Planning a 70B Build in BuildSmith: Start With VRAM, Not Parts
Every BuildSmith plan starts with the same question: what do you want to run? The app takes the model first, because the model decides the memory, and the memory decides almost everything else in the build. This walkthrough plans the build readers ask about most: a home machine for a 70B model.
The math is printed in the open. A 70B model needs its parameter count times the bytes per parameter, plus about 20 percent for the KV cache and runtime. At FP16 that is roughly 168 GB, beyond any single card. At INT8 it is about 84 GB, workstation card territory. At INT4 it lands near 42 GB, which is why the classic home answer is a 24 GB card with system RAM offload, or a 48 GB and up configuration for full residency. BuildSmith shows all three lines before it shows a single part, so the trade is yours, not the app's.
The memory lines BuildSmith draws
| Model | Params | FP16 needs | INT8 needs | INT4 needs |
|---|---|---|---|---|
| Llama 3.1 8B | 8B | 19.2 GB | 9.6 GB | 4.8 GB |
| Qwen 2.5 32B | 32B | 76.8 GB | 38.4 GB | 19.2 GB |
| Llama 3.1 70B | 70B | 168 GB | 84 GB | 42 GB |
| Llama 3.1 405B | 405B | 972 GB | 486 GB | 243 GB |
VRAM needed at INT4, by model
From the memory line, the plan assembles itself: GPU first, then the CPU tier that will not bottleneck it, the RAM that covers offload headroom, and the storage that holds the weights. Each part comes from the live directory data, with its source named, and parts without a published price are labeled as such instead of being quietly estimated. The result for the 70B walkthrough is close to our hand built 70B inference build, which is the point: the app and the editors are reading the same numbers.
One refinement worth knowing: the 20 percent overhead allowance is a planning figure for typical chat context lengths. Long context work, 128K tokens and up, grows the KV cache well past that allowance, and BuildSmith asks for your expected context before it finalizes the memory line. At very long contexts the app will push a 70B plan from the 24 GB tier toward workstation memory or a smaller model, and it shows the cache math that forced the move rather than quietly changing the recommendation. That transparency is the design rule for the whole app: every number on the plan can be traced to a formula or a dataset row.
BuildSmith is in development now. Join the waitlist from the homepage and the launch note goes to your inbox, or try the same math yourself in the VRAM Fit Checker.