AI GPU performance ladder

Updated: Sep 2026
Ranks: 26
Series: 10
Chips: 40

NVIDIA 29
AMD 9
Intel 2

Ranked by AI index (H100 SXM = 100). Hover a chip for its score.

Rank 1

NVIDIA Blackwell

GB300 GB200

AMD Instinct MI350

MI355X

Rank 2

AMD Instinct MI350

MI350X

Rank 3

NVIDIA Blackwell

B300 B200

Rank 4

AMD Instinct MI300

MI325X

Rank 5

AMD Instinct MI300

MI300X

Intel Gaudi

Gaudi 3

Rank 6

AMD Instinct MI300

MI300A

Rank 7

NVIDIA Hopper

GH200 H200 SXM

Rank 8

NVIDIA Hopper

H200 NVL

Rank 9

NVIDIA Hopper

H800 SXM H100 SXM H100 NVL

Rank 10

NVIDIA Hopper

H100 PCIe

Rank 11

AMD Instinct MI200

MI250X MI250

Rank 12

Intel Gaudi

Gaudi 2

Rank 13

NVIDIA Blackwell

RTX PRO 6000 Server

Rank 14

NVIDIA Ampere

A100 SXM 80GB A800 SXM A100 PCIe 80GB

Rank 15

NVIDIA Hopper

H20

Rank 16

NVIDIA Ampere

A100 SXM 40GB

Rank 17

NVIDIA Ada Lovelace

L40S

AMD Instinct MI200

MI210

Rank 18

AMD Instinct MI100

MI100

Rank 19

NVIDIA Ampere

A30

NVIDIA Ada Lovelace

L40

Rank 20

NVIDIA Volta / Turing

V100S PCIe

Rank 21

NVIDIA Volta / Turing

V100 SXM2

Rank 22

NVIDIA Volta / Turing

V100 PCIe

NVIDIA Ampere

A40

NVIDIA Ada Lovelace

L20

Rank 23

NVIDIA Ampere

A10

Rank 24

NVIDIA Ada Lovelace

Rank 25

NVIDIA Volta / Turing

Rank 26

NVIDIA Ampere

Rank
Rank	Volta / Turing	Ampere	Ada Lovelace	Hopper	Blackwell	Instinct MI350	Instinct MI300	Instinct MI200	Instinct MI100	Gaudi
1					GB300 GB200	MI355X
2						MI350X
3					B300 B200
4							MI325X
5							MI300X			Gaudi 3
6							MI300A
7				GH200 H200 SXM
8				H200 NVL
9				H800 SXM H100 SXM H100 NVL
10				H100 PCIe
11								MI250X MI250
12										Gaudi 2
13					RTX PRO 6000 Server
14		A100 SXM 80GB A800 SXM A100 PCIe 80GB
15				H20
16		A100 SXM 40GB
17			L40S					MI210
18									MI100
19		A30	L40
20	V100S PCIe
21	V100 SXM2
22	V100 PCIe	A40	L20
23		A10
24			L4
25	T4
26		A2

How the AI GPU ladder is ranked

Data-center accelerators have no shared editorial ladder, so this one is computed from spec-sheet figures. Each part gets an AI index; parts within 3% of a tier's best score share a rank.

AI index: √(16-bit tensor TFLOPS × memory bandwidth), scaled so the H100 SXM scores 100. Training tends to be limited by tensor throughput and inference by memory bandwidth, so the geometric mean weighs both equally.
16-bit tensor throughput: The faster of FP16 and BF16 tensor throughput, dense (without structured sparsity). Vendor figures quoted "with sparsity" are halved.
Coverage: Data-center GPUs and accelerators sold by NVIDIA, AMD, and Intel since Volta (2017). Cloud-only chips such as Google TPU, AWS Trainium, and Microsoft Maia are not listed. GB200 and GB300 are rated per Blackwell GPU.
Limits: Peak figures ignore interconnect, software maturity, and low-precision formats. B300 and B200 share a tier because their gains are FP4 throughput and memory capacity; within a tier, higher FP4, FP8, and memory rank first.

Data sources

Epoch AI — Data on Machine Learning Hardware CC BY license, retrieved 2026-09-23
Vendor datasheets and product pages fill gaps and corrections; linked on each accelerator page