Reference · NVIDIA Line Card 2025

The NVIDIA Data Center GPU
comparison guide

Every shipping NVIDIA accelerator at a glance — memory, bandwidth, interconnect, power, and which workloads each one is built for. Sourced from the official NVIDIA Data Center Platform line card and our own deployment experience.

Section 01

Data Center Compute Portfolio

Built for pretraining, post-training, and large-scale inference. NVLink/NVSwitch-class systems for the biggest models on earth.

Price-performance is relative within each column
SolutionNetworkingPretrainingPost-TrainingInference / Test-TimeData AnalyticsHPC / ScientificAI Video Generation
GB300 NVL72
Blackwell · Rack-scale (72 GPU, liquid)
Quantum-X800 InfiniBand / Spectrum-X
GB200 NVL72
Blackwell · Rack-scale (72 GPU, liquid)
Quantum-X800 InfiniBand / Spectrum-X
GB200 NVL4
Blackwell · 4-GPU Grace+Blackwell module
Quantum-2 InfiniBand / Spectrum-X
GH200 NVL2 / NVL4
Hopper · 2 or 4-GPU Grace+Hopper
Quantum-2 InfiniBand / Spectrum-X
HGX B300 (8-GPU)
Blackwell · 8-GPU HGX baseboard
Quantum-X800 InfiniBand / Spectrum-X
HGX B200 (8-GPU)
Blackwell · 8-GPU HGX baseboard
Quantum-X800 InfiniBand / Spectrum-X
HGX H200 (8-GPU)
Hopper · 8-GPU HGX baseboard
Quantum-2 InfiniBand / Spectrum-X
H200 NVL (PCIe)
Hopper · PCIe dual-slot (NVLink bridge)
Quantum-2 InfiniBand / Spectrum-X
Section 02

Universal Graphics & Compute Portfolio

Versatile PCIe accelerators that span training, inference, rendering, Omniverse, and virtualization in a single SKU.

SolutionNetworkingPretrainingPost-TrainingInference / Test-TimeData AnalyticsHPC / ScientificOmniverse / RenderVirtualization (vWS/VDI)Far Edge / AI Video
RTX PRO 6000 Blackwell (Server Edition)
Blackwell · PCIe dual-slot
Quantum-2 InfiniBand / Spectrum-X
L40S
Ada Lovelace · PCIe dual-slot
Quantum-2 InfiniBand / Spectrum-3
L4
Ada Lovelace · PCIe single-slot, low-profile
Quantum-2 InfiniBand / Spectrum-3

vGPU is supported across all products shown in this table.

Section 03

Specifications at a glance

Memory, bandwidth, interconnect, and power for every GPU above.

GB300 NVL72
Blackwell
~120 kW / rack

Flagship rack for frontier reasoning & trillion-parameter training.

Memory
21 TB HBM3e (rack)
Bandwidth
Up to 576 TB/s aggregate
Interconnect
5th-Gen NVLink + NVL72
Form factor
Rack-scale (72 GPU, liquid)
GB200 NVL72
Blackwell
~120 kW / rack

Production rack for LLM training and real-time trillion-parameter inference.

Memory
13.5 TB HBM3e (rack)
Bandwidth
576 TB/s NVLink
Interconnect
5th-Gen NVLink + NVL72
Form factor
Rack-scale (72 GPU, liquid)
GB200 NVL4
Blackwell
~5.4 kW

Right-sized Grace+Blackwell node for mid-scale training & HPC.

Memory
768 GB HBM3e
Bandwidth
32 TB/s NVLink (module)
Interconnect
5th-Gen NVLink (NVL4)
Form factor
4-GPU Grace+Blackwell module
GH200 NVL2 / NVL4
Hopper
~2.4–4.8 kW

Coherent CPU+GPU memory for HPC, RAG, and graph workloads.

Memory
Up to 480 GB LPDDR5X + 282 GB HBM3e
Bandwidth
900 GB/s NVLink-C2C
Interconnect
NVLink + NVLink-C2C
Form factor
2 or 4-GPU Grace+Hopper
HGX B300 (8-GPU)
Blackwell
~14 kW (chassis)

Ultra-class Blackwell baseboard for the largest single-node training jobs.

Memory
2.3 TB HBM3e
Bandwidth
64 TB/s aggregate HBM
Interconnect
5th-Gen NVLink + NVSwitch
Form factor
8-GPU HGX baseboard
HGX B200 (8-GPU)
Blackwell
~10 kW (chassis)

The new training workhorse — 3× training, 15× inference vs H100 generation.

Memory
1.5 TB HBM3e (180 GB × 8)
Bandwidth
64 TB/s aggregate HBM
Interconnect
5th-Gen NVLink + NVSwitch
Form factor
8-GPU HGX baseboard
HGX H200 (8-GPU)
Hopper
~6.5 kW (chassis)

Hopper refresh: 1.8× LLM inference vs H100 thanks to HBM3e.

Memory
1.1 TB HBM3e (141 GB × 8)
Bandwidth
38 TB/s aggregate HBM
Interconnect
4th-Gen NVLink + NVSwitch
Form factor
8-GPU HGX baseboard
H200 NVL (PCIe)
Hopper
600 W

PCIe Hopper with HBM3e — best PCIe option for enterprise LLM inference.

Memory
141 GB HBM3e
Bandwidth
4.8 TB/s
Interconnect
NVLink bridge (up to 4-way)
Form factor
PCIe dual-slot (NVLink bridge)
RTX PRO 6000 Blackwell (Server Edition)
Blackwell
600 W

Universal Blackwell card: training, inference, rendering, VDI — one SKU.

Memory
96 GB GDDR7 ECC
Bandwidth
1.6 TB/s
Interconnect
PCIe Gen5
Form factor
PCIe dual-slot
L40S
Ada Lovelace
350 W

Best $/perf for generative inference, fine-tuning, and Omniverse rendering.

Memory
48 GB GDDR6 ECC
Bandwidth
864 GB/s
Interconnect
PCIe Gen4
Form factor
PCIe dual-slot
L4
Ada Lovelace
72 W

Low-power universal accelerator for video, edge inference, and VDI.

Memory
24 GB GDDR6
Bandwidth
300 GB/s
Interconnect
PCIe Gen4
Form factor
PCIe single-slot, low-profile
Section 04

Networking glossary

QTM-X800

NVIDIA Quantum-X800 InfiniBand switch + BlueField-3 DPUs and ConnectX-8 SuperNICs. The fabric for Blackwell-class clusters.

QTM-2

NVIDIA Quantum-2 InfiniBand switch + BlueField-3 DPUs and ConnectX-7 NICs. The mainstream 400G AI fabric.

Spectrum-X

NVIDIA Spectrum-X Ethernet + SuperNICs + BlueField-3 DPUs. Ethernet built for AI workloads.

Spectrum-3

NVIDIA Spectrum-3 Ethernet + ConnectX-6 NICs + BlueField-2 DPUs. Cost-effective fabric for inference and Omniverse.

5th-Gen NVLink

1.8 TB/s per-GPU NVLink in Blackwell. Up to 72 GPUs in a single NVL72 domain.

NVLink-C2C

900 GB/s coherent CPU↔GPU link in Grace+Hopper / Grace+Blackwell superchips.

Section 05

How to choose

Frontier LLM training (>100B params)

GB300 NVL72 or GB200 NVL72 rack-scale. Liquid-cooled, 5th-Gen NVLink across 72 GPUs, Quantum-X800 fabric.

Production LLM training (7B–70B)

HGX B200 or HGX H200 8-GPU. Our Scalar L40S is our most-rented inference and fine-tuning platform.

Fine-tuning & RAG inference

H200 NVL (PCIe) or RTX PRO 6000 Blackwell. 96–141 GB per card with NVLink bridging.

Generative inference at $/token scale

L40S or RTX PRO 6000 Blackwell in dense PCIe servers. Pair with Spectrum-X.

Omniverse, USD, rendering

RTX PRO 6000 Blackwell or L40S — the only cards with full RT cores + DLSS in the data center.

Edge AI video & VDI

L4 in 1U servers — 72 W, single-slot, NVENC × 4 per GPU.

Ready to build

Configure any of these in a SuperMicro chassis.

Pick a platform, dial in GPU count, CPU, memory, storage, and fabric. We quote every custom order within one business day.

Start a custom build