B200 vs B100 vs H200 vs H100 vs A100 SXM

Before we get into the details, it is worth noting that A100 is EOL, and the H100 has been supplanted by the H200. Existing stock of H100 may still be available; the H200 will supersede any new orders. The NVIDIA B300 and B200 are now available in 2025. For those who need FP64, B200. For those who need more memory and increased FP4 Dense, choose the NVIDIA B300.

Although some GPUs are EOL, for those who have existing deployments, we include the legacy NVIDIA A100 and NVIDIA H100 for easy comparison.

Architecture	Blackwell Ultra	Blackwell	Hopper	Hopper	Ampere
GPU Name	NVIDIA B300	NVIDIA B200	NVIDIA H200	NVIDIA H100	NVIDIA A100
FP64	1.2 teraFLOPS	37 teraFLOPS	34 teraFLOPS	34 teraFLOPS	9.7 teraFLOPS
FP64 Tensor Core	1.2 teraFLOPS	37 teraFLOPS	67 teraFLOPS	67 teraFLOPS	19.5 teraFLOPS
FP32	75 teraFLOPS	75 teraFLOPS	67 teraFLOPS	67 teraFLOPS	19.5 teraFLOPS
FP32 Tensor Core	2.2 petaFLOPS	2.2 petaFLOPS	989 teraFLOPS	989 teraFLOPS	312 teraFLOPS
FP16/BF16 Tensor Core	4.5 petaFLOPS	4.5 petaFLOPS	2 peraFLOPS	2 peraFLOPS	624 teraFLOPS
INT8 Tensor Core	9 petaOPs	9 petaOPs	4 petaOPs	4 petaOPs	1248 teraOPs
FP8/FP6 Tensor Core	9 petaFLOPS	9 petaFLOPS	4 petaFLOPS	4 petaFLOPS	-
FP4 Tensor Core (Sparse \| Dense)	18 \| 14 petaFLOPS	18 \| 9 petaFLOPS	-	-	-
GPU Memory	270GB HBM3e	192GB HBM3e	141GB HBM3e	80GB HBM3	80GB HBM2e
NVIDIA HGX 8x GPU Memory	2.1TB Total	1.4TB Total	1.1TB Total	640GB Total	640GB Total
Memory Bandwidth	Up to 7.7TB/s	Up to 7.7TB/s	4.8TB/s	3.2TB/s	2TB/s
Decoders	7 NVDEC 7 JPEG	7 NVDEC 7 JPEG	7 NVDEC 7 JPEG	7 NVDEC 7 JPEG	5 NVDEC 5 JPEG
Multi-Instance GPUs	Up to 7 MIGs @23GB	Up to 7 MIGs @23GB	Up to 7 MIGs @16.5GB	Up to 7 MIGs @16.5GB	Up to 7 MIGs @ 10GB
Interconnect	NVLink 1.8TB/s	NVLink 1.8TB/s	NVLink 900GB/s	NVLink 900GB/s	NVLink 600GB/s
NVIDIA AI Enterprise	Yes	Yes	Yes	Yes	EOL

Looking at the raw performance defined by the number of floating-point operations performed per second at a given precision, the NVIDIA Blackwell GPUs sacrifice FP64 Tensor Core performance in favor of heavily increased Tensor Core performance in FP32 and below. The NVIDIA B300 doubles down and has extremely low FP64 performance in favor for additional FP4 performance.

Training AI doesn’t require the high 64-bit precision for calculating weights and parameters. By sacrificing FP64 Tensor Core performance, NVIDIA squeezes more juice out when calculating on the more standard 32-bit, 16-bit, FP8/FP6, and FP4 precision formats.

The NVIDIA B300 and B200 throughput is over 2x in TF32, FP16, and FP8 than the last generation H200. It also features a new transformer engine that supports FP4. These lower-precision floating-point formats won’t be used in entire calculations, but, when incorporated into Mixed Precision workloads, the performance gains realized are unparalleled.

iStarUSA TC-2U70PD8UC 700 W PSU - 2U - 80+ - High Efficiency Switching PSU