NVIDIA B30A vs NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip 252GB
Side-by-side specifications, performance, and rental pricing for datacenter AI workloads.
B30A: Blackwell Ultra, 2025 GB300: Blackwell Ultra, 2025
NVIDIA B30A
Full specs →Architecture
Blackwell Ultra
Year
2025
NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip 252GB
Full specs →Architecture
Blackwell Ultra
Year
2025
Memory
252 GB
Specifications
B30A GB300
Architecture
Blackwell UltraBlackwell Ultra
Launch Year
20252025
Form Factor
—Superchip
Memory
—252 GB
Memory Bandwidth
—7.1 TB/s
TDP
——
Process Node
—4nm
Spec B30A GB300
Architecture Blackwell UltraBlackwell Ultra
Launch Year 20252025
Form Factor —Superchip
Memory —252 GB
Memory Bandwidth —7.1 TB/s
TDP ——
Process Node —4nm
Performance (TFLOPS)
B30A GB300
FP64
No verified data available1.3 TFLOPS
FP32
No verified data available80 TFLOPS
TF32
No verified data available1,250 TFLOPS
2,500 TFLOPS sparse
BF16
No verified data available2,500 TFLOPS
5,000 TFLOPS sparse
FP16
No verified data available2,500 TFLOPS
5,000 TFLOPS sparse
FP8
No verified data available5,000 TFLOPS
10,000 TFLOPS sparse
FP6
No verified data available5,000 TFLOPS
10,000 TFLOPS sparse
FP4
No verified data available15,000 TFLOPS
20,000 TFLOPS sparse
INT8
No verified data available165 TOPS
330 TOPS sparse
Precision B30A GB300
FP64 No verified data available1.3 TFLOPS
FP32 No verified data available80 TFLOPS
TF32 No verified data available1,250 TFLOPS
2,500 TFLOPS sparse
BF16 No verified data available2,500 TFLOPS
5,000 TFLOPS sparse
FP16 No verified data available2,500 TFLOPS
5,000 TFLOPS sparse
FP8 No verified data available5,000 TFLOPS
10,000 TFLOPS sparse
FP6 No verified data available5,000 TFLOPS
10,000 TFLOPS sparse
FP4 No verified data available15,000 TFLOPS
20,000 TFLOPS sparse
INT8 No verified data available165 TOPS
330 TOPS sparse
FLOPS by Precision
What actually differs
Both GPUs launched in 2025: the B30A on Blackwell Ultra and the GB300 on Blackwell Ultra.
For LLM training and inference, weight the FP8 and FP16 rows and memory capacity most heavily. For scientific and HPC workloads, the FP64 row is the one to read.
Get Comparison Updates
New GPUs added weekly. Be the first to see how they compare.