Is the Arc A770 Worth Buying in 2026?
The Arc A770 launched in 2022 to middling reviews and a reputation for rough drivers. That reputation is out of date. The drivers matured. The card did not change.
What you get at $360–400 used: 16GB of GDDR6 VRAM, an open-source xe driver that lives in the mainline Linux kernel, and the ability to run Llama 3 70B Q4 — fully, in VRAM, without CPU offload. No other card in this price range does that.
The RTX 4070 costs more and ships with 12GB. The RTX 3090 with 24GB runs $450–600 used. The A770 sits between them on VRAM at a price point neither reaches.
This is not the card you buy for value. The GTX 1080 Ti at $97 wins on value. This is the card you buy when 16GB is non-negotiable and you will not pay RTX 3090 prices to get there.
Current market: US sellers have clean units at $359–$380. Avoid China-shipped listings — $50 shipping and 3-week delivery at $400 total makes no sense when domestic options exist at the same price.
Current Arc A770 Prices on eBay
Listed as for parts or untested — skip.
Only 13 feedback — limited seller history.
Ships from China — long delivery window, no US warranty, possible China-market SKU.
Only 5 feedback — too little history to assess seller reliability. eBay Money Back Guarantee applies.
Ships from China — long delivery window, no US warranty, possible China-market SKU.
Ships from China — long delivery window, no US warranty, possible China-market SKU.
Only 5 feedback — too little history to assess seller reliability. eBay Money Back Guarantee applies.
Ships from China — long delivery window, no US warranty, possible China-market SKU.
Only 0 feedback — too little history to assess seller reliability. eBay Money Back Guarantee applies.
Arc A770 Specifications
Arc A770 on Linux — Driver Guide
Intel's xe driver lives in the mainline Linux kernel — the same model as AMD's amdgpu. No company decision can end-of-life it. Ubuntu 24.04 LTS loads it automatically on boot. You do not install a driver. For LLM inference, you need the OpenCL compute runtime as one additional step.
lspci -k | grep -A 2 'VGA'
# expected: Kernel driver in use: xe
# OpenCL compute runtime — required for LLM inference
sudo apt install intel-opencl-icd
# Verify compute runtime sees the GPU
clinfo | grep 'Device Name'
xe driver. Workaround: fall back to i915 or install compute-runtime packages manually from intel/compute-runtime releases. Solvable — but requires Linux fluency.Gaming on Linux: the A770 runs roughly 19% faster on Linux than Windows — the Mesa/ANV Vulkan stack is better optimized for this hardware than the proprietary Windows driver. DirectX 9 games via Proton require compatibility checking.
Running Local AI on the Arc A770 — Ollama Performance
Token throughput is slower than the GTX 1080 Ti on CUDA — IPEX-LLM is not as mature as the CUDA/Ollama stack. The tradeoff: you run larger models. Llama 3 70B Q4 fits in 16GB where both the 1080 Ti and RX 6700 XT hit the wall and offload to CPU.
4 tokens/second on 70B is usable for drafting and summarization. Not fast. For users who need 70B locally, no other option at this price exists.
Test environment: Ubuntu 24.04 · IPEX-LLM 2.1 · xe driver · Ryzen 5 5600X · 32GB DDR4. Results approximate.
16GB VRAM — How Does It Compare?
VRAM determines which models you can run at full speed. The A770 at 16GB sits in a tier almost nothing else reaches at this price point.
Should You Buy the Arc A770?
- You need 70B models to run fully in VRAM — non-negotiable
- You care about driver longevity and mainline kernel support
- You won't pay RTX 3090 prices for 24GB
- You're comfortable with Linux setup beyond apt install
- Most VRAM per dollar vs. any new card at this price point
- You want the simplest Ollama/CUDA path — get the GTX 1080 Ti
- You want best value for 12GB — get the RX 6700 XT
- Raw tok/s on small models matters more than model size
- You run DX9 games via Proton without researching compatibility
- Budget is under $300 — better options exist below this price
- 4 tok/s on 70B is too slow for your workflow
- RTX 3090 drops below $380 in your market
- You need CUDA for tools that don't support IPEX-LLM
- You move to serious 4K gaming workloads