GPU upgrade vs full system: AI workloads demand 16GB VRAM or new AM5
Ryzen 3600 owner debates 5070 Ti vs AM5 build for image gen and LLMs
A Reddit user seeks advice on upgrading for AI workloads, specifically image and video generation (e.g., Stable Diffusion), TTS, consistent character generation, and future local LLM experiments. Their current rig: Ryzen 3600 CPU, AMD Radeon RX 5700 (8GB VRAM), 32GB DDR4 RAM, MSI B450 Gaming Plus motherboard (PCIe 3.0), and a 650W Corsair RMx PSU. They’re considering an RTX 5060 Ti 16GB (€600) or RTX 5070 Ti 16GB (€1000) but are unsure whether to drop it into the existing system or build a new AM5 platform with 64-128GB RAM, a Taichi motherboard, and a PSU supporting two GPUs (budget ~€3000).
Key technical concerns: PCIe 3.0 bandwidth may limit GPU performance for data-heavy AI tasks, but for inference it’s often less critical than VRAM capacity. CPU (Zen 2) might bottleneck in stable diffusion batch processing but not drastically. The user also considers a used RTX 3090 (€800) with 24GB VRAM for larger models. The wife’s PC (Ryzen 5600, RTX 5060 8GB, 16GB RAM, B550 PCIe 4.0) offers a potential testbed. The real question: will 16GB VRAM suffice for current image gen and character consistency, or will 32GB+ system RAM be needed for LLM inference? Given that LLMs can use CPU+RAM fallback, but performance plummets without GPU VRAM. The consensus from replies likely leans toward a GPU-only upgrade first, testing on wife’s system, before a full AM5 jump.
- Current system (Ryzen 3600, RX 5700 8GB, PCIe 3.0, 32GB RAM) may bottleneck 16GB VRAM GPUs but still usable for AI inference
- RTX 5070 Ti 16GB (€1000 vs new AM5 build €3000 with 64-128GB RAM) – VRAM is primary constraint for local image gen and LLMs
- Used RTX 3090 24GB (€800) offers more VRAM for larger models but lacks modern features like tensor cores for FP8
Why It Matters
AI enthusiasts face a common dilemma: incremental GPU upgrade vs platform overhaul – VRAM and memory often trump CPU generation gains.