NVIDIA GPU Families

Cross-reference for all data center GPU products. See Domain 1.6 for workload mapping.

Architecture Overview

GPU Architectures Overview

Blackwell / Blackwell Ultra — B200, B300

B200 / B300 Introduction

Blackwell B200/B300 Intro

B200 / B300 Specifications

Blackwell B200/B300 Specs

Spec	Value
Architecture	Blackwell / Blackwell Ultra
Transistors	208 billion
Transformer Engine	2nd generation (FP4/FP8/BF16/FP16)
NVLink	5th generation (1.8 TB/s bidirectional)
Memory	HBM3e
B200 TDP	~1,000 W (SXM form factor)
Key features	2nd-gen Transformer Engine, Secure AI, Decompression Engine

Primary workloads: Generative AI (LLM training + inference), AI reasoning, biology/drug discovery

Hopper — H100, H200

H100 Hopper

	H100 SXM5	H100 PCIe	H200 SXM5
Memory	80 GB HBM3	80 GB HBM2e	141 GB HBM3e
Memory BW	3.35 TB/s	2.0 TB/s	4.8 TB/s
Peak TF (BF16)	989 TFLOPS	756 TFLOPS	989 TFLOPS
NVLink	4th gen, 900 GB/s	4th gen	4th gen
TDP	700 W	350 W	700 W

Primary workloads: LLMs, data analytics, conversational AI, NLP, autonomous vehicles

Ada Lovelace — L40S, L40, L4

L40S (Data Center)

L40S Intro L40S Specs

Spec	L40S	L40	L4
Memory	48 GB GDDR6	48 GB GDDR6	24 GB GDDR6
Memory BW	864 GB/s	864 GB/s	300 GB/s
Tensor Core gen	4th	4th	4th
RT Core gen	3rd	3rd	3rd
TDP	350 W	300 W	72 W
Video codec	AV1 enc/dec	H.264/H.265	AV1 enc/dec

L40S use cases: Generative AI inference + graphics, mainstream data center compute, AI video
L4 use cases: VDI, edge inference, AI video at low power

Grace CPU

Spec	Value
ISA	Arm Neoverse V2 (72 cores)
Memory	LPDDR5X (proprietary subsystem design)
Interface	NVLink-C2C (to GPU in Superchip configs)
TDP	~500 W (in GB200 Superchip config)

Primary use: HPC, hyperscale cloud, building block for Grace Superchips

RTX PRO Server

Spec	RTX PRO 6000 Blackwell Server
CUDA cores	24,064
Tensor Cores	752 (5th gen)
RT Cores	188 (4th gen)
Architecture	Blackwell
Purpose	AI + professional visualization in data center