Your PS5 Could Run AI. Sony Didn’t Plan That

Summary

The video examines whether the PS5 is a practical machine for running local AI. Its conclusion is mixed: the console can generate text quickly, but access depends on homebrew software and a vulnerable firmware setup. As a result, it is not a sensible purchase solely for AI.

Why the PS5 can run language models

Text generation is often limited by how quickly a system can move model weights from memory, rather than by raw computing power. The PS5 has 16 GB of shared GDDR6 memory, rated at up to 448 GB/s, which can serve both its CPU and GPU. This high-bandwidth shared memory suits this type of workload.

Two ways to run software

The initial PS5 LM project ported llama.cpp to run as a homebrew “payload”—a program running alongside Sony’s system software. This approach has tight memory limits: although a payload can map more, only about 1.6 GB can reportedly be kept safely available for the model. A larger model running this way was reported to generate about one token per minute.

A native app, or “title,” can access substantially more direct memory and the GPU. Developer Abhiram’s work used a community RADV Vulkan driver to run the model on the graphics chip, with a little over 11 GB set aside for it.

Reported speed and trade-offs

The developers’ table reports about 20 tokens per second for a heavily compressed 27-billion-parameter Qwen model on the GPU—over 1,000 times the reported CPU-payload speed. Smaller models can be faster; the video cites more than 100 tokens per second for a much smaller Llama 3.2 model. A dashboard displayed speed, chip activity, power use, and memory use.

The 27B model must be compressed to roughly two bits per weight to fit. The video notes that this can reduce quality, and that there is no independent quality evaluation of this particular build. The reported speed of roughly 20 tokens per second is also developer-reported, not an independently verified benchmark.

Estimating performance

For a dense model, each generated token requires reading roughly the whole model file. The narrator estimates that the PS5’s memory bandwidth and the model’s size imply a theoretical ceiling of about 45 tokens per second. The measured result is lower, and the cause of the gap is unclear.

The video suggests using memory bandwidth divided by model size as a first-pass estimate when assessing other hardware, then checking whether the model fits in usable memory.

Long prompts are a bottleneck

The developers reportedly measured prompt processing at about 35–50 tokens per second. Because a long document can contain far more tokens than the generated response, processing it may take several minutes before the model produces its first word.

One cited test estimates roughly 7–9.5 minutes for a long report and about 10 minutes for a code-editor task, with most of that time spent processing the input. The developer’s proposed explanation involves RADV driver limitations and graphics-work timeouts; the video notes that this diagnosis has not been independently confirmed.

Using the PS5 as a home-network model server

The software reportedly supports request formats used by OpenAI and Anthropic APIs, so other devices on a home network could send it requests. The changelog also gives an idle power figure of about 22 W, with higher consumption during generation—an important consideration if the console is left running as a server.

Project development

Abhiram offered the GPU work back to the original project. After requesting reproducible builds, fixes for model-switching freezes, and clearer documentation, the original developer merged it in mid-October, according to the video.

Historical caution

The video compares the PS5 project with the US Air Force’s 2010 Condor cluster, built from 1,760 PS3s. The Air Force described the system as substantially cheaper than equivalent computing capability and intended it for research, including AI-related and radar-image work.

However, Sony’s 2010 removal of the PS3’s “Other OS” feature made it difficult for the Air Force to replace failed consoles. The example illustrates the risk of relying on console hardware when platform access is controlled by the manufacturer.

Bottom line

The PS5 is a striking homebrew AI demonstration, especially for people already involved in console homebrew. The narrator advises against buying or jailbreaking one solely for AI, citing firmware dependence, warranty and online-ban risks, compressed model quality, and the lack of independent performance testing.

Main speakers and sources

• Main source: An unnamed video narrator providing the analysis and performance calculations.
• Project and developer sources cited: Mert Çulhan’s PS5 LM project; Abhiram’s fork and GPU work; posts or notes attributed to Vibhav Sisanty and Neotoy.
• Other sources cited: Sony’s PS5 specifications; the US Air Force Research Laboratory’s Condor announcement; WCCF Tech.