I'm in the process of building a powerful PC with a budget of around $5,000 to $6,000, although I'm open to spending a bit more if it leads to better performance and longevity. This computer will primarily be used for local AI generation tasks including creating images, videos, text, and engaging in complex video workflows for a marketing brand. Therefore, I need reliable and fast AI processing.
I've done some research and noticed a significant difference between consumer gaming hardware and professional components suited for AI and machine learning, particularly regarding VRAM, tensor capabilities, and supporting multiple models simultaneously.
Here are my main points of concern:
1. **GPU Options** - What should I focus on? From my understanding, the RTX 5090 with 32GB VRAM seems like a good choice for intensive local generation. However, I'm also considering higher-end workstation cards like the RTX 6000 Ada with 48GB VRAM or RTX PRO 6000, though I'm not certain they justify the price for my use case. Are consumer GPUs like the 5090 reliable for long-term serious work in AI video generation, or should I opt for professional-grade options if possible? Would it be beneficial to set up multiple GPUs or start with one single powerful unit?
2. **CPU/RAM/Storage Needs** - Beyond the GPU, how crucial is a high-core-count CPU (like Ryzen 9 or Threadripper) for tasks like preprocessing and video batching? Is investing in 128GB+ of DDR5 RAM necessary for future-proofing? And is a fast NVMe storage solution of 4TB or more ideal for managing models and project files?
3. **Multi-GPU vs Single GPU** - I've seen various configurations ranging from two consumer GPUs to a single high-end GPU or even server-class accelerators like A100/H100. Given my budget, I'm curious about the feasibility of multi-GPU setups relating to power, cooling, and scaling benefits.
4. **Use Cases** - My intention is to run large local LLMs, engage in image and AI video generation/editing, and develop long-form video workflows while managing multiple models or processes for marketing content. If anyone has experience with similar setups, I'd love to hear about performance issues or improvements you would suggest.
In terms of budget allocation:
- GPU: ~50-70%
- CPU + RAM: ~20-30%
- Storage, PSU, cooling: ~10-20%
- Case and airflow are also high priorities.
I'm willing to invest more upfront to avoid a major upgrade soon, but I want to ensure I'm not overspending on unnecessary features.
3 Answers
I see you're focused on large local models. Definitely consider getting a system with unified memory like the AMD AI 395 or an Apple Studio Mac. The 32GB VRAM of the RTX 5090 might be limiting for larger models like what you mentioned.
Looking at your needs, the Intel B60 cards might be something to consider too. They cost around $600 each and have 24GB VRAM. You could set up a few of those, perhaps even five, for the price of one RTX 5090. It's worth checking out if you're balancing cost and performance!
If you plan to run heavy AI workloads, I recommend investing in an RTX PRO 6000. It provides the necessary VRAM for large models and can handle multiple tasks well. You might consider 128GB or even 256GB of RAM, especially for models that offload some tasks to system memory. If your work is document-heavy, you’ll benefit from an octa-channel RAM setup with EPYC systems. Just keep in mind that multi-GPU setups aren't as effective for AI video tasks compared to using a single high-end GPU.
Exactly! I was looking into EPYC systems too. They can really speed up processing for complex tasks. Plus, you're right about multi-GPU setups; they might not always provide the benefits you expect for AI workloads.
Agreed! The unified memory approach can make a significant difference in performance and efficiency, especially with larger AI frameworks.