Yes, a quick back of the envelope math is 0.65 * (memory bandwidth of the card / (model weights in bytes + kv cache in bytes) ~ practical decode tps. Below context around 32k (depends upon the model but again can be used as a placeholder number) you can ignore the kv cache in bytes and the math becomes just about memory bandwidth and model weights in bytes.
Comments
Yes, a quick back of the envelope math is 0.65 * (memory bandwidth of the card / (model weights in bytes + kv cache in bytes) ~ practical decode tps. Below context around 32k (depends upon the model but again can be used as a placeholder number) you can ignore the kv cache in bytes and the math becomes just about memory bandwidth and model weights in bytes.