Most small AI teams standardize on one card: the RTX 3090.
Cheap, 24 GB, easy to rent.
The catch was always that NVIDIA capped its FP16 and BF16 math at roughly half the rate of the professional cards.
That cap just came off.
Zagita is an open-source project that loads a Linux kernel module before the NVIDIA driver binds the card, then raises the FP/BF16 instruction issue rate on GA102, GA103, GA104 and GA106 to the professional level.
No vBIOS flash, no hardware modding, and driver updates do not wipe it.
The measured gap it closes: an RTX 3090 Ti moves from 87.5 to 157 TFLOPS in BF16, and from 42.6 to 82.6 in TF32.
Users across the 30 series report up to 80%.
Read that as a price cut.
The same card now finishes close to double the work per hour, so a fine-tune job costs less without anyone buying new hardware.
Our RTX 3090s run at $0.15/hr, and Buy Now Pay Later keeps it at zero upfront with payment after one month.
Plenty of room to test a bigger job before you commit budget.
What would you throw at double the throughput first: a longer context evaluation pass, or another training run on the same data?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w43&utm_content=morning
Originally posted on linkedin_personal