{"id":28414,"date":"2025-02-10T12:31:13","date_gmt":"2025-02-10T04:31:13","guid":{"rendered":"https:\/\/www.1ai.net\/?p=28414"},"modified":"2025-02-10T12:35:40","modified_gmt":"2025-02-10T04:35:40","slug":"%e5%9c%a8%e6%9c%ac%e5%9c%b0%e8%bf%90%e8%a1%8c-deepseek-r1-%e7%9a%84%e6%88%90%e6%9c%ac%e6%98%af%e5%a4%9a%e5%b0%91%ef%bc%9fdeepseek-r1%e6%9c%ac%e5%9c%b0%e8%bf%90%e8%a1%8c%e6%88%90%e6%9c%ac","status":"publish","type":"post","link":"https:\/\/www.1ai.net\/en\/28414.html","title":{"rendered":"How much does it cost to run DeepSeek-R1 locally?DeepSeek-R1 Local Running Costs"},"content":{"rendered":"<p><a href=\"https:\/\/www.1ai.net\/en\/tag\/deepseek\" title=\"[View articles tagged with [DeepSeek]]\" target=\"_blank\" >DeepSeek<\/a> has taken the race to generate models to another level.<\/p>\n<p>Now there are even people ready to run models with 671B parameters locally.<\/p>\n<p>But running such a large model locally is not child's play; you'll need to make significant hardware upgrades just to perform inference operations.<\/p>\n<p>Here's a rough breakdown of what it would cost to run DeepSeek-R1 on your PC.<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-28415\" title=\"786540e4j00srg9se00fjd000n000n0p\" src=\"https:\/\/www.1ai.net\/wp-content\/uploads\/2025\/02\/786540e4j00srg9se00fjd000n000n0p.jpg\" alt=\"786540e4j00srg9se00fjd000n000n0p\" width=\"828\" height=\"828\" \/><\/p>\n<p><strong>hardware cost<\/strong><\/p>\n<p>Most of the cost is spent on hardware. We will discuss GPUs, CPUs, memory (RAM), SSD storage, cooling systems, and more.<\/p>\n<p><strong>The configuration you need is as follows (at today's exchange rate 1 USD = 7.29 CNY):<\/strong><\/p>\n<p>GPU<\/p>\n<ul>\n<li><strong>4 NVIDIA H100 80GB GPUs<\/strong>(Each $25,000)<\/li>\n<li><strong>Total cost:<\/strong>US$ 100,000 (approximately $729,000)<\/li>\n<li><strong>reason:<\/strong>These GPUs are state-of-the-art gas pedals optimized for AI workloads, enabling faster training and inference for large models like DeepSeek-R1.<\/li>\n<\/ul>\n<blockquote>\n<ul>\n<li><strong>NVIDIA H100:<\/strong>The NVIDIA H100 is an advanced GPU based on the Hopper architecture with a fourth-generation tensor core and Transformer engine that enables up to 9x faster AI training and 30x faster inference than the previous A100 GPU.<\/li>\n<\/ul>\n<\/blockquote>\n<p>CPU<\/p>\n<ul>\n<li><strong>Intel Xeon Platinum<\/strong><\/li>\n<li><strong>Total cost:<\/strong>\u00a0US$ 1,550 (approximately $11,299.50)<\/li>\n<li><strong>reason:<\/strong>High-end CPUs ensure smooth multitasking and system stability during resource-intensive operations.<\/li>\n<\/ul>\n<blockquote>\n<ul>\n<li>The Intel Xeon Platinum is essential for DeepSeek-R1 inference because of its advanced AI acceleration features, such as Intel AMX and AVX-512, which dramatically improve the performance of deep learning tasks.<\/li>\n<li>It delivers up to 42% of AI inference performance boost over its predecessor, making it ideal for handling high-load jobs. In addition, its optimized memory and interconnect technologies ensure efficient processing of large datasets and complex models.<\/li>\n<\/ul>\n<\/blockquote>\n<p><strong>Memory (RAM)<\/strong><\/p>\n<ul>\n<li>512GB DDR4 ($6,399.98)<\/li>\n<li>Total Cost: $6,399.98 (approximately $46,655.85)<\/li>\n<li><strong>reason:<\/strong>\u00a0Large memory capacity is critical for processing massive datasets and model parameters without performance bottlenecks.<\/li>\n<\/ul>\n<p><strong>Storage<\/strong><\/p>\n<ul>\n<li>4TB NVMe SSD ($249.99)<\/li>\n<li>Total Cost: $249.99 (approximately $1,822.43)<\/li>\n<li><strong>reason:<\/strong>\u00a0Fast storage ensures rapid reading of data during training.<\/li>\n<\/ul>\n<blockquote>\n<ul>\n<li>SSDs (Solid State Drives) are devices that utilize flash memory to store data with faster read and write speeds, higher endurance, and better energy efficiency than traditional mechanical hard drives (HDDs).<\/li>\n<li>4TB NVMe SSDs refer specifically to high-capacity (4 TB) SSDs that utilize the NVMe (Non-Volatile Memory Express) protocol, which utilizes the PCIe interface to achieve faster data transfer rates than older SATA interface-based SSDs.<\/li>\n<li>NVMe SSDs are particularly well suited for tasks where speed and high-capacity storage are critical, such as gaming, video editing, or server applications.<\/li>\n<\/ul>\n<\/blockquote>\n<p><strong>Power Supply Unit (PSU)<\/strong><\/p>\n<ul>\n<li>2000W PSU ($259.99)<\/li>\n<li>Total Cost: $259.99 (approximately $1,895.33)<\/li>\n<li><strong>reason:<\/strong>\u00a0High-wattage power supplies can reliably provide stable power to multiple GPUs.<\/li>\n<\/ul>\n<p><strong>Cooling System<\/strong><\/p>\n<ul>\n<li>Customized Liquid Cooling Circulation ($500)<\/li>\n<li>Total Cost: $500 (approximately $3,645.00)<\/li>\n<li><strong>reason:<\/strong>\u00a0The GPU generates a lot of heat while running, and the liquid cooling system helps prevent overheating.<\/li>\n<\/ul>\n<p><strong>Motherboard<\/strong><\/p>\n<ul>\n<li>ASUS S14NA-U12 ($500)<\/li>\n<li>Total Cost: $500 (approximately $3,645.00)<\/li>\n<li><strong>reason:<\/strong>\u00a0Supports dual-socket GPUs and high-end CPUs.<\/li>\n<\/ul>\n<p><strong>Chassis<\/strong><\/p>\n<ul>\n<li>Cooler Master Cosmos C700M ($482)<\/li>\n<li>Total cost: $482 (approximately $3,513.78)<\/li>\n<li><strong>reason:<\/strong>\u00a0The spacious chassis can accommodate custom cooling systems and multiple GPUs.<\/li>\n<\/ul>\n<blockquote>\n<ul>\n<li><strong>Total hardware cost: approximately $109,941 (approximately $801,469.89)<\/strong><\/li>\n<\/ul>\n<\/blockquote>\n<p><strong>Software Costs<\/strong><\/p>\n<p>The software required to run DeepSeek-R1 is all free, but you will need to have the following components on hand:<\/p>\n<ul>\n<li><strong>Operating System:<\/strong>\u00a0Debian Linux (free)<\/li>\n<li><strong>Programming Languages:<\/strong>\u00a0Python 3.10+ (free)<\/li>\n<li><strong>DeepSeek-R model:<\/strong>\u00a070B Parametric modeling (free)<\/li>\n<li><strong>CUDA Toolkit &amp; cuDNN:<\/strong>\u00a0NVIDIA's deep learning library (free)<\/li>\n<li><strong>Deep Learning Framework:<\/strong>\u00a0PyTorch with CUDA support (free)<\/li>\n<\/ul>\n<blockquote>\n<ul>\n<li><strong>Software totals:<\/strong>\u00a00 dollars<\/li>\n<\/ul>\n<\/blockquote>\n<p><strong>Key Takeaways<\/strong><\/p>\n<ul>\n<li><strong>Hardware costs dominate:<\/strong>\u00a0The GPU, memory and cooling system account for approximately 99% of the total cost.<\/li>\n<li><strong>Expertise is required:<\/strong>\u00a0Building such a system requires knowledge of high performance computing.<\/li>\n<li><strong>Alternatives:<\/strong>\u00a0For short-term projects, cloud services (e.g. AWS, Google Cloud) may be more economical, but will incur ongoing costs.<\/li>\n<\/ul>\n<p><strong>Is it worth it?<\/strong><\/p>\n<p>For well-funded researchers, corporations, or enthusiasts with specific needs (e.g., privacy protection, offline use), building a system locally provides unmatched control and speed; for other users, using a cloud platform or choosing a smaller model may be more practical.<\/p>\n<p>but,<strong>Since the overall price is close to $110,000 (about $801,900.00)<\/strong>, for individuals, such a commitment is indeed unaffordable.<\/p>\n<p>However, you can try the distilled, more affordable model versions.<\/p>\n<p>So, are you going to run DeepSeek-R1 locally? Think again!<\/p>","protected":false},"excerpt":{"rendered":"<p>DeepSeek has raised the competition for model generation to another level. Now, there are even people who are going to run local models with 671B parameters. But it's not a joke to run a model so big locally; it's just a reasoning exercise, and you need to make a major hardware upgrade. The cost of running DeepSeek-R1 on your PC is roughly broken down below. Hardware costs, mostly on hardware. We will discuss GPU, CPU, memory (RAM), SSD storage, distillation systems, etc. The configurations you need are as follows (at today's rate of exchange of US$ 1 = RMB 7.29): GPU 4 NVIDIA H100 80GB GPU (1 TP4T25 per block)<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[144],"tags":[3606],"collection":[5669],"class_list":["post-28414","post","type-post","status-publish","format-standard","hentry","category-baike","tag-deepseek","collection-deepseek"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/28414","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/comments?post=28414"}],"version-history":[{"count":0,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/28414\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/media?parent=28414"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/categories?post=28414"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/tags?post=28414"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/collection?post=28414"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}