gpus.py covers 20 cards but skips the mid-range ada parts people actually own: 4070, 4070 super, 4070 ti, 4070 ti super.
adding a card is one line in DATABASE (sm count, boost mhz, bus width, per-pin gbps, tdp) plus one line in tests/test_gpus.py asserting the derived bandwidth and fp32 tflops against the vendor spec sheet. the test is the point: if the entry is wrong, ci fails.
grab the numbers from the nvidia product page or techpowerup, and cross-check that derived == published before opening the pr.
gpus.py covers 20 cards but skips the mid-range ada parts people actually own: 4070, 4070 super, 4070 ti, 4070 ti super.
adding a card is one line in DATABASE (sm count, boost mhz, bus width, per-pin gbps, tdp) plus one line in tests/test_gpus.py asserting the derived bandwidth and fp32 tflops against the vendor spec sheet. the test is the point: if the entry is wrong, ci fails.
grab the numbers from the nvidia product page or techpowerup, and cross-check that derived == published before opening the pr.