GB300 SXM6 vs RTX 3090

The GB300 SXM6 delivers 2250 TFLOPS FP16 dense performance against the RTX 3090 35.6 TFLOPS FP16 figure, producing a ratio that supports substantially larger batch sizes during training. Memory bandwidth of 8000 GB/s on the GB300 SXM6 versus 936 GB/s on the RTX 3090 further enables sustained throughput when model parameters exceed the 24 GB limit of the RTX 3090. FP32 performance stands at 75 TFLOPS for the GB300 SXM6 family versus 35.6 TFLOPS for the RTX 3090 family. When sparsity is applied the GB300 SXM6 reaches 4500 TFLOPS FP16 with sparsity while the RTX 3090 figure remains without a sparsity designation and must be compared only to dense values.

The GB300 SXM6 variant suits training runs that require 288 GB HBM3e capacity and 8000 GB/s bandwidth to accommodate models that exceed 24 GB. Its 2250 TFLOPS FP16 dense rating and 1400 W TDP support sustained operation on large scale scientific computing tasks where the SXM6 form factor and InfiniBand interconnect maintain data movement efficiency.

The RTX 3090 variant fits inference workloads that remain within 24 GB GDDR6X and operate under a 350 W TDP envelope. Its PCIe form factor and 35.6 TFLOPS FP16 rating allow deployment in standard server slots where the lower power draw and NVLink interconnect suffice for moderate batch sizes.