fix wrong b200 flops number (#1393)

This pr fix what seems to be a wrong estimation of the peak flops for
the B200. With the current code the peak flops of B200 is 4.5x bigger
that H100 which seems off. It seems that the number reported
https://nvdam.widen.net/s/wwnsxrhm2w/blackwell-datasheet-3384703 are
with 2:4 sparsity ?

Author: samsja

torchtitan/tools/utils.py

@@ -97,7 +97,7 @@ def get_peak_flops(device_name: str) -> int:
         return 989e12
     elif "B200" in device_name:
         # data from https://nvdam.widen.net/s/wwnsxrhm2w/blackwell-datasheet-3384703
-        return 4.5e15
+        return 2.25e15
     elif "MI300X" in device_name or "MI325X" in device_name:
         # MI300X data from https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html
         # MI325X data from https://www.amd.com/en/products/accelerators/instinct/mi300/mi325x.html