常见GPU算力比较（历代游戏卡皇）

一、硬件参数

	2080Ti	3090	3090Ti	4090	4090D	5090	5090D
核心	TU102-300A	GA102-300	GA102-350	AD102-300	AD102-250	GB202-300	GB202-250
架构	Turing	Ampere	Ampere	Ada Lovelace	Ada Lovelace	Blackwell	Blackwell
SM	68	82	84	128	114	170	170
CUDA Cores / SM	64	128	128	128	128	128	128
CUDA Cores / GPU	4352	10496	10752	16384	14592	21760	21760
Tensor Core	2nd	3rd	3rd	4th	4th	5th	5th
Tensor Cores / SM	8	4	4	4	4	4	4
Tensor Cores / GPU	544	328	336	512	456	680	680
GPU 加速频率 (MHz)	1545	1695	1860	2520	2520	2407	2407
显存	11 / 22 GB (GDDR6)*	24 GB (GDDR6X)	24 GB (GDDR6X)	24 GB (GDDR6X)	24 GB (GDDR6X)	32 GB (GDDR7)	32 GB (GDDR7)
显存位宽 (bit)	352	384	384	384	384	512	512
显存速率 (Gbps)	14	19.5	21	21	21	28	28
显存带宽 (GBps)	616	936.2	1008	1008	1008	1792	1792
一缓 (KB per SM)	64	128	128	128	128	128	128
二缓 (MB)	6	6	6	72	72	96	96
TGP (W)	250	350	450	450	425	575	575
制程	TSMC 12nm FFN	Samsung 8N (8nm)	Samsung 8N (8nm)	TSMC 4N (5nm)	TSMC 4N (5nm)	TSMC 4N (5nm)	TSMC 4N (5nm)

* 22 GB 是常见的手动扩显存的魔改卡

二、算力

1、CUDA Core 算力

浮点：TFLOPS

整型：TIOPS

取 4090 的算力为100%

	2080Ti	3090	3090Ti	4090	4090D	5090	5090D
FP32	13.45	35.58	40.00	82.6	73.5	104.8	104.8
FP16	26.9	35.58	40.00	82.6	73.5	104.8	104.8
FP64	0.4202	0.556	0.625	1.29	1.149	1.64	1.64
BF16	NA	35.58	40.00	82.6	73.5	104.8	104.8
INT32	13.45	17.79	20.00	41.3	36.8	104.8	104.8

	2080Ti	3090	3090Ti	4090	4090D	5090	5090D
FP32	16.3%	43.1%	48.4%	100%	89.0%	126.9%	126.9%
FP16	32.6%	43.1%	48.4%	100%	89.0%	126.9%	126.9%
FP64	32.6%	43.1%	48.4%	100%	89.0%	126.9%	126.9%
BF16	NA	43.1%	48.4%	100%	89.0%	126.9%	126.9%
INT32	32.6%	43.1%	48.4%	100%	89.0%	253.6%	253.6%

2、Tensor Core 算力

浮点：TFLOPS

整型：TIOPS

稠密/稀疏

取 4090 的算力为100%

	2080Ti	3090	3090Ti	4090	4090D	5090	5090D*
FP4	NA	NA	NA	NA	NA	1676 / 3352	NA / 2375
FP8	NA	NA	NA	660.6 / 1321.2	588.4 / 1176.8	838 / 1676	NA / NA
FP16	107.6	142 / 284	160 / 320	330.3 / 660.6	294.2 / 588.4	419 / 838	NA / NA
BF16	NA	71 / 142	80 / 160	165.2 / 330.4	147.1 / 294.2	209.5 / 419	NA / NA
TF32	NA	35.6 / 71	40 / 80	82.6 / 165.2	73.5 / 147.1	104.8 / 209.5	NA / NA
INT8	215.2	284 / 568	320 / 640	660.6 / 1321.2	588.4 / 1176.8	838 / 1676	NA / NA
INT4	430.3	568 / 1136	640 / 1280	1321.2 / 2642.4	1176.8 / 2353.6	1676 / 3352	NA / NA

	2080Ti	3090	3090Ti	4090	4090D	5090	5090D*
FP4	NA	NA	NA	NA	NA	NA	NA
FP8	NA	NA	NA	100%	89.0%	126.9%	NA
FP16	32.6%	43.1%	48.4%	100%	89.0%	126.9%	NA
BF16	NA	43.1%	48.4%	100%	89.0%	126.9%	NA
TF32	NA	43.1%	48.4%	100%	89.0%	126.9%	NA
INT8	32.6%	43.1%	48.4%	100%	89.0%	126.9%	NA
INT4	32.6%	43.1%	48.4%	100%	89.0%	126.9%	NA

*5090D 的 Tensor Core 算力有待考证