@sudoingX This is the part I wish more build guides covered: what can you upgrade later, and what forces a whole new build?
The VRAM-per-dollar comparison should be useful.
A model fitting in GPU memory is only the starting point for sizing local AI.
Longer documents and simultaneous requests need additional KV cache capacity. Quantizing the weights frees memory, but doesn’t automatically shrink that cache.
Test the workload your team will actually run.
@sudoingX First questions would be what runs on it today, how painful deployment is, and who handles a failed unit. Purchase price is only part of what we have to build around.
@Hikari_07_jp Have you checked socket contact and cooler mounting pressure? I wouldn’t rule out the board yet.
If one or a few of the pins aren't contacted or seated properly, you can see this issue.