87 GPUsUpdated July 2026

GPU Specs Database — H100, A100, RTX 4090 & All AI Cloud GPU Specs

Complete technical specifications for every major AI and cloud GPU — NVIDIA Hopper, Blackwell, Ampere, Ada Lovelace, AMD CDNA 3, and more. Click any GPU to see live cloud pricing across 94+ providers.

Total GPUs

Generations

Max FP16

7.5kT

Max VRAM

288GB

87 results

Sort by:

NVIDIABlackwell Ultra2026

GB300 NVL72

FP16 Performance7.5k TFLOPS

Next-gen frontier trainingTrillion-param inference+1

NVIDIABlackwell2025

GB200 NVL72

FP16 Performance4.5k TFLOPS

Frontier LLM trainingTrillion-param models+1

NVIDIABlackwell2025

B200 192GB

FP16 Performance3.5k TFLOPS

Next-gen LLM trainingUltra-large model inference+1

NVIDIABlackwell2025

B200 SXM 192GB

FP16 Performance3.5k TFLOPS

Next-gen LLM trainingUltra-large model inference+1

NVIDIABlackwell2025

B200 NVL 192GB

FP16 Performance3.5k TFLOPS

Large-scale inferenceMulti-GPU NVLink clusters+1

NVIDIABlackwell Ultra2025

B300 SXM6

FP16 Performance2.8k TFLOPS

Frontier LLM trainingMassive multi-GPU clusters+1

NVIDIABlackwell Ultra2025

GB300 SXM6

FP16 Performance2.8k TFLOPS

Frontier AI trainingInference at scale+1

NVIDIABlackwell Ultra2025

B300 SXM 262GB

FP16 Performance2.8k TFLOPS

Frontier LLM trainingMassive multi-GPU clusters+1

NVIDIABlackwell Ultra2025

GB300 SXM 288GB

FP16 Performance2.8k TFLOPS

Frontier AI trainingInference at scale+1

NVIDIAHopper2024

H200 141GB

FP16 Performance2.0k TFLOPS

70B+ model inferenceMemory-bound LLM serving+1

NVIDIAHopper2022

H100 80GB

FP16 Performance2.0k TFLOPS

LLM trainingLarge-scale inference+1

NVIDIAHopper2022

H100 SXM5 80GB

FP16 Performance2.0k TFLOPS

LLM trainingLarge-scale inference+1

NVIDIAHopper2023

H100 NVL 94GB

FP16 Performance2.0k TFLOPS

Large context LLM inferenceMemory-bound workloads+1

NVIDIAHopper2024

H200 SXM 141GB

FP16 Performance2.0k TFLOPS

70B+ model inferenceMemory-bound LLM serving+1

NVIDIAHopper2024

H200 NVL 141GB

FP16 Performance2.0k TFLOPS

Large model inferenceMulti-GPU NVLink clusters+1

NVIDIAHopper2023

GH200 Grace Hopper 96GB

FP16 Performance2.0k TFLOPS

Memory-bound LLM inferenceLarge context windows+1

NVIDIAHopper2023

H100 40GB

FP16 Performance1.5k TFLOPS

Mid-size LLM inferenceCost-efficient training+1

NVIDIAHopper2022

H100 PCIe 80GB

FP16 Performance1.5k TFLOPS

Inference servingCost-efficient training+1

AMDCDNA 3.52024

MI325X 256GB

FP16 Performance1.5k TFLOPS

Massive context LLM inferenceMemory-bound workloads+1

AMDCDNA 3.52024

MI325X 256GB

FP16 Performance1.5k TFLOPS

Massive context LLM inferenceMemory-bound workloads+1

AMDCDNA 32023

MI300X 192GB

FP16 Performance1.3k TFLOPS

Memory-bound LLM inferenceLarge context windows+1

AMDCDNA 32023

MI300X 192GB

FP16 Performance1.3k TFLOPS

Memory-bound LLM inferenceLarge context windows+1

AMDCDNA 42025

MI355X

FP16 Performance1.3k TFLOPS

Large-scale LLM trainingMemory-bound inference+1

AMDCDNA 42025

MI355X 288GB

FP16 Performance1.3k TFLOPS

Large-scale LLM trainingMemory-bound inference+1

NVIDIAGaudi 22022

Gaudi 2 96GB

FP16 Performance432 TFLOPS

Cost-effective LLM trainingTransformer workloads+1

NVIDIAGaudi 22022

Gaudi 2

FP16 Performance432 TFLOPS

Cost-effective LLM trainingTransformer workloads+1

NVIDIABlackwell2025

RTX 5090

FP16 Performance419.6 TFLOPS

Consumer AI inferenceFine-tuning up to 30B+1

AMDCDNA 22021

MI250X 128GB

FP16 Performance383 TFLOPS

Large model trainingMemory-bound HPC+1

NVIDIAAmpere2020

A100 80GB

FP16 Performance312 TFLOPS

ML trainingLarge model inference+1

NVIDIAAmpere2020

A100 40GB

FP16 Performance312 TFLOPS

ML trainingMid-size model inference+1

NVIDIAAmpere2020

A100 SXM4 80GB

FP16 Performance312 TFLOPS

ML trainingLarge model inference+1

NVIDIAAmpere2021

A100 PCIe 80GB

FP16 Performance312 TFLOPS

Inference servingCost-efficient training+1

NVIDIAAmpere2020

A100 PCIe 40GB

FP16 Performance312 TFLOPS

Mid-size model inferenceCost-efficient training+1

NVIDIAAmpere2020

A100 SXM4 40GB

FP16 Performance312 TFLOPS

ML trainingMulti-GPU NVLink setups+1

NVIDIABlackwell2025

RTX 5080

FP16 Performance275.4 TFLOPS

Budget consumer inferenceSmall model fine-tuning+1

NVIDIABlackwell2025

RTX PRO 6000 Blackwell

FP16 Performance250 TFLOPS

Professional AI workloadsLarge model inference+1

NVIDIABlackwell2025

RTX 5070 Ti

FP16 Performance215.2 TFLOPS

Budget inferenceExperimentation+1

NVIDIAAda Lovelace2023

L40S

FP16 Performance183 TFLOPS

Inference servingGenerative AI+1

NVIDIAAda Lovelace2022

RTX 6000 Ada

FP16 Performance182.2 TFLOPS

Large model inferenceMulti-GPU professional workloads+1

NVIDIAAda Lovelace2022

L40

FP16 Performance181 TFLOPS

VisualizationInference+1

NVIDIAAda Lovelace2024

RTX 5880 Ada

FP16 Performance181 TFLOPS

Professional inferenceLarge model workstation use+1

NVIDIAAda Lovelace2022

RTX 4090

FP16 Performance165.2 TFLOPS

Fine-tuning small modelsInference up to 13B+1

NVIDIAAmpere2021

A30

FP16 Performance165 TFLOPS

Inference at scaleLow-power HPC+1

NVIDIABlackwell2025

RTX 5070

FP16 Performance144.2 TFLOPS

Ultra-budget inferenceExperimentation+1

NVIDIAAda Lovelace2023

RTX 5000 Ada

FP16 Performance130.6 TFLOPS

Mid-range professional inference30B model serving+1

NVIDIATuring2018

Quadro RTX 6000 24GB

FP16 Performance130.5 TFLOPS

Budget 13B inferenceNVLink multi-GPU setups+1

NVIDIATuring2018

Quadro RTX 8000 48GB

FP16 Performance130.5 TFLOPS

Budget 30B inferenceLarge VRAM workloads+1

NVIDIAAmpere2021

A10G

FP16 Performance125 TFLOPS

Inference servingGraphics workloads+1

NVIDIAAmpere2021

A10 24GB

FP16 Performance125 TFLOPS

Inference servingGraphics + compute+1

NVIDIAVolta2018

V100 32GB

FP16 Performance112 TFLOPS

Legacy ML trainingHPC workloads+1

NVIDIAVolta2017

V100 16GB

FP16 Performance112 TFLOPS

Legacy ML trainingBudget inference+1

NVIDIAVolta2017

TITAN V 12GB

FP16 Performance110 TFLOPS

Legacy ML researchBudget Volta workloads+1

NVIDIAAda Lovelace2022

RTX 4080

FP16 Performance97.5 TFLOPS

Budget inferenceSmall model fine-tuning+1

NVIDIAAda Lovelace2023

RTX 4500 Ada

FP16 Performance97.4 TFLOPS

Enterprise inferenceBudget professional AI+1

NVIDIATuring2018

Quadro RTX 5000 16GB

FP16 Performance89.2 TFLOPS

Budget inferenceNVLink multi-GPU setups+1

NVIDIAAda Lovelace2023

RTX 4070 Ti

FP16 Performance80.2 TFLOPS

Budget inferenceSmall model fine-tuning+1

NVIDIAAmpere2020

RTX A6000

FP16 Performance77.4 TFLOPS

Large model inferenceMulti-GPU NVLink setups+1

NVIDIAAmpere2020

A40

FP16 Performance74.8 TFLOPS

Visualization + computeMid-size inference+1

NVIDIAAmpere2020

RTX 3090

FP16 Performance71 TFLOPS

Budget inferenceSmall model fine-tuning+1

NVIDIABlackwell2025

RTX 5060 Ti

FP16 Performance70.4 TFLOPS

Budget AI inferenceSmall model serving+1

NVIDIAAmpere2021

RTX 3080 Ti

FP16 Performance68.2 TFLOPS

Budget inferenceSmall model fine-tuning+1

NVIDIAAmpere2021

RTX A5500 24GB

FP16 Performance68.2 TFLOPS

Professional inferenceNVLink multi-GPU setups+1

NVIDIATuring2018

T4

FP16 Performance65 TFLOPS

Cost-efficient inferenceHigh-density serving+1

NVIDIAAmpere2021

A16

FP16 Performance62.4 TFLOPS

VDI / virtual workstationsMulti-tenant inference+1

NVIDIAAda Lovelace2023

L4

FP16 Performance60.6 TFLOPS

Efficient inferenceEdge deployments+1

NVIDIAAmpere2020

RTX 3080

FP16 Performance59.6 TFLOPS

Ultra-budget inferenceExperimentation+1

NVIDIAAda Lovelace2023

RTX 4070

FP16 Performance58.2 TFLOPS

Ultra-budget inferenceExperimentation+1

NVIDIATuring2018

Quadro RTX 4000 8GB

FP16 Performance57 TFLOPS

Ultra-budget inferenceSmall model experimentation+1

NVIDIAAmpere2021

RTX A5000

FP16 Performance55.6 TFLOPS

Professional inferenceBudget NVLink multi-GPU+1

NVIDIAAda Lovelace2023

RTX 4000 Ada

FP16 Performance53.4 TFLOPS

Entry enterprise inferenceLow-power deployments+1

NVIDIABlackwell2025

RTX 5060

FP16 Performance48 TFLOPS

Entry-level AI inferenceStable Diffusion+1

NVIDIAAmpere2021

RTX A4500 20GB

FP16 Performance47.4 TFLOPS

Professional inferenceMid-range AI workloads+1

NVIDIAAda Lovelace2023

RTX 4060 Ti 16GB

FP16 Performance44.2 TFLOPS

Budget inferenceSmall model fine-tuning+1

NVIDIAAmpere2020

RTX 3070 8GB

FP16 Performance40.6 TFLOPS

Ultra-budget inferenceSmall model experimentation+1

NVIDIAAda Lovelace2023

RTX 4000 SFF Ada

FP16 Performance38.4 TFLOPS

Ultra-low-power inferenceEdge deployments+1

NVIDIAAmpere2021

RTX A4000 16GB

FP16 Performance38.4 TFLOPS

Entry professional inferenceBudget AI workloads+1

NVIDIAAmpere2020

RTX 3060 Ti 8GB

FP16 Performance32.4 TFLOPS

Ultra-budget inferenceSmall model experimentation+1

NVIDIAAda Lovelace2023

RTX 4060 8GB

FP16 Performance30.2 TFLOPS

Ultra-budget inferenceSmall model experimentation+1

NVIDIAAmpere2021

RTX 3060 12GB

FP16 Performance25.4 TFLOPS

Ultra-budget inferenceSmall model experimentation+1

NVIDIAAda Lovelace2023

RTX 2000 Ada

FP16 Performance24 TFLOPS

Entry professional inferenceLow-power AI workloads+1

NVIDIAPascal2016

Tesla P100 16GB

FP16 Performance18.7 TFLOPS

Legacy ML workloadsBudget HPC+1

NVIDIAAmpere2021

RTX A2000 12GB

FP16 Performance16 TFLOPS

Ultra-budget inferenceEdge AI+1

NVIDIAPascal2017

TITAN Xp 12GB

FP16 Performance12.1 TFLOPS

Legacy inferenceUltra-budget workloads+1

NVIDIAPascal2016

Quadro P6000 24GB

FP16 Performance12 TFLOPS

Legacy inferenceBudget 13B model serving+1

NVIDIAPascal2016

Quadro P5000 16GB

FP16 Performance8.9 TFLOPS

Legacy inferenceBudget 7B model serving+1

NVIDIAPascal2017

Quadro P4000 8GB

FP16 Performance5.3 TFLOPS

Legacy inferenceUltra-budget workloads+1

NVIDIAMaxwell2015

Quadro M4000 8GB

FP16 Performance2.6 TFLOPS

Legacy visualizationUltra-budget compute+1

Live GPU Prices Compare GPUs Head-to-Head Multi-GPU Comparison Tool

GPU specs guide for AI and cloud compute

NVIDIA Hopper: H100 and H200

NVIDIA's Hopper architecture (2022–2023) remains the dominant GPU for large-scale AI training and inference in 2026. The H100 SXM delivers 989 TFLOPS FP16 with 80GB HBM3 at 3.35 TB/s bandwidth and 900 GB/s NVLink 4.0 interconnect — the standard for 70B+ LLM training. The H100 PCIe variant offers the same compute at lower bandwidth (2 TB/s) and is 10–20% cheaper on cloud. The H200 upgrades memory to 141GB HBM3e at 4.8 TB/s bandwidth — a significant advantage for inference serving of large models that previously required multi-GPU configurations. H100 on-demand pricing ranges from $3.50–$5.50/hr depending on provider.

NVIDIA Blackwell: B200 and GB200

NVIDIA's Blackwell architecture (2024–2025) delivers a 2.5–4× performance improvement over Hopper for AI inference. The B200 provides 180GB HBM3e memory — enough to serve a 70B model on a single GPU — with 4.6 TB/s bandwidth and 1.8 PFLOPS FP8 performance. The GB200 NVL72 rack-scale system pairs 36 Grace CPUs with 72 B200 GPUs connected via NVLink Switch, delivering 1.4 exaFLOPS FP8 for training. Blackwell cloud availability is expanding in 2026 but remains limited compared to Hopper. B200 on-demand pricing is currently $8–$12/hr where available.

NVIDIA Ada Lovelace: L40S, L4, RTX 4090

Ada Lovelace (2022–2023) targets inference and professional workloads. The L40S (48GB GDDR6, 733 TFLOPS FP16) is the primary data-centre inference GPU for 7B–34B models, offering a better cost-per-token than H100 for models that fit in 48GB. The L4 (24GB GDDR6, 242 TFLOPS FP16) is the most cost-effective option for 7B model inference and video processing. The RTX 4090 (24GB GDDR6X, 330 TFLOPS FP16) is popular for fine-tuning and development — available from $0.44/hr on Vast.ai and RunPod, making it the cheapest path to 24GB VRAM in the cloud.

NVIDIA Ampere: A100, A40, A10, A10G

Ampere (2020–2021) remains widely available and cost-effective for many AI workloads. The A100 80GB SXM (312 TFLOPS FP16, 2 TB/s bandwidth, 600 GB/s NVLink 3.0) is the previous-generation training standard — still capable for 13B–70B model training at $1.89–$3.20/hr, roughly 40–60% cheaper than H100. The A100 40GB is available from $1.50/hr. The A40 (48GB GDDR6) and A10/A10G (24GB GDDR6) are popular inference GPUs — A10G is the standard GPU in AWS g5 instances. For workloads that don't require H100's bandwidth or NVLink scale, Ampere offers the best price-to-performance ratio in 2026.

AMD Instinct: MI300X and MI355X

AMD's Instinct MI300X (192GB HBM3, 5.3 TB/s bandwidth, 1.3 PFLOPS FP16) is the primary NVIDIA alternative for large-model inference. Its 192GB memory capacity — 2.4× the H100's 80GB — enables serving 70B models on a single GPU and 405B models on a 4-GPU node, reducing the multi-GPU overhead that H100 requires for large models. The MI355X (288GB HBM3e) extends this advantage further. ROCm software support has improved significantly in 2025–2026, with PyTorch, vLLM, and TGI all supporting AMD GPUs. Cloud availability is growing — Oracle Cloud, Fluidstack, and TensorWave offer MI300X instances.

Key specs to compare: VRAM and bandwidth

For AI workloads, VRAM capacity and memory bandwidth are often more important than raw FLOPS. VRAM determines the maximum model size that fits on a single GPU — a 70B model in FP16 requires 140GB, necessitating multi-GPU configurations on H100 (80GB) but fitting on a single MI300X (192GB). Memory bandwidth determines how fast weights can be loaded during inference — the primary bottleneck for autoregressive generation. H100 SXM's 3.35 TB/s bandwidth enables 2× faster inference than A100's 2 TB/s for the same model, even at similar FLOPS.

FP8 and quantisation in 2026

FP8 precision (supported on H100, H200, B200) doubles effective throughput versus FP16 for inference with minimal quality degradation. A 70B model in FP8 requires 70GB VRAM — fitting on a single H100 80GB — versus 140GB in FP16. INT4 quantisation (GPTQ, AWQ, GGUF) reduces this further to 35GB, enabling 70B inference on 2× A100 40GB or a single MI300X. The practical implication: for inference workloads, effective VRAM after quantisation is the relevant constraint, not raw VRAM capacity. The GPU specs database shows raw VRAM; factor in your target precision when selecting hardware.

NVLink and multi-GPU scaling

NVLink bandwidth determines how efficiently multiple GPUs can collaborate on a single large model. H100 SXM's NVLink 4.0 provides 900 GB/s bidirectional bandwidth between GPUs in an 8-GPU node — fast enough to treat 8× H100 as a single 640GB memory pool for tensor parallelism. H100 PCIe uses PCIe 5.0 (128 GB/s) instead of NVLink, making it significantly less efficient for multi-GPU training. A100 SXM uses NVLink 3.0 (600 GB/s). For single-GPU inference or training jobs that fit on one GPU, NVLink is irrelevant — PCIe variants are cheaper and equally capable.