blackwell datacenter parts are deliberately missing from gpus.py. the database rule is that every entry derives its peaks from physical parameters (sm count, boost clock, bus, per-pin rate) and a test asserts the result against the vendor sheet.
for b200 the public numbers don't yet pin down a consistent sm-count + boost-clock pair the way they do for the cards in the db (we already hit this with h100, where the published fp16 tensor number implies a different clock than the published fp32 number).
if you have access to a b200: kernelmeter info will print the real sm count and boost clock, which settles it. paste the output here and the entry becomes trivial.
blackwell datacenter parts are deliberately missing from gpus.py. the database rule is that every entry derives its peaks from physical parameters (sm count, boost clock, bus, per-pin rate) and a test asserts the result against the vendor sheet.
for b200 the public numbers don't yet pin down a consistent sm-count + boost-clock pair the way they do for the cards in the db (we already hit this with h100, where the published fp16 tensor number implies a different clock than the published fp32 number).
if you have access to a b200: kernelmeter info will print the real sm count and boost clock, which settles it. paste the output here and the entry becomes trivial.