← sources

Gemma 4 E2B on a 8GB RAM netbook

source: reddit (reddit:LocalLLaMA)published 2026-11-13 00:31ingested 2026-11-14 02:42

retained text

Hey everyone,

Not a native speaker, but I write everything by hand. Please correct my grammer if there are any issues! Can only learn from it.

# Background

A long-term project I've had was getting my netbook up and running for basic LLM use. I bought the netbook initially due to my poor eyesight, the screen is small enough that it fits in my cone of vision.

The notebook's specs were practically obsolete when it released:

* Release: December 2021
* Device: ASUS E210MA
* CPU: Intel Celeron N4020
* RAM: 4GB 2400MHz DDR4 single-channel
* iGPU: Intel UHD Graphics 600
* Storage: 64GB eMMC
* OS: Windows 10 Home
* Screen: 11'6" TN 1366x768

With the help of friends who could do soldering and reballing, we upgraded the netbook:

* CPU: Intel Pentium Silver N5000
* RAM: 8GB 2400MHz DDR4 single-channel
* iGPU: Intel UHD Graphics 605
* Storage: 128B eMMC + Intel 660P 512GB NVME
* OS: Windows 11 IoT LTSC Enterprise 24H2

After finding proper drivers for the iGPU, I managed to get llama.cpp running on it on windows with Vulkan. Gemma 4 E2B IT QAT could run at roughly 2.1 t/s gen, and prefill speed was "grab a cup of tea" slow on 300 token prompts. The Vulkan backend was faster than the CPU backend by \~30%.

The past months I've been trying to move my devices from Windows 11 IoT LTSC Enterprise 24H2 to Ubuntu 26.04.1 LTS, and so I tried to switch the netbook over to it.

There is only one problem: The Vulkan backend hard crashes every time.

# Getting vulkan to work

After doing a bit of digging with `vulkaninfo` by myself, some instructions are advertised as supported but are in practice not stable. Using this fixed the crashing on startup:

    export GGML_VK_DISABLE_ASYNC=1

Disabling the following features helped me regain some performance and stability:

    # UHD 605 compat
    export GGML_VK_DISABLE_COOPMAT=1
    export GGML_VK_DISABLE_COOPMAT2=1
    export GGML_VK_DISABLE_INTEGER_DT_PRODUCT=1
    export GGML_VK_DISABLE_BFLOAT16=1
    export GGML_VK_DISABLE_GRAPH_OPTIMIZE=1
    # UMA
    export GGML_VK_DISABLE_MMVQ=1
    export GGML_VK_PREFER_HOST_MEMORY=1

My config is also pretty specific, since some combinations would cause crashes too:

    flash-attn = off
    cache-type-k = f16
    cache-type-v = f16

FA is not supported by the device. Running Q8\_0 on both K an V would crash with FA enabled. Trying to run Gemma 4 E2B IT QAT Q2\_K\_XL on ubuntu always resulted in crashing (either during startup or when generating), On windows those things never caused crashes.

The end result is that I could run Gemma 4 E2B IT QAT Q4\_K\_XL at 1.5 t/s gen and 16-25 t/s prefill. The laptop used \~12W in total.

# What about the CPU?

Turned out that on Ubuntu 26.04.1 LTS, the CPU backend was faster for generating but not for prefill. I got 2 t/s gen and 10-16 t/s prefill. The laptop used \~13W in total.

As a sidenote, since I am used to windows slowing to a crawl when all cores are utilized, I was pleasantly suprised when this didn't happen on ubuntu.

# As a benchmark?

As a fun experiment, I gave a couple of LLMs the problem:

    Hey <LLM>,
    
    I have a laptop I am trying to run llama.cpp with vulkan on Ubuntu 26.06.1 LTS.
    The specs are: <specs>.
    I build it with these commands: <commands>,
    The problem is that it's crashing whenever I try to load a model.
    I've included the full terminal output in the attachments.
    
    How can I fix the issue?

My goal with the prompt is to test the model's understanding of the problem and to ask for additional information when required (e.g. vulkaninfo).

Qwen 4 max preview on [chat.qwen.ai](http://chat.qwen.ai) and Qwen3.8-Flash-Next Q4\_K\_XL + BF16 engram on llama.cpp v0.9.0 failed to find the issue after multiple separate chats. They got focussed on disabling intel's hang detection. With websearch, it got even more confused as it looked at irrelevant issues.

Gemma 4 31B IT QAT Q4\_K\_XL + MTP on llama.cpp, DeepSeek v4 Flash Vision Exp Q2\_K\_XL on llama.cpp v0.9.0, DeepSeek v4.1 1017 on [chat.deepseek.com](http://chat.deepseek.com) were able to find and fix the issue. They asked for the `vulkaninfo` output, which made them find async being advertised incorrectly.

# Appendix

>run-llama-server.sh

    #!/bin/bash
    
    # Intel UHD Graphics 600 series
    export GGML_VK_DISABLE_ASYNC=1
    export GGML_VK_DISABLE_COOPMAT=1
    export GGML_VK_DISABLE_COOPMAT2=1
    export GGML_VK_DISABLE_INTEGER_DT_PRODUCT=1
    export GGML_VK_DISABLE_BFLOAT16=1
    export GGML_VK_DISABLE_GRAPH_OPTIMIZE=1
    # UMA
    export GGML_VK_DISABLE_MMVQ=1
    export GGML_VK_PREFER_HOST_MEMORY=1
    
    ./llama.cpp/build/bin/llama-server \
    --list-devices
    
    ./llama.cpp/build/bin/llama-server \
    --host 127.0.0.1 \
    --port 5001 \
    --api-key "API_KEY_HERE" \
    --webui-mcp-proxy \
    --offline \
    --cache-ram 0 \
    --ctx-checkpoints 0 \
    --parallel 1 \
    --models-max 1 \
    --models-preset ./configs/llama-config.ini

>llama-config

note: remove the device line to use the igpu on the vulkan build.

    [*]
    device = none
    load-mode = none
    mmproj-offload = off
    gpu-layers = 999
    batch-size = 2048
    ubatch-size = 512
    threads = 4
    fit = off
    flash-attn = off
    kv-unified = on
    cache-type-k = f16
    cache-type-v = f16
    reasoning = off
    reasoning-preserve = off
    
    [gemma-4-e2b-q4_k_xl]
    model = ./models/gemma-4-e2b/gemma-4-E2B-it-qat-UD-Q4_K_XL.gguf
    mmproj = ./models/gemma-4-e2b/gemma-4-E2B-it-qat-mmproj-F16.gguf
    lazy-mode = on
    load-mode = mmap
    fit-ctx = 8192
    ctx-size = 8192
    predict = 8192
    image-min-tokens = 70
    image-max-tokens = 70
    temp = 1.0
    top-k = 64
    top-p = 0.95
    min-p = 0.0

>Other info

Device: [https://www.asus.com/ca-en/laptops/for-home/vivobook/vivobook-go-12-l210/](https://www.asus.com/ca-en/laptops/for-home/vivobook/vivobook-go-12-l210/)  
OS: [https://ubuntu.com/blog/upgrade-your-desktop-ubuntu-26-04-lts](https://ubuntu.com/blog/upgrade-your-desktop-ubuntu-26-04-lts)  
Llama.cpp: [https://github.com/ggml-org/llama.cpp/releases/b12110](https://github.com/ggml-org/llama.cpp/releases/b12110)  
GGUF: [https://huggingface.co/unsloth/gemma-4-E2B-it-qat-GGUF](https://huggingface.co/unsloth/gemma-4-E2B-it-qat-GGUF)

# TL;DR

Getting the Intel UHD 605 to work with the Vulkan backend was a pain on Ubuntu 26.04.1 LTS compared to Windows 11 IoT LTSC 24H2. The issue was vulkan async causing crashes, disabling it fixed it.

On ubuntu the CPU backend was faster, on windows the iGPU over Vulkan was faster. Ubuntu is far more responsive than windows, even under full utilization.

After solving it myself, as a fun benchmark I made LLMs try diagnose and fix the issue. While the Qwen series failed to fix the issue from intent, Gemma 4 and DeepSeek v4(.1) Flash (Vision Exp / 1017) did manage it.

# Thanks for reading!

I know the post is quite different from the Strata stuff usually here, but I thought it would be fun to share regardless.

I'm sure models like LFM 3 will run wonderfully on it, I know IBM Granite 4 H 1B and Qwen3.5-2B did on windows.

document fetched 2026-11-14 00:47 · canonical https://reddit.com/r/LocalLLaMA/comments/1xutkqv/gemma_4_e2b_on_a_8gb_ram_netbook