Million Miner Logo
Searching…

No products found for ""

AI GPU Benchmark Comparison: GPUs Ranked by Performance

Compare AI inference speed, VRAM capacity, memory bandwidth, FP16 TFLOPS, and training throughput side by side. Filter and rank graphics cards by real-world AI performance and hardware efficiency.

Still unsure? Compare any two graphics cards of your choice side by side. Jump to the comparison ↓

78 GPUs

More columns
Pick GPU ↕ Memory ↕ Bandwidth ↕ FP16 TFLOPS ↕ AI Index ▼ Power ↕ Price ↕ Precision Form factor Interconnect Training
NVIDIA · Blackwell · 2025
LLMImageVideoTraining
96 GB
GDDR7
1,597 GB/s 500.0
403
600 W $23,000 FP8 · FP4 PCIe MIG —
RTX Pro 5000 Blackwell
NVIDIA · Blackwell · 2025
LLMImageVideoTraining
48 GB
GDDR7
1,344 GB/s 258.0
285
300 W FP8 · FP4 PCIe MIG —
RTX 4090 Pro In Shop
NVIDIA · Ada Lovelace · 2022
LLMImageVideoTraining
48 GB
GDDR6X
1,008 GB/s 165.2
238
450 W $4,099 FP8 PCIe — —
RTX 5090 In Shop
NVIDIA · Blackwell · 2025
LLMImageVideoTraining
32 GB
GDDR7
1,792 GB/s 209.5
207
575 W FP8 · FP4 PCIe — —
A100 40GB In Shop
NVIDIA · Ampere · 2020
LLMImageVideoTraining
40 GB
HBM2
1,555 GB/s 312.0
152
250 W $9,500 — PCIe NVLink Bridge · MIG 1,396 img/s
NVIDIA · Blackwell · 2025
LLMImageVideoTraining
32 GB
GDDR7
896 GB/s 202.1
146
200 W $4,400 FP8 · FP4 PCIe — —
RTX 4080 Super Pro
NVIDIA · Ada Lovelace
LLMImageVideoTraining
32 GB
GDDR6X
736 GB/s 104.5
139
320 W FP8 PCIe — —
RTX 4090 In Shop
NVIDIA · Ada Lovelace · 2022
LLMImageVideoTraining
24 GB
GDDR6X
1,008 GB/s 165.2
133
450 W $3,200 FP8 PCIe — 1,301 img/s
RTX 3090 In Shop
NVIDIA · Ampere · 2020
LLMImageVideoTraining
24 GB
GDDR6X
936 GB/s 71.0
100
350 W $2,200 — PCIe NVLink Bridge 905 img/s
V100 32GB
NVIDIA · Volta · 2018
LLMImageVideoTraining
32 GB
HBM2
900 GB/s 125.0
84
300 W — SXM NVLink —
NVIDIA · Blackwell · 2025
LLMImageVideo
24 GB
GDDR7
672 GB/s —
83
145 W $2,600 FP8 · FP4 PCIe — —
V100 16GB
NVIDIA · Volta · 2017
LLMImage
16 GB
HBM2
900 GB/s 125.0
62
300 W — SXM NVLink 833 img/s
RTX 4070 Ti Super
NVIDIA · Ada Lovelace · 2024
LLMImage
16 GB
GDDR6X
672 GB/s 88.3
56
285 W FP8 PCIe — —
RTX A4000
NVIDIA · Ampere · 2021
LLMImage
16 GB
GDDR6
448 GB/s 76.7
51
140 W — PCIe — —
NVIDIA · Blackwell · 2025
LLMImageVideoTraining
96 GB
GDDR7
1,792 GB/s 438.9 — 300 W FP8 · FP4 PCIe MIG —
NVIDIA · Blackwell · 2025
LLMImageVideoTraining
96 GB
GDDR7
1,792 GB/s 503.8 — 600 W FP8 · FP4 PCIe MIG —
B200 SXM
NVIDIA · Blackwell · 2024
LLMImageVideoTraining
180 GB
HBM3e
7,700 GB/s 2,250.0 — 1000 W FP8 · FP4 SXM NVLink · MIG —
B100 SXM In Shop
NVIDIA · Blackwell · 2024
LLMImageVideoTraining
192 GB
HBM3e
8,000 GB/s 1,750.0 — 700 W $32,000 FP8 · FP4 SXM NVLink · MIG —
Instinct MI350X
AMD · CDNA4 · 2025
LLMImageVideoTraining
288 GB
HBM3e
8,000 GB/s 2,306.9 — 1000 W FP8 · FP4 OAM Infinity Fabric —
Instinct MI325X
AMD · CDNA3 · 2024
LLMImageVideoTraining
256 GB
HBM3e
6,000 GB/s 1,307.4 — 1000 W FP8 OAM Infinity Fabric —
H200 SXM In Shop
NVIDIA · Hopper · 2024
LLMImageVideoTraining
141 GB
HBM3e
4,800 GB/s 989.5 — 700 W $36,500 FP8 SXM NVLink · MIG —
H200 NVL In Shop
NVIDIA · Hopper · 2024
LLMImageVideoTraining
141 GB
HBM3e
4,800 GB/s 835.5 — 600 W $46,899 FP8 PCIe NVLink Bridge · MIG —
Gaudi 3
Intel · Gaudi 3 · 2024
LLMImageVideoTraining
128 GB
HBM2e
3,700 GB/s 459.0 — 900 W FP8 OAM RoCE Ethernet —
Instinct MI300X
AMD · CDNA3 · 2023
LLMImageVideoTraining
192 GB
HBM3
5,300 GB/s 1,307.4 — 750 W FP8 OAM Infinity Fabric —
H100 SXM5 In Shop
NVIDIA · Hopper · 2022
LLMImageVideoTraining
80 GB
HBM3
3,350 GB/s 989.5 — 700 W $26,500 FP8 SXM NVLink · MIG —
RTX 6000 Ada
NVIDIA · Ada Lovelace · 2022
LLMImageVideoTraining
48 GB
GDDR6
960 GB/s 364.0 — 300 W FP8 PCIe — —
RTX 5090 Blower In Shop
NVIDIA · Blackwell · 2025
LLMImageVideoTraining
32 GB
GDDR7
1,792 GB/s 209.5 — 575 W $6,099 FP8 · FP4 PCIe — —
NVIDIA · Blackwell · 2025
LLMImageVideo
24 GB
GDDR7
1,344 GB/s — — 575 W $5,200 FP8 · FP4 PCIe — —
H100 NVL In Shop
NVIDIA · Hopper · 2023
LLMImageVideoTraining
94 GB
HBM3
3,900 GB/s 835.5 — 400 W $26,500 FP8 PCIe NVLink Bridge · MIG —
H100 PCIe
NVIDIA · Hopper · 2022
LLMImageVideoTraining
80 GB
HBM2e
2,000 GB/s 756.5 — 350 W FP8 PCIe NVLink Bridge · MIG —
H800 In Shop
NVIDIA · Hopper · 2023
LLMImageVideoTraining
80 GB
HBM2e
2,000 GB/s 756.5 — 350 W $23,500 FP8 PCIe NVLink Bridge · MIG —
A800 80GB In Shop
NVIDIA · Ampere · 2022
LLMImageVideoTraining
80 GB
HBM2e
1,935 GB/s 312.0 — 300 W $12,500 — PCIe NVLink Bridge · MIG —
A100 80GB In Shop
NVIDIA · Ampere · 2020
LLMImageVideoTraining
80 GB
HBM2e
2,039 GB/s 312.0 — 400 W $13,500 — SXM NVLink · MIG —
A800 40GB In Shop
NVIDIA · Ampere · 2023
LLMImageVideoTraining
40 GB
HBM2
1,555 GB/s 311.9 — 240 W $12,500 — PCIe NVLink Bridge · MIG —
L40S In Shop
NVIDIA · Ada Lovelace · 2023
LLMImageVideoTraining
48 GB
GDDR6
864 GB/s 362.1 — 350 W $13,000 FP8 PCIe — —
RTX 5000 Ada In Shop
NVIDIA · Ada Lovelace · 2023
LLMImageVideoTraining
32 GB
GDDR6
576 GB/s 261.1 — 250 W $5,400 FP8 PCIe — —
Instinct MI250X
AMD · CDNA2 · 2021
LLMImageVideoTraining
128 GB
HBM2e
3,277 GB/s 383.0 — 560 W — OAM Infinity Fabric —
RTX 4070 Ti
NVIDIA · Ada Lovelace · 2023
Image
12 GB
GDDR6X
504 GB/s 80.1 — 285 W FP8 PCIe — —
L4
NVIDIA · Ada Lovelace · 2023
LLMImageVideoTraining
24 GB
GDDR6
300 GB/s 121.0 — 72 W FP8 PCIe — —
L40 In Shop
NVIDIA · Ada Lovelace · 2022
LLMImageVideoTraining
48 GB
GDDR6
864 GB/s 181.1 — 300 W $10,500 FP8 PCIe — —
Instinct MI210 In Shop
AMD · CDNA2 · 2022
LLMImageVideoTraining
64 GB
HBM2e
1,638 GB/s 181.0 — 300 W $6,500 — PCIe Infinity Fabric —
Instinct MI100
AMD · CDNA · 2020
LLMImageVideoTraining
32 GB
HBM2
1,229 GB/s 184.6 — 300 W — PCIe Infinity Fabric —
L20 In Shop
NVIDIA · Ada Lovelace · 2024
LLMImageVideoTraining
48 GB
GDDR6
864 GB/s 119.5 — 275 W $6,500 FP8 PCIe — —
Radeon RX 7900 XTX
AMD · RDNA3 · 2022
LLMImageVideoTraining
24 GB
GDDR6
960 GB/s 123.0 — 355 W — PCIe — —
RTX 4080 Super
NVIDIA · Ada Lovelace · 2024
LLMImage
16 GB
GDDR6X
736 GB/s 104.5 — 320 W FP8 PCIe — —
Radeon RX 7900 XT
AMD · RDNA3 · 2022
LLMImage
20 GB
GDDR6
800 GB/s 103.0 — 315 W — PCIe — —
RTX 4080
NVIDIA · Ada Lovelace · 2022
LLMImage
16 GB
GDDR6X
717 GB/s 97.5 — 320 W FP8 PCIe — —
Quadro RTX 8000 In Shop
NVIDIA · Turing · 2018
LLMImageVideoTraining
48 GB
GDDR6
672 GB/s 130.5 — 295 W $4,500 — PCIe NVLink Bridge —
A40
NVIDIA · Ampere · 2020
LLMImageVideoTraining
48 GB
GDDR6
696 GB/s 149.7 — 300 W — PCIe NVLink Bridge —
A30 In Shop
NVIDIA · Ampere · 2021
LLMImageVideoTraining
24 GB
HBM2
933 GB/s 165.0 — 165 W $5,700 — PCIe NVLink Bridge · MIG —
A10 In Shop
NVIDIA · Ampere · 2021
LLMImageVideoTraining
24 GB
GDDR6
600 GB/s 125.0 — 150 W $3,700 — PCIe — —
RTX 4500 Ada In Shop
NVIDIA · Ada Lovelace · 2023
LLMImageVideoTraining
24 GB
GDDR6
432 GB/s 158.5 — 210 W $4,400 FP8 PCIe — —
RTX A6000
NVIDIA · Ampere · 2020
LLMImageVideoTraining
48 GB
GDDR6
768 GB/s 154.8 — 300 W — PCIe NVLink Bridge —
RTX 3090 Ti
NVIDIA · Ampere · 2022
LLMImageVideoTraining
24 GB
GDDR6X
1,008 GB/s 80.0 — 450 W — PCIe NVLink Bridge —
RTX 3080 Ti
NVIDIA · Ampere · 2021
Image
12 GB
GDDR6X
912 GB/s 68.2 — 350 W — PCIe — —
Arc A770
Intel · Xe-HPG · 2022
LLMImage
16 GB
GDDR6
560 GB/s 131.0 — 225 W — PCIe — —
T4
NVIDIA · Turing · 2018
LLMImage
16 GB
GDDR6
300 GB/s 65.0 — 70 W — PCIe — —
RTX A5000
NVIDIA · Ampere · 2021
LLMImageVideoTraining
24 GB
GDDR6
768 GB/s 111.1 — 230 W — PCIe NVLink Bridge —
RTX A4500 In Shop
NVIDIA · Ampere · 2021
LLMImage
20 GB
GDDR6
640 GB/s 94.6 — 200 W $1,545 — PCIe NVLink Bridge —
RTX 4000 Ada In Shop
NVIDIA · Ada Lovelace · 2023
LLMImage
20 GB
GDDR6
360 GB/s 81.9 — 130 W $2,600 FP8 PCIe — —
NVIDIA · Ada Lovelace · 2023
LLMImage
20 GB
GDDR6
280 GB/s 76.7 — 70 W $1,550 FP8 PCIe — —
RTX 3080
NVIDIA · Ampere · 2020
10 GB
GDDR6X
760 GB/s 59.5 — 320 W — PCIe — —
RTX 2000 Ada In Shop
NVIDIA · Ada Lovelace · 2024
LLMImage
16 GB
GDDR6
224 GB/s 48.0 — 70 W $1,850 FP8 PCIe — —
Quadro RTX 5000 In Shop
NVIDIA · Turing · 2018
LLMImage
16 GB
GDDR6
448 GB/s 89.2 — 265 W $2,400 — PCIe NVLink Bridge —
Intel Flex 170 In Shop
Intel · Xe-HPG · 2022
LLMImage
16 GB
GDDR6
576 GB/s 128.0 — 150 W $1,634 — PCIe — —
RTX 3070
NVIDIA · Ampere · 2020
8 GB
GDDR6
448 GB/s 40.6 — 220 W — PCIe — —
RTX 4060 Ti 16GB
NVIDIA · Ada Lovelace · 2023
LLMImage
16 GB
GDDR6
288 GB/s 44.1 — 165 W FP8 PCIe — —
RTX A2000 In Shop
NVIDIA · Ampere · 2021
Image
12 GB
GDDR6
288 GB/s 32.0 — 70 W $2,569 — PCIe — —
RTX 4070
NVIDIA · Ada Lovelace · 2023
Image
12 GB
GDDR6X
504 GB/s 58.3 — 200 W FP8 PCIe — —
Radeon RX 7800 XT
AMD · RDNA3 · 2023
LLMImage
16 GB
GDDR6
624 GB/s 74.6 — 263 W — PCIe — —
Tesla P100 In Shop
NVIDIA · Pascal · 2016
16 GB
HBM2
732 GB/s 18.7 — 250 W $1,932 — PCIe — —
RTX 4060
NVIDIA · Ada Lovelace · 2023
8 GB
GDDR6
272 GB/s 30.3 — 115 W FP8 PCIe — —
Instinct MI50
AMD · Vega · 2018
32 GB
HBM2
1,024 GB/s 26.5 — 300 W — PCIe Infinity Fabric —
Radeon RX 7600
AMD · RDNA3 · 2023
8 GB
GDDR6
288 GB/s 43.5 — 165 W — PCIe — —
T1000 In Shop
NVIDIA · Turing · 2021
8 GB
GDDR6
160 GB/s 5.0 — 50 W $2,072 — PCIe — —
Quadro P2200 In Shop
NVIDIA · Pascal · 2019
5 GB
GDDR5X
200 GB/s 0.06 — 75 W $2,397 — PCIe — —
T400 In Shop
NVIDIA · Turing · 2021
4 GB
GDDR6
80 GB/s 2.2 — 30 W $2,646 — PCIe — —
Quadro P1000 In Shop
NVIDIA · Pascal · 2017
4 GB
GDDR5
82 GB/s 0.03 — 47 W $2,168 — PCIe — —

AI Index: real-world inference throughput across LLM, image generation and vision benchmarks, normalized to the RTX 3090 = 100.
FP16 is the dense tensor peak from the official datasheet, without sparsity. Prices are current in-stock prices; cards without stock are quoted on request. Data updated: October 2026

Head-To-Head

Put Any Two Cards In The Ring

Already spotted a couple of favorites above? Pick any two GPUs and watch them face off on VRAM, inference index, FP16 compute and training throughput, with a plain-language verdict on which one wins and why.

Fighter A
RTX Pro 6000 Blackwell
Flagship
403Inference Index
VS
Fighter B
RTX 3090
Mid-range
100Inference Index
96 GB VRAM 24 GB
403 AI Inference Index 100
500 Compute (FP16 TFLOPS) 71
1,597 GB/s Memory Bandwidth 936 GB/s
— Training Throughput 905 img/s
GDDR7Memory typeGDDR6X
600 WPower draw350 W
FP8 · FP4Precision—
PCIeForm factorPCIe
2025Launch year2020
$23,000Price$2,200
Verdict RTX Pro 6000 Blackwell leads RTX 3090 by about 4.0× on AI inference. It also carries more VRAM (96 GB vs 24 GB), so it fits larger models.
Data indexed to RTX 3090 = 100 · "—" means no published benchmark for that exact card
GPU hardware being benchmarked on the test bench
Inside a multi-GPU AI server running real workloads
Rows of GPU racks in the data center

How is the score calculated?

Our AI performance index combines verified benchmark data from LLM inference (8B–70B+), image generation and vision/OCR. All results are normalized to the RTX 3090 = 100, so you can compare GPUs directly at a glance.

Custom hardware & clusters

Need a configuration not listed above?

From single enterprise accelerators to custom 8-GPU rack servers: our engineers spec, validate and ship AI infrastructure worldwide.

  • Transparent sourcing for hard-to-find GPUs (H100, H200, Blackwell, Instinct)
  • Pre-configured HGX/DGX rack systems ready for data center deployment
  • Direct B2B consultation: talk directly with a systems engineer within 2 hours
Known from
  • BTC-Echo
  • Bitcoin MENA
  • NiceHash
  • ViaBTC
  • CryptoCompare
  • WhatToMine
  • ASIC Miner Value
  • Mining Now
Thanks, your request is in. Our team replies within 2 hours.
We never share your data. Replies typically within 2 hours, 24/7.

FAQ

FP16 TFLOPS is a theoretical compute ceiling published by chip manufacturers. The AI Inference Index reflects real-world throughput across LLM token generation, diffusion and vision pipelines. It accounts for memory bandwidth bottlenecks and architecture differences that theoretical numbers ignore.

For single-GPU inference and local model development, cards like the RTX 4090 and RTX 5090 deliver the highest throughput per dollar. For serving 70B+ parameters without aggressive quantization, professional cards with 48 GB to 96 GB VRAM (like the RTX 6000 Ada or RTX Pro Blackwell series) eliminate multi-GPU latency and reduce hosting overhead.

VRAM dictates whether your model loads at all; compute dictates how fast it generates tokens. If an LLM exceeds your available VRAM, offloading to system RAM drops generation speed sharply. Always size your VRAM first based on model parameters and context window, then optimize for bandwidth and TFLOPS.

Fine-tuning smaller models (8B to 14B) works reliably on 24 GB consumer GPUs via LoRA/QLoRA. Full pre-training or running large continuous batches requires enterprise cards (A100, H100, H200 or MI300X) due to faster HBM memory bandwidth, FP8/FP4 tensor support and the high-speed interconnects (NVLink) required for multi-node clusters.

Individual GPU stock is updated continuously with live in-stock pricing. For high-density enterprise cards and multi-GPU rack systems (4x to 8x HGX/DGX configurations), pricing is quoted on request to match current supply, custom cooling requirements and worldwide DDP logistics.

AI GPUs and AI servers from our shop