开发者将破解后的索尼 PS5 当作本地 AI 服务器 利用 GPU 加速运行 27B 参数 Qwen3.8-27B 模型 速度达 20.9 token/sDevelopers use the cracked Sony PS5 as a local AI server using GPU acceleration to run the 27b parameter Qwen3.8-27B model at a speed of 20.9 token/s


AI free token/AI免费积分
感谢IT之家网友 华南吴彦祖、前Apple使用者 的线索投递!

IT之家 10 月 17 日消息,据外媒 Wccftech 报道,开发者 abhi245y 基于资深 AI 工程师 Mert Cobanov 的 PS5LM(llama.cpp 移植项目),开发了原生 GPU 应用 PS5LM Plus,让经过破解的 PS5 主机可以在本地运行 Qwen3.8-27B 等大型语言模型。

IT之家获悉,PS5LM 的早期版本主要依靠 CPU 运行模型,执行 27B 参数模型时,平均每分钟只能生成约 1 个 token。为提高表现,开发者 abhi245y  移植了 Mesa 开源项目中的 AMD RADV 驱动,并借助 Vulkan 计算接口让 llama.cpp 可以作为原生应用在 PS5 上运行。这项改动将可用内存从约 6GB 提高至 11.4GB,使主机可以运行经过高度量化的 Qwen3.8-27B 模型。

开发者将破解后的索尼 PS5 当作本地 AI 服务器 利用 GPU 加速运行 27B 参数 Qwen3.8-27B 模型 速度达 20.9 token/s

开发者将破解后的索尼 PS5 当作本地 AI 服务器 利用 GPU 加速运行 27B 参数 Qwen3.8-27B 模型 速度达 20.9 token/s

得益于 PS5 较高的内存带宽(448GB/s),Qwen3.8-27B 的生成速度可达每秒 20.9 个 token。测试期间,GPU 使用率达到 100%,CPU 使用率仅为 2%。对于规模更小的模型,PS5 还能达成更高的生成速度,例如 Gemma 4 E4B 约为每秒 61 个 token,Llama 3.2 3B 则达到每秒 124 个 token。

除了利用 PS5 的 GPU 加速模型推理,PS5LM Plus 还内置网页聊天界面,并可借助局域网 8081 端口给出 API 服务,使用者可借助同一网络中的其他设备访问页面,与本地运行的模型对话。该项目还兼容 OpenAI 和 Anthropic API,方便 opencode、Codex、Claude Code 等工具调用 PS5 上运行的模型。

对于一台主要面向游戏外观的主机而言,PS5 可以充当本地 AI 服务器,为大模型推理给出 GPU 加速,展示了其硬件在游戏之外的应用潜力。不过,该机配备的 16GB RAM 统一内存仍然限制了模型的挑选和运行质量。另外,运行原生版 llama.cpp 需要先对 PS5 开展越狱,可能会导致游戏机及 PSN 账号被封禁等风险。

本文来源:https://www.ithome.com/1/014/372.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

IT House October 17 news, according to foreign media Wccftech reported that the developer abhi245y based on senior AI engineer Mert Cobanov's PS5LM (llama.cpp porting project), developed a native GPU application PS5LM Plus, so that the cracked PS5 host can run Qwen3.8-27B and other large language models locally.

IT House learned that the earlier version of PS5LM mainly relied on the CPU to run the model. When executing the 27b parameter model, only about 1 token can be generated per minute on average. In order to improve performance, the developer abhi245y ported the AMD RADV driver in the Mesa open source project, and made llama.cpp run on PS5 as a native application with the help of the Vulkan computing interface. This change increases the available memory from about 6GB to 11.4GB, allowing the host to run the highly quantified Qwen 3.8-27B model.

Thanks to the PS5's high memory bandwidth (448GB/s), the Qwen3.8-27B can generate up to 20.9 tokens per second. During the test period, the GPU usage reached 100% and the CPU usage was only 2%. For smaller models, the PS5 can also achieve higher generation speeds, for example, Gemma 4 E4B is about 61 tokens per second, and Llama 3.2 3B reaches 124 tokens per second.

In addition to using the PS5's GPU-accelerated model inference, the PS5LM Plus also has a built-in web chat interface and can give API services through the LAN port 8081. Users can use other devices in the same network to access the page and talk to the locally running model. The project is also compatible with OpenAI and Anthropic APIs, making it easy for tools such as opencode, Codex, and Claude Code to call models running on PS5.

For a console that is primarily game-oriented, the PS5 can act as a local AI server, giving GPU acceleration to large model inference and demonstrating the potential of its hardware for applications outside the game. However, the machine's 16GB RAM unified memory still limits the selection of models and the quality of operation. In addition, running the native version of llama.cpp requires jailbreak on PS5, which may lead to gaming consoles and PSN accounts being banned.

This article is from: https://www.ithome.com/1/014/372.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry