新しいMac miniのM4 ProプロセッサのAI性能をGeekbench AIで測定してみた

2024年11月に発売された新型Mac miniに搭載されたM4 ProのAIワークロードのパフォーマンスを、Geekbench AIを利用して測定する機会がありましたので、紹介します。

Geekbench AIバージョン

• Geekbench AI 1.1.0 for macOS AArch64
  • ダウンロードURL

測定対象

比較のため、Mac mini 2023 (M2)とMacBook Pro (13-inch Late 2020)でも、Geekbench AIを実行しました。

-	Mac mini (M4PRO)	Mac mini (2023)	MacBook Pro (13-inch Late 2020)
Operating System	macOS 15.1 (Build 24B2083)	macOS 15.1 (Build 24B83)	macOS 15.1 (Build 24B83)
CPU Information
Topology	1 Processor, 12 Cores	1 Processor, 8 Cores	1 Processor, 8 Cores
Identifier	Apple M4 Pro	Apple M2	Apple M1
Base Frequency	4.51 GHz	3.48 GHz	1.95 GHz
Cluster 1	8 Cores	4 Cores	4 Cores
Cluster 2	4 Cores	4 Cores	4 Cores
Memory Information
Size	24.00 GB	16.00 GB	16.00 GB

Geekbench AI Score

「バックエンドを「CPU」「GPU」「Neural Engine」に指定、それぞれで得られた結果を求めました。

CPU

	Apple M4 Pro	Apple M2	Apple M1
Single Precision Score	5807	3456	2214
Half Precision Score	8889	5976	3549
Quantized Score	6899	4878	2872

GPU

	Apple M4 Pro	Apple M2	Apple M1
Single Precision Score	12797	6680	4180
Half Precision Score	14510	7520	4778
Quantized Score	12756	6583	4362

Neural Engine

	Apple M4 Pro	Apple M2	Apple M1
Single Precision Score	5814	3433	2215
Half Precision Score	37717	27857	12356
Quantized Score	51383	31717	12165

AI Performance

• 個別ワークロード別のスコアを集計しました。

CPU

Workload	Precision	M4 Pro		M2		M1
		Score	Rate	Score	Rate	Score	Rate
Image Classification	SP	2,093	389.2 IPS	1,382	257.0 IPS	938	174.4 IPS
	HP	3,051	567.5 IPS	2,532	470.8 IPS	1,408	261.8 IPS
	Q	2,202	411.3 IPS	1,776	331.7 IPS	993	185.3 IPS
Image Segmentation	SP	2,564	41.6 IPS	1,513	24.5 IPS	1,000	16.2 IPS
	HP	3,530	57.2 IPS	2,481	40.2 IPS	1,485	24.1 IPS
	Q	3,934	63.8 IPS	2,868	46.5 IPS	1,668	27.0 IPS
Pose Estimation	SP	18,952	22.1 IPS	6,813	7.95 IPS	4,301	5.02 IPS
	HP	30,048	35.1 IPS	14,900	17.4 IPS	9,133	10.7 IPS
	Q	24,112	28.3 IPS	12,993	15.2 IPS	7,821	9.17 IPS
Object Detection	SP	2,501	198.4 IPS	1,553	123.2 IPS	1,016	80.6 IPS
	HP	3,860	307.3 IPS	2,887	229.8 IPS	1,612	128.3 IPS
	Q	2,683	213.8 IPS	2,095	167.0 IPS	1,167	93.0 IPS
Face Detection	SP	4,524	53.8 IPS	3,071	36.5 IPS	2,171	25.8 IPS
	HP	7,880	93.6 IPS	5,870	69.8 IPS	3,727	44.3 IPS
	Q	4,675	55.5 IPS	3,526	41.9 IPS	2,066	24.5 IPS
Depth Estimation	SP	8,869	68.3 IPS	5,670	43.7 IPS	3,791	29.2 IPS
	HP	13,910	107.5 IPS	11,756	90.9 IPS	7,157	55.3 IPS
	Q	10,031	78.3 IPS	8,173	63.9 IPS	4,846	37.9 IPS
Style Transfer	SP	35,035	45.0 IPS	18,624	23.9 IPS	11,352	14.6 IPS
	HP	49,911	64.2 IPS	28,684	36.9 IPS	17,335	22.3 IPS
	Q	38,030	49.1 IPS	24,017	31.0 IPS	14,228	18.4 IPS
Image Super-Resolution	SP	7,055	260.5 IPS	4,127	152.4 IPS	2,553	94.3 IPS
	HP	9,596	354.4 IPS	4,857	179.3 IPS	2,985	110.2 IPS
	Q	5,362	198.0 IPS	3,617	133.6 IPS	2,185	80.7 IPS
Text Classification	SP	2,942	3.93 KIPS	2,261	3.02 KIPS	1,558	2.08 KIPS
	HP	5,301	7.08 KIPS	3,778	5.04 KIPS	2,636	3.52 KIPS
	Q	5,248	7.04 KIPS	3,748	5.03 KIPS	2,627	3.53 KIPS
Machine Translation	SP	5,878	101.3 IPS	3,632	62.6 IPS	1,865	32.1 IPS
	HP	8,864	153.2 IPS	5,927	102.4 IPS	2,835	49.0 IPS
	Q	8,688	150.1 IPS	5,867	101.4 IPS	3,089	53.4 IPS

GPU

Workload	Precision	M4 Pro		M2		M1
		Score	Rate	Score	Rate	Score	Rate
Image Classification	SP	7,244	1.35 KIPS	3,754	698.1 IPS	2,603	484.1 IPS
	HP	7,796	1.45 KIPS	3,697	687.5 IPS	2,168	403.1 IPS
	Q	6,604	1.23 KIPS	3,046	568.5 IPS	2,052	382.9 IPS
Image Segmentation	SP	10,793	175.0 IPS	4,015	65.1 IPS	2,409	39.1 IPS
	HP	11,345	183.9 IPS	5,010	81.2 IPS	3,725	60.4 IPS
	Q	11,457	185.7 IPS	5,114	82.9 IPS	3,096	50.2 IPS
Pose Estimation	SP	46,287	54.0 IPS	24,654	28.8 IPS	10,411	12.1 IPS
	HP	48,335	56.4 IPS	26,345	30.7 IPS	11,582	13.5 IPS
	Q	45,849	53.7 IPS	24,509	28.7 IPS	10,997	12.9 IPS
Object Detection	SP	6,648	527.4 IPS	3,539	280.7 IPS	1,928	153.0 IPS
	HP	7,213	572.1 IPS	3,702	293.7 IPS	2,651	210.3 IPS
	Q	6,025	481.5 IPS	3,091	247.0 IPS	2,220	177.4 IPS
Face Detection	SP	22,353	265.6 IPS	10,404	123.6 IPS	5,403	64.2 IPS
	HP	26,207	311.4 IPS	12,031	143.0 IPS	6,383	75.8 IPS
	Q	18,406	218.7 IPS	8,437	100.2 IPS	4,048	48.1 IPS
Depth Estimation	SP	31,660	243.9 IPS	16,759	129.1 IPS	10,695	82.4 IPS
	HP	33,130	255.3 IPS	17,873	137.7 IPS	9,989	77.0 IPS
	Q	26,962	210.7 IPS	14,868	116.2 IPS	10,883	85.1 IPS
Style Transfer	SP	105,013	135.0 IPS	42,372	54.5 IPS	27,702	35.6 IPS
	HP	112,970	145.2 IPS	48,536	62.4 IPS	24,885	32.0 IPS
	Q	107,312	138.5 IPS	45,326	58.5 IPS	31,249	40.3 IPS
Image Super-Resolution	SP	16,911	624.4 IPS	9,776	361.0 IPS	6,228	230.0 IPS
	HP	17,778	656.4 IPS	10,097	372.8 IPS	7,240	267.3 IPS
	Q	14,481	534.7 IPS	8,007	295.7 IPS	5,769	213.0 IPS
Text Classification	SP	2,956	3.95 KIPS	2,231	2.98 KIPS	1,552	2.07 KIPS
	HP	5,396	7.20 KIPS	3,761	5.02 KIPS	2,639	3.52 KIPS
	Q	5,413	7.27 KIPS	3,734	5.01 KIPS	2,620	3.52 KIPS
Machine Translation	SP	1,318	22.7 IPS	835	14.4 IPS	837	14.4 IPS
	HP	1,426	24.6 IPS	808	13.9 IPS	826	14.2 IPS
	Q	1,308	22.6 IPS	762	13.2 IPS	774	13.4 IPS

Neural Engine

Workload	Precision	M4 Pro		M2		M1
		Score	Rate	Score	Rate	Score	Rate
Image Classification	SP	2,105	391.4 IPS	1,388	258.1 IPS	941	175.0 IPS
	HP	18,470	3.43 KIPS	14,775	2.75 KIPS	5,084	945.5 IPS
	Q	23,527	4.39 KIPS	15,635	2.92 KIPS	4,729	883.0 IPS
Image Segmentation	SP	2,562	41.5 IPS	1,506	24.4 IPS	998	16.2 IPS
	HP	23,164	375.5 IPS	19,222	311.6 IPS	7,550	122.4 IPS
	Q	32,504	526.9 IPS	23,594	382.5 IPS	7,959	129.0 IPS
Pose Estimation	SP	18,826	22.0 IPS	6,811	7.95 IPS	4,274	4.99 IPS
	HP	123,705	144.3 IPS	106,538	124.3 IPS	48,795	56.9 IPS
	Q	232,400	272.4 IPS	106,262	124.5 IPS	46,889	55.0 IPS
Object Detection	SP	2,409	191.1 IPS	1,532	121.5 IPS	1,009	80.0 IPS
	HP	20,688	1.64 KIPS	15,850	1.26 KIPS	5,347	425.3 IPS
	Q	24,911	1.99 KIPS	17,979	1.44 KIPS	5,162	412.5 IPS
Face Detection	SP	4,593	54.6 IPS	3,039	36.1 IPS	2,186	26.0 IPS
	HP	45,299	538.3 IPS	33,535	398.5 IPS	11,607	137.9 IPS
	Q	62,123	738.2 IPS	42,746	507.9 IPS	11,217	133.3 IPS
Depth Estimation	SP	8,895	68.5 IPS	5,652	43.5 IPS	3,809	29.3 IPS
	HP	111,724	860.8 IPS	87,815	676.6 IPS	37,037	285.3 IPS
	Q	183,718	1.44 KIPS	111,102	868.2 IPS	36,941	288.7 IPS
Style Transfer	SP	35,660	45.8 IPS	18,613	23.9 IPS	11,354	14.6 IPS
	HP	239,728	308.2 IPS	202,604	260.4 IPS	92,269	118.6 IPS
	Q	303,659	391.9 IPS	194,470	251.0 IPS	87,488	112.9 IPS
Image Super-Resolution	SP	7,128	263.2 IPS	4,132	152.6 IPS	2,534	93.6 IPS
	HP	47,078	1.74 KIPS	43,093	1.59 KIPS	18,448	681.2 IPS
	Q	81,246	3.00 KIPS	51,332	1.90 KIPS	18,180	671.3 IPS
Text Classification	SP	2,960	3.95 KIPS	2,254	3.01 KIPS	1,560	2.08 KIPS
	HP	5,511	7.36 KIPS	3,794	5.06 KIPS	2,667	3.56 KIPS
	Q	6,106	8.19 KIPS	3,797	5.10 KIPS	2,638	3.54 KIPS
Machine Translation	SP	5,879	101.3 IPS	3,560	61.3 IPS	1,883	32.4 IPS
	HP	16,905	291.2 IPS	6,777	116.7 IPS	4,245	73.1 IPS
	Q	16,856	291.3 IPS	7,314	126.4 IPS	4,484	77.5 IPS

注釈：

• SP: Single Precision
• HP: Half Precision
• Q: Quantized
• IPS: Inferences Per Second
• KIPS: Kilo Inferences Per Second

まとめ

最新のM4 Proは、前世代のM2、M1プロセッサと比較し、高いAI性能のベンチマーク結果が得られました。

「M1と比較した場合、CPU処理は約2.6倍、GPU処理は約3.1倍、Neural Engineは約4.2倍の性能向上が確認できました。AppleシリコンのGPUやNeural Engineに最適化されたAIモデルの実行環境として、より高性能な利用が期待できます。

Apple Intelligence は2024年11月現在 日本語に対応していませんが、英語モードでの評価や、Appleシリコンをサポートするモデルについても今後評価していきたいと思います。