (llama.cpp issue #15601).

ROCm/HIP Setup Guide (gabriwar, Aug 2026)

Repository: GabriWar/bc250-rocm-working ↗

Stable Diffusion via ComfyUI with PyTorch ROCm native backend is now possible — the first published complete working guide. Requires kernel patches + rocBLAS gfx1013 kernels (gabriwar: "the kernel patch from neoney IS needed", 04/08/2026). Published by gabriwar on Aug 5, 2026 with full investigation notes. Root cause of prior failures identified: hipFree requests a TLB invalidation the board never performs, so reused virtual addresses keep translating through their previous mapping; the bc250_flush_by_runlist=1 patch (rebuild the runlist on unmap) fixed it — 13/18 dirty runs → 0/18 (p = 3.7e-06).

Benchmarks (SD 1.5, 512×512, dpmpp_2m + karras): Sampler runs at 1.50 it/s (24 steps in 16.0s). Full pipeline: 35.3 seconds with VAE decode on CPU, 17.9 seconds without decode (load + sample only). ~17 seconds per image reported by gabriwar (Aug 2026); repo later measured 14.1-14.5s warm with VAE on GPU.

Performance scaling: 8-core unlock gives ~15-20% more performance; 40 CU gives marginal gains in shader computation (gabriwar, Aug 2026).

Caveats: Early versions required a GPU warmup before generation (first-gen could produce trash output); the later runlist TLB fix removed that need. VRAM OOM on larger workloads. System freezes possible under heavy compute loads (see ROCm #6313 regression — geenight, May 2026).

Limitation: Unlike Vulkan which works out of box, ROCm still requires kernel modifications and manual compilation for gfx1013.

Quick Start: llama.cpp + Vulkan (Recommended)

BIOS and Kernel Configuration

Set VRAM to 512MB dynamic in BIOS. The rest is allocated dynamically via GTT:

BASH

# Add to kernel parameters for ~14 GB VRAM
amdgpu.gttsize=14750 ttm.pages_limit=3959290 ttm.page_pool_size=3959290

Install llama.cpp

BASH

# Create directory and download pre-built Vulkan binary
mkdir -p ~/llama && cd ~/llama
wget https://github.com/ggerganov/llama.cpp/releases/download/b7150/llama-b7150-bin-ubuntu-vulkan-x64.zip
7z e llama-b7150-bin-ubuntu-vulkan-x64.zip && rm llama-b7150-bin-ubuntu-vulkan-x64.zip

Run the Server

BASH

./llama-server -m /path/to/model.gguf -c 40960 -ngl 999 -fa -ctk q4_0 -ctv q4_0 --jinja --host 0.0.0.0 --port 8080

Key flags:

-ngl 999 -- offload ALL layers to GPU (critical for performance)

-fa -- Flash Attention (requires Mesa 25.1+ with Wave32 support)

-ctk q4_0 -ctv q4_0 -- KV cache quantization

--jinja -- use chat template from model file

Avoid OOM Crashes

The -cram flag prevents OOM by limiting context RAM allocation. Without it, llama.cpp defaults to 8 GB system RAM assumption which will cause OOM on the BC-250 (Discord community).

BASH

export GGML_VK_FORCE_MAX_ALLOCATION_SIZE=2000000000  # 2 GB chunks

Headless Mode (Maximum RAM)

BASH

sudo systemctl set-default multi-user.target   # saves ~600 MB vs GUI

Performance Benchmarks

Single Board -- llama.cpp (Vulkan)

Community-tested performance (Discord):

Model	Quant	Prompt (pp512)	Generation (tg128)	User
TinyLlama 1.1B	Q4_K_M	326 tok/s	49.8 tok/s	hammercoral
Llama 3.1 8B	Q4_K_M	281 tok/s	52.3 tok/s	__nightfox
Gemma 3 12B	Q8	--	~30 tok/s	xseol
GPT-OSS 20B	Q4	--	~70 tok/s	_fanoush_
Qwen3-Coder-30B-A3B (MoE)	Q4	--	~70 tok/s	machinezer0
MiroThinker-14B	IQ4_XS	--	12-15 tok/s	steinbeks
Qwen3-4B	Q4_K_XL	--	~50 tok/s (after fixing layer offload)	birdetta
GLM 4.6V	Q4_K_M	--	12 tok/s (6->12 after llama.cpp update)	xseol

BC-250 in llama-bench (Standardized Benchmark)

From the official llama.cpp Vulkan benchmark thread (llama.cpp issue #10879, ggml-org):

Model	Size	Backend	Prompt (pp512)	Generation (tg128)
Llama 2 7B Q4_0	3.56 GB	Vulkan	370.66 tok/s	62.32 tok/s
Llama 2 7B Q4_0	3.56 GB	Vulkan (later build)	356.87 tok/s	63.14 tok/s

The BC-250's generation speed is comparable to an RTX 3070 Mobile (63.64 tok/s) and GTX 1080 Ti (64.63 tok/s) on the same benchmark.

ROCm vs Vulkan Comparison

hammercoral's benchmarks on a single BC-250 (Discord):

Model	Metric	Vulkan	ROCm/HIP	Speedup
TinyLlama 1.1B	pp512	326 tok/s	709 tok/s	2.17x
TinyLlama 1.1B	tg128	49.8 tok/s	151 tok/s	3.03x
Llama 3.1 8B	pp512	46.6 tok/s	115 tok/s	2.48x
Llama 3.1 8B	tg128	12.2 tok/s	30 tok/s	2.46x

Note: On newer RDNA hardware, Vulkan has been consistently matching or beating ROCm for token generation due to Vulkan-specific Wave32 optimizations (llama.cpp issue #20934). These benchmarks may not reflect the current state.

Multi-Board RPC (Distributed)

xseol's multi-board data (Discord):

Model	Boards	Generation	User
GLM-4.5-Air Q4_K_S	6	Starts at 9.5 tok/s	xseol
ERNIE-4.5-21B	2 (Q8)	Starts at 30 tok/s	xseol
GPT-OSS-20B	2 (Q8)	Starts at 23 tok/s	xseol
Gemma 3 27B	2 (Q6_K)	Starts at 9 tok/s	xseol
DeepSeek Coder V2 Q4	2	Prompt 48.6, Gen 20.7 tok/s	adaptive__manipulator

Model Compatibility (Single Board, 16 GB)

Fits Comfortably

Qwen 3.5 4B-9B (Q4_K_M) -- ~50 tok/s

Llama 3.1 8B (Q4_K_M) -- ~52 tok/s

Gemma 3 12B (Q8) -- ~30 tok/s

Mistral Nemo 12B (IQ4_K or Q4_K_M)

Phi-4 (various quants)

Qwen3-Coder-30B-A3B (MoE, fits as mixture of experts)

GPT-OSS 20B (Q4) -- ~70 tok/s, cuts it close

Requires Careful Quantization

Gemma 4 26B MoE (Q3) -- barely fits, 48K context

MiroThinker-14B (IQ4_XS)

Requires Multiple Boards

70B models -- requires 5+ boards via RPC

GPT-OSS 120B -- requires 6+ boards via RPC

Alternative: Disk-Streamed MoE (colibri)

JustVugg/colibri ↗ can run GLM-5.2 (744B MoE) on a 25GB-RAM machine by streaming experts from disk — pure C, zero dependencies. For BC-250 LLM nodes with 16GB unified memory, this allows running far larger models than fit in RAM by trading speed for capacity. (Note: community testing on BC-250 pending as of Jul 2026.)

ROCm/HIP Status

ROCm on gfx1013 (Cyan Skillfish) is partial and has known regressions:

hammercoral achieved working ROCm/HIP/PyTorch compute with "firmware on MEC changed, bios settings, recompiling multiple software stacks" but published no setup guide (Discord).

neoney (May 2026): Got a single token via ROCm but failed when attempting to offload more layers (Discord).

ROCm regression (geenight, May 2026): Issue filed as ROCm #6313. BC-250 system freezes completely after compute workloads. Worked with ROCm 5.2 on Ubuntu 20.04, but newer versions cause full system freezes requiring power cycle. scallion_9883 confirmed building a custom HIP kernel and warned that memcpy + worker reaping can wedge the board.

rocBLAS aborts on detection because gfx1013 is not in the GPU list (elektricM radv.md).

RDNA1 architectures gfx1010/gfx1011 are build-passing in TheRock but gfx1013 is absent (ROCm/TheRock SUPPORTED_GPUS.md).

Community sentiment: "The last person who had ROCm working promised a guide then disappeared forever" (Discord).

HSA_OVERRIDE_GFX_VERSION=10.3.0 may enable partial ROCm stack support but comes with risks and no guarantees.

For most users, Vulkan is the recommended backend for all AI inference workloads on the BC-250 (elektricM radv.md).

Ollama Notes

Ollama works with its Vulkan backend on the BC-250 but has two important limitations (Discord community, akandr/bc250 repo):

1. Outdated vendored llama.cpp: As of early 2026, Ollama vendors llama.cpp at b7437 (Dec 2025), missing Wave32 flash attention (PR #19625) and graphics queue fixes (PR #20551) that provide ~56% improvement on AMD Vulkan (llama.cpp issue #15601).

2. VRAM/system RAM balancing: Unified memory on the BC-250 can cause Ollama to overallocate system RAM under its default assumptions.

For best performance, use standalone llama.cpp directly instead of Ollama.

Stable Diffusion

Stable Diffusion runs via stablediffusion.cpp with the Vulkan backend (Discord community):

One community member reported 2x performance vs an RX 6600 on the same workload.

Video/latent models are very slow on a single board.

The main limitation is the lack of PyTorch ROCm support, which blocks ComfyUI and similar tools.

40 CU Unlock (Active Community Project)

Repository: duggasco/bc250-40cu-unlock ↗

Status: Working, 19 stars, 1 fork (May 2026). 1.61x compute scaling verified.

The BC-250 ships with 24 of 40 RDNA2 CUs active (16 harvested). These can be re-enabled via bc250-cu-live-manager (UMR-based, no kernel patch/reboot, recommended) or the duggasco kernel patch (legacy). See 02-BIOS & Firmware for full procedures.

LLM Performance with 40 CUs (llama.cpp Vulkan, Qwen3.5-9B Q4_K_XL)

Config	pp512 tok/s	Power	Temp	SCLK
Stock 24 CU	230	95W	79C	1500 MHz
40 CU unlocked	372	125W	83C	1500 MHz
Ratio	1.61x	+30W	+4C	same
40 CU @ 2 GHz (governor)	466	181W	96C	2000 MHz

Qwen3.5-9B @ 40 CU (May 2026): +25% tok/s over 24 CU on llama-server with Q4_K_M.gguf (community test, May 2026). Furmark showed ~50% FPS gain, so LLM workload scales differently than gaming — prompt batching and context size tuning needed for optimal results.

Recommended sweet spot: 1500 MHz / 900 mV via cyan-skillfish-governor. 40 CU at 2 GHz hits 96C and requires upgraded cooling (duggasco, scallion_9883).

Community Resources

akandr/bc250 (GitHub) -- Ollama + Vulkan inference guide for BC-250

kingofgames0880 VC-250 guide (Discord bc250-resources) -- Complete llama.cpp + Crush CLI setup

hammercoral ROCm benchmarks (Discord bc250-chat) -- Best ROCm performance data available

NexGen-3D SteamMachine repo (GitHub) -- Automated setup scripts for Bazzite including governor + swap

scallion_9883 HIP kernel work (Discord bc250-chat) -- Custom HIP kernel for BC-250, ROCm debugging

ROCm issue #6313 -- BC-250 system freeze after compute workloads. Filed by geenight, community investigation ongoing

elektricM radv driver guide -- RADV setup and LLM inference notes (elektricM.github.io/amd-bc250-docs/drivers/radv/)

Sources: elektricM radv.md (primary), Discord bc250-chat (hammercoral, __nightfox, xseol, _fanoush_, steinbeks, deathstalkerjr, adaptive__manipulator, birdetta, machinezer0, neoney), llama.cpp GitHub benchmark thread #10879, Ollama issue #15601, ROCm/TheRock SUPPORTED_GPUS.md, Phoronix coverage.

Last modified: 2026-08-13