CPU Architecture Analysis

Correlating 55 CPU architecture families (1989–2025) with the 24 Livermore Loops kernels, from 130 EvoBench measurements. Normalization: single-core = score per 1 GHz (CLOCKRATE_MAX); multi-core = score per Watt (TDP_MAX).

The 24 Livermore kernels

The Livermore Loops (1986, Lawrence Livermore National Laboratory) are 24 numerical inner loops with very different characters. That difference is exactly what makes them a micro-architecture probe: some loops vectorize perfectly (SIMD width and FMA throughput dominate), others are serial recurrences or branchy scans where only latency, branch prediction and out-of-order depth help.

#	Kernel	Character
L1	Hydro fragment	vectorizable
L2	ICCG (Incomplete Cholesky Conj. Grad.)	partially vectorizable
L3	Inner product	vectorizable
L4	Banded linear equations	partially vectorizable
L5	Tri-diagonal elimination	serial recurrence
L6	General linear recurrence	serial recurrence
L7	Equation of state fragment	vectorizable
L8	ADI integration	vectorizable
L9	Integrate predictors	vectorizable
L10	Difference predictors	vectorizable
L11	First sum (prefix sum)	serial recurrence
L12	First difference	vectorizable
L13	2-D Particle-In-Cell	gather/scatter (irregular memory)
L14	1-D Particle-In-Cell	gather/scatter (irregular memory)
L15	Casual Fortran (branchy matrix)	branch-heavy
L16	Monte Carlo search	branch-heavy
L17	Implicit conditional computation	branch-heavy
L18	2-D explicit hydrodynamics	vectorizable
L19	General linear recurrence (2)	serial recurrence
L20	Discrete ordinates transport	serial recurrence
L21	Matrix * Matrix product	vectorizable
L22	Planckian distribution	vectorizable
L23	2-D implicit hydrodynamics	partially vectorizable
L24	First minimum location	serial recurrence

35 years of IPC, kernel by kernel

Single-core OPTIMIZED score per 1 GHz (IPC proxy). Geometric mean over every result in the family. Colors are scaled per column (per kernel) on a log scale: red = lowest IPC ever measured for that kernel, green = highest. Reading a row left-to-right shows the strong/weak kernels of that architecture; reading a column top-to-bottom shows 35 years of progress on one loop.

Architecture family (by year)	Year	Overall	L1	L2	L3	L4	L5	L6	L7	L8	L9	L10	L11	L12	L13	L14	L15	L16	L17	L18	L19	L20	L21	L22	L23	L24
P4	1989	36.2	52.5	48.9	35.1	54.6	36.9	29.6	53.7	48.3	58.5	18.6	23.3	26.7	14.8	17.9	38.7	44.1	55.7	42.8	45.0	52.5	37.2	22.1	42.8	29.1
P5 aka i586 / P54C	1993	121	221	164	197	187	154	83.1	239	173	219	110	103	95.4	32.9	49.9	83.2	124	164	142	202	122	94.4	53.5	199	64.0
K5	1996	89.8	155	132	144	126	124	71.7	156	97.5	137	67.0	109	82.9	37.5	43.4	53.8	114	137	86.6	179	72.9	85.0	26.5	121	43.3
K6	1997	110	208	123	208	220	169	139	244	228	262	48.6	57.4	101	15.3	32.4	85.2	95.2	117	140	240	102	135	47.4	196	41.0
P6 (Klamath and Deschutes)	1997	216	406	386	371	380	187	205	474	205	502	232	222	262	36.3	65.9	129	194	418	278	254	163	249	62.5	305	140
P55C aka i586	1997	122	242	186	221	230	140	92.9	220	211	319	119	63.4	64.2	42.0	44.9	82.3	127	120	141	187	120	118	53.6	184	66.0
K7, Athlon, "Argon", "Orion"	1999	316	699	545	441	365	233	313	755	824	948	371	221	311	43.7	134	244	260	474	520	235	199	460	102	437	175
P6	1999	262	533	424	374	404	240	261	531	552	636	375	235	301	40.0	89.3	137	207	517	336	261	163	349	64.1	411	145
P6	1999	228	430	401	378	391	195	217	489	241	546	246	227	275	38.0	71.8	133	200	454	301	258	168	269	63.7	336	143
K7, Athlon, "Thunderbird"	2000	326	764	565	461	385	240	321	736	817	1,044	419	231	303	50.4	145	246	250	486	526	242	201	482	95.1	442	164
NetBust / Northwood	2002	220	591	452	300	345	162	211	504	276	599	202	191	244	52.1	67.5	186	191	354	314	166	115	330	65.4	308	132
K8 (Newcastle / and others...)	2003	331	669	347	438	388	195	313	860	782	867	227	220	232	91.0	206	330	343	415	425	208	254	492	221	451	198
Banias / Dothan / Centrino	2003	311	755	479	502	416	248	330	792	904	947	466	278	415	34.5	83.1	156	229	581	404	272	166	455	69.8	460	191
Yonah	2005	331	758	476	506	413	249	331	795	910	960	467	284	417	67.9	161	157	235	582	404	273	166	456	69.7	462	192
Conroe	2006	408	905	399	618	475	193	387	1,253	1,242	1,379	261	346	272	199	333	185	351	429	643	213	189	622	240	514	292
Wolfdale	2007	436	917	402	626	516	193	389	1,272	1,247	1,389	263	346	272	204	331	409	355	428	737	213	250	624	288	515	293
Nehalem / (Bloomfield or Lynnfield)	2008	548	2,313	595	652	614	245	442	2,541	1,383	1,495	439	346	623	147	309	304	444	582	994	252	248	1,526	310	545	318
K10 (Deneb, Thuban)	2009	417	1,379	494	459	427	162	322	1,772	1,050	1,057	362	233	473	116	214	393	381	412	785	188	250	1,049	261	457	201
Sandy Bridge	2011	599	2,439	771	626	659	260	501	2,917	1,407	1,709	527	332	699	227	376	371	526	519	1,078	245	244	1,712	316	569	307
ARM11	2012	87.7	149	113	181	171	73.1	66.0	241	140	167	86.6	56.3	55.9	29.1	37.7	92.9	75.6	116	102	127	81.4	83.6	47.4	88.6	33.7
Haswell	2013	728	2,787	927	769	808	323	521	3,470	1,455	1,827	786	398	915	226	417	593	705	650	1,520	273	269	2,131	464	594	410
Cyclone	2013	507	933	402	427	568	211	255	1,490	1,494	1,261	518	227	778	240	385	667	686	411	1000	173	548	755	475	374	296
Amd E1	2014	419	1,361	490	601	490	238	288	1,528	741	900	307	294	419	152	237	309	346	461	606	272	280	1,013	187	533	207
Typhoon	2014	590	1,072	523	440	607	250	271	1,934	1,642	1,267	605	339	753	244	415	665	772	690	1,048	256	673	757	528	618	314
Skylake	2015	759	2,891	1,005	731	710	318	431	3,308	1,780	1,698	870	357	865	346	402	1,487	830	550	1,871	236	298	2,256	658	527	250
ARM Cortex A7	2015	151	181	147	240	203	149	101	254	223	230	188	160	143	66.0	86.9	141	121	217	163	171	147	138	76.3	186	104
Twister	2015	760	1,487	540	606	831	313	339	2,697	2,188	2,112	757	399	947	399	497	1,240	807	682	1,641	229	989	880	709	832	344
ARM Cortex A53	2016	222	380	227	378	295	161	116	614	395	388	190	120	157	88.3	145	216	198	291	380	222	197	280	146	264	123
Hurricane + Zephyr	2016	813	1,715	997	601	857	293	367	2,585	2,433	2,010	1,027	394	960	384	500	1,473	775	619	1,708	200	939	1,050	837	861	459
Zen1	2017	831	3,275	1,009	910	947	356	492	3,894	1,943	2,069	944	407	933	232	436	1,210	778	713	2,039	285	364	2,257	648	649	308
Kaby Lake (Skylake variant)	2017	798	3,008	1,043	766	724	333	448	3,457	1,854	1,770	903	374	876	356	470	1,559	804	579	1,956	248	311	2,347	680	557	316
Coffee Lake (Skylake variant)	2017	796	3,016	1,048	768	723	325	446	3,457	1,828	1,781	896	371	881	370	463	1,565	793	575	1,978	247	312	2,252	687	552	315
Monsoon + Mistral	2017	989	1,817	1,141	663	1,004	367	434	3,168	2,715	2,678	1,320	445	1,617	457	692	2,506	1,142	610	2,125	197	1,007	1,199	1,045	914	600
Coffee Lake refresh (Skylake variant)	2018	800	3,323	997	753	779	330	446	3,424	1,812	1,753	891	364	946	383	483	1,544	681	559	2,021	247	310	2,431	656	554	335
Vortex + Tempest	2018	971	1,707	758	650	902	383	415	3,552	2,771	2,664	1,374	439	1,526	466	544	1,945	1,209	668	2,225	235	1,244	1,128	1,092	930	572
Zen2 "Matisse"	2019	867	3,823	1,116	1,041	1,019	320	517	4,195	1,977	2,153	657	430	1,202	248	462	1,175	822	634	2,419	269	375	2,570	749	724	299
ARM Cortex A72	2019	672	1,638	845	868	691	345	532	1,859	2,006	1,295	601	436	733	197	404	661	586	629	936	361	629	1,383	513	667	260
Lightning + Thunder	2019	1,057	2,011	898	723	1,048	459	447	3,709	2,870	3,162	1,310	448	1,511	457	724	2,471	1,246	718	2,306	249	1,305	1,252	1,363	840	579
Zen3 "Vermeer"	2020	975	4,090	1,128	1,038	1,042	440	522	4,488	2,333	2,256	1,223	517	1,510	341	519	1,136	933	800	1,970	321	386	2,845	834	728	336
Comet Lake (Skylake variant)	2020	786	3,042	974	772	793	330	444	3,217	1,707	1,755	863	382	838	361	442	1,514	777	572	1,919	240	302	2,215	648	546	376
Firestorm + Icestorm	2020	1,447	2,951	2,099	948	1,253	570	573	5,657	4,423	4,863	1,966	634	1,985	669	1,057	3,010	1,562	822	3,519	275	1,354	1,787	1,970	1,104	702
Firestorm + Icestorm	2020	1,119	2,064	1,383	734	1,015	495	530	3,752	2,901	3,157	1,373	495	1,513	552	839	2,310	1,143	698	2,681	241	1,345	1,415	1,313	924	591
Rocket Lake / Cypress Cove / Sunny Cove	2021	813	2,905	998	758	874	367	431	3,450	1,768	1,744	1,057	353	795	323	556	1,442	898	613	2,264	221	283	2,169	762	531	414
Alder Lake / Golden Cove	2021	1,282	4,704	1,527	1,590	1,422	568	804	5,699	2,869	2,932	1,307	725	1,667	567	798	1,627	1,065	1,043	3,832	332	360	3,911	1,072	937	469
Avalanche + Blizzard	2021	1,162	2,063	1,376	810	1,039	513	577	4,014	3,005	3,408	1,422	545	1,518	595	813	2,291	1,169	870	2,684	220	1,355	1,347	1,366	1,073	601
Zen4 "Raphael"	2022	1,076	4,543	1,164	1,143	1,182	500	562	4,904	2,501	2,459	1,273	543	1,517	507	592	1,195	1,083	839	2,783	329	379	2,932	839	757	439
Raptor Lake / Golden Cove	2022	1,218	3,891	1,255	1,538	1,421	561	831	5,447	2,701	2,659	1,281	775	1,575	556	749	1,637	1,173	1,025	3,234	332	357	3,064	1,092	928	384
Avalanche + Blizzard	2022	1,482	2,893	2,009	1,018	1,247	583	615	5,587	4,339	4,895	2,014	703	2,182	729	1,041	3,039	1,589	996	3,488	252	1,353	1,779	1,869	1,263	715
Everest + Sawtooth	2022	1,321	2,556	1,676	864	1,174	535	607	4,870	3,603	4,263	1,646	602	1,662	673	943	2,752	1,399	925	3,186	240	1,348	1,568	1,606	1,155	673
Raptor Lake Refresh / Golden Cove	2023	1,209	4,080	1,292	1,575	1,443	568	845	5,385	2,571	2,691	1,308	577	1,353	635	762	1,632	1,043	1,056	3,096	333	360	3,120	1,082	935	429
Macbook Pro M3 (2023)	2023	1,526	2,892	1,980	1,048	1,538	628	628	5,562	4,285	4,928	1,817	696	1,901	798	1,073	4,267	1,495	1,001	3,435	309	1,495	1,618	2,040	1,251	685
ARM Cortex A76	2023	829	2,552	721	1,030	791	376	575	3,323	1,852	1,715	728	420	818	277	459	1,110	623	668	1,837	430	870	1,901	625	799	334
Zen5 "Granite Ridge"	2024	1,310	4,623	1,116	1,551	1,543	551	724	6,119	2,755	3,135	1,570	693	1,492	660	867	1,335	1,567	964	3,503	359	393	3,012	1,259	959	742
(?)	2024	1,615	2,775	1,943	1,065	1,668	669	741	5,384	4,197	6,127	2,014	719	2,116	857	1,056	4,577	1,669	1,027	3,045	302	1,560	1,587	2,256	1,956	687
Apple M5	2025	1,630	2,789	1,965	1,013	1,632	686	753	5,505	4,277	6,259	2,030	676	1,970	867	1,082	4,588	1,901	895	3,481	283	1,860	1,600	2,288	1,997	671

What did the generations actually speed up?

Score/GHz growth factor from the P5 aka i586 / P54C (1993) baseline to the newest measured families: Raptor Lake Refresh / Golden Cove (Intel), Zen5 "Granite Ridge" (AMD), Apple M5 (Apple).

Fastest-improving kernels: L22 Planckian distribution ×27.3; L21 Matrix * Matrix product ×26.1; L15 Casual Fortran (branchy matrix) ×25.9.
The big winners are the loops modern hardware could attack in parallel or predict away: vector-friendly kernels collected every SIMD dividend (SSE→AVX2/NEON width, FMA units, more load ports, giant caches), while branchy loops like Casual Fortran were rescued by modern branch predictors that make well-patterned conditions nearly free.

Slowest-improving kernels: L20 Discrete ordinates transport ×5.25; L5 Tri-diagonal elimination ×3.89; L19 General linear recurrence (2) ×1.60.
Serial recurrences (each iteration needs the previous result) cannot be vectorized or reordered; per-clock gains here come only from out-of-order depth and latency improvements, a few percent per generation. This is the “legacy code” speed limit of every architecture.

#◂	Kernel◂	Character◂	Intel◂	AMD◂	Apple◂	Average growth◂
L22	Planckian distribution	vectorizable	×20.2	×23.5	×42.8	×27.3
L21	Matrix * Matrix product	vectorizable	×33.0	×31.9	×16.9	×26.1
L15	Casual Fortran (branchy matrix)	branch-heavy	×19.6	×16.0	×55.1	×25.9
L7	Equation of state fragment	vectorizable	×22.6	×25.6	×23.1	×23.7
L18	2-D explicit hydrodynamics	vectorizable	×21.7	×24.6	×24.4	×23.6
L13	2-D Particle-In-Cell	gather/scatter (irregular memory)	×19.3	×20.1	×26.3	×21.7
L8	ADI integration	vectorizable	×14.8	×15.9	×24.7	×18.0
L14	1-D Particle-In-Cell	gather/scatter (irregular memory)	×15.3	×17.4	×21.7	×17.9
L9	Integrate predictors	vectorizable	×12.3	×14.3	×28.6	×17.1
L1	Hydro fragment	vectorizable	×18.4	×20.9	×12.6	×16.9
L12	First difference	vectorizable	×14.2	×15.6	×20.7	×16.6
L10	Difference predictors	vectorizable	×11.9	×14.2	×18.4	×14.6
L16	Monte Carlo search	branch-heavy	×8.41	×12.6	×15.3	×11.8
L24	First minimum location	serial recurrence	×6.70	×11.6	×10.5	×9.34
L6	General linear recurrence	serial recurrence	×10.2	×8.72	×9.06	×9.29
L2	ICCG (Incomplete Cholesky Conj. Grad.)	partially vectorizable	×7.86	×6.79	×12.0	×8.61
L4	Banded linear equations	partially vectorizable	×7.70	×8.24	×8.71	×8.21
L3	Inner product	vectorizable	×8.01	×7.88	×5.15	×6.88
L11	First sum (prefix sum)	serial recurrence	×5.63	×6.75	×6.58	×6.30
L23	2-D implicit hydrodynamics	partially vectorizable	×4.71	×4.83	×10.1	×6.12
L17	Implicit conditional computation	branch-heavy	×6.44	×5.88	×5.46	×5.91
L20	Discrete ordinates transport	serial recurrence	×2.95	×3.22	×15.2	×5.25
L5	Tri-diagonal elimination	serial recurrence	×3.69	×3.58	×4.46	×3.89
L19	General linear recurrence (2)	serial recurrence	×1.65	×1.78	×1.40	×1.60

The IPC race

Overall single-core optimized score per GHz, by year of architecture introduction (log scale). Hover the dots.

Kernel correlation clusters

Pearson correlation of log(score/GHz) between kernel pairs, across the 130 CPUs with complete data. Kernels that always rise and fall together respond to the same hardware resources.

Clusters (r ≥ 0.985)

L7 (Equation of state fragment), L18 (2-D explicit hydrodynamics)

Most correlated pairs

Kernel A	Kernel B	r
L7 Equation of state fragment	L18 2-D explicit hydrodynamics	0.991
L1 Hydro fragment	L21 Matrix * Matrix product	0.985
L1 Hydro fragment	L7 Equation of state fragment	0.981
L16 Monte Carlo search	L18 2-D explicit hydrodynamics	0.980
L13 2-D Particle-In-Cell	L14 1-D Particle-In-Cell	0.977
L15 Casual Fortran (branchy matrix)	L22 Planckian distribution	0.976
L16 Monte Carlo search	L22 Planckian distribution	0.976
L14 1-D Particle-In-Cell	L22 Planckian distribution	0.974
L13 2-D Particle-In-Cell	L22 Planckian distribution	0.973
L7 Equation of state fragment	L16 Monte Carlo search	0.972

Least correlated pairs (the architecture separators)

Kernel A	Kernel B	r
L13 2-D Particle-In-Cell	L19 General linear recurrence (2)	0.457
L19 General linear recurrence (2)	L20 Discrete ordinates transport	0.476
L19 General linear recurrence (2)	L22 Planckian distribution	0.488
L15 Casual Fortran (branchy matrix)	L19 General linear recurrence (2)	0.506
L14 1-D Particle-In-Cell	L19 General linear recurrence (2)	0.507
L8 ADI integration	L19 General linear recurrence (2)	0.563
L16 Monte Carlo search	L19 General linear recurrence (2)	0.585
L19 General linear recurrence (2)	L24 First minimum location	0.587
L18 2-D explicit hydrodynamics	L19 General linear recurrence (2)	0.611
L9 Integrate predictors	L19 General linear recurrence (2)	0.621

Low-correlation pairs are the most useful diagnostics: a CPU can be great at one and poor at the other, so these two loops measure genuinely different parts of the machine (e.g. SIMD throughput vs. branchy serial latency).

Performance per Watt (multi-core)

Multi-core OPTIMIZED score per Watt of TDP. Same color rules. This view rewards efficiency: phone SoCs and chiplet-era CPUs glow green, NetBurst glows red.

Architecture family (by year)	Year	Overall	L1	L2	L3	L4	L5	L6	L7	L8	L9	L10	L11	L12	L13	L14	L15	L16	L17	L18	L19	L20	L21	L22	L23	L24
P4	1989	0.34	0.50	0.47	0.33	0.52	0.35	0.28	0.51	0.46	0.56	0.18	0.22	0.25	0.14	0.17	0.37	0.42	0.53	0.41	0.43	0.50	0.35	0.21	0.41	0.28
P5 aka i586 / P54C	1993	1.33	2.43	1.81	2.16	2.06	1.69	0.91	2.62	1.91	2.41	1.21	1.13	1.05	0.36	0.55	0.92	1.36	1.80	1.57	2.22	1.34	1.04	0.59	2.18	0.70
K5	1996	0.64	1.11	0.94	1.03	0.90	0.89	0.51	1.11	0.70	0.98	0.48	0.78	0.59	0.27	0.31	0.38	0.81	0.98	0.62	1.28	0.52	0.61	0.19	0.86	0.31
K6	1997	2.61	4.95	2.92	4.95	5.24	4.03	3.31	5.81	5.43	6.24	1.16	1.37	2.40	0.36	0.77	2.03	2.27	2.78	3.34	5.71	2.43	3.22	1.13	4.67	0.98
P6 (Klamath and Deschutes)	1997	3.12	5.87	5.57	5.36	5.48	2.70	2.97	6.83	2.96	7.25	3.35	3.21	3.78	0.52	0.95	1.87	2.80	6.03	4.01	3.67	2.35	3.59	0.90	4.40	2.02
P55C aka i586	1997	1.59	3.14	2.41	2.87	2.99	1.81	1.21	2.85	2.73	4.14	1.55	0.82	0.83	0.54	0.58	1.07	1.64	1.56	1.82	2.42	1.56	1.53	0.70	2.39	0.86
K7, Athlon, "Argon", "Orion"	1999	5.49	12.1	9.46	7.64	6.33	4.04	5.43	13.1	14.3	16.5	6.43	3.84	5.40	0.76	2.33	4.23	4.51	8.22	9.01	4.08	3.45	7.99	1.77	7.57	3.03
P6	1999	9.71	19.8	15.8	13.9	15.0	8.90	9.69	19.7	20.5	23.6	13.9	8.72	11.2	1.49	3.32	5.09	7.67	19.2	12.5	9.70	6.07	13.0	2.38	15.2	5.37
P6	1999	4.06	7.66	7.15	6.74	6.97	3.48	3.86	8.72	4.29	9.72	4.37	4.05	4.90	0.68	1.28	2.38	3.56	8.08	5.36	4.60	2.98	4.79	1.13	5.99	2.54
K7, Athlon, "Thunderbird"	2000	6.04	14.1	10.5	8.54	7.13	4.44	5.94	13.6	15.1	19.3	7.76	4.28	5.61	0.93	2.69	4.56	4.63	9.00	9.74	4.48	3.72	8.93	1.76	8.19	3.04
NetBust / Northwood	2002	8.94	24.0	18.3	12.2	14.0	6.59	8.58	20.4	11.2	24.3	8.19	7.75	9.91	2.12	2.74	7.57	7.75	14.3	12.7	6.75	4.67	13.4	2.65	12.5	5.35
K8 (Newcastle / and others...)	2003	7.45	15.0	7.81	9.85	8.71	4.37	7.02	19.3	17.6	19.5	5.11	4.96	5.21	2.04	4.63	7.42	7.70	9.31	9.54	4.67	5.72	11.1	4.97	10.1	4.45
Banias / Dothan / Centrino	2003	17.1	41.5	26.3	27.6	22.8	13.6	18.1	43.5	49.6	52.0	25.6	15.3	22.8	1.89	4.56	8.59	12.6	31.9	22.2	14.9	9.10	25.0	3.83	25.3	10.5
Yonah	2005	46.2	106	66.5	70.6	57.8	34.8	46.1	111	126	134	65.3	39.5	58.2	9.52	22.4	22.0	32.6	81.4	56.4	38.2	23.2	63.7	9.74	64.5	26.8
Conroe	2006	28.4	63.0	27.8	43.0	33.0	13.5	27.1	86.2	86.5	95.5	18.2	24.1	18.9	13.9	22.4	13.0	24.5	30.1	44.8	14.9	13.3	43.4	16.7	36.0	20.1
Wolfdale	2007	40.0	84.4	37.0	57.6	47.5	17.7	35.9	117	114	128	24.2	31.8	25.1	18.8	29.5	37.7	32.6	39.4	67.6	19.6	23.0	57.4	26.5	47.4	26.6
Nehalem / (Bloomfield or Lynnfield)	2008	94.9	281	81.4	141	96.8	59.7	63.2	322	172	188	68.4	64.4	95.8	35.0	55.3	68.1	69.1	122	151	66.9	61.7	218	52.0	132	52.5
K10 (Deneb, Thuban)	2009	56.2	186	66.5	62.0	57.6	21.8	43.5	237	141	142	48.8	31.3	63.7	15.6	28.9	53.0	51.1	55.3	106	25.3	33.7	141	35.2	61.6	27.1
Sandy Bridge	2011	117	372	123	174	133	67.9	98.7	438	212	244	82.5	83.0	137	46.4	63.6	62.7	99.8	141	210	74.3	59.8	268	52.9	148	60.5
ARM11	2012	20.5	34.7	26.4	42.3	40.0	17.1	15.4	56.3	32.6	39.0	20.2	13.1	13.0	6.80	8.80	21.7	17.6	27.0	23.7	29.7	19.0	19.5	11.1	20.7	7.87
Haswell	2013	191	651	229	255	214	96.2	144	795	340	420	178	111	225	53.6	101	151	174	196	385	89.4	85.5	494	123	183	107
Cyclone	2013	370	769	346	346	425	161	191	1,095	1,099	917	385	162	514	167	270	459	471	298	699	121	393	548	343	270	213
Amd E1	2014	113	367	132	162	132	63.9	77.7	412	200	243	82.6	79.4	112	41.0	63.5	83.3	92.7	125	164	73.3	75.6	273	50.4	144	55.3
Typhoon	2014	420	819	417	326	424	184	209	1,449	1,223	924	444	240	530	171	289	464	527	458	713	175	459	517	358	422	213
Skylake	2015	233	711	261	254	216	128	168	823	390	431	204	134	272	94.7	120	365	214	240	488	114	139	545	185	222	75.4
ARM Cortex A7	2015	135	162	132	217	182	134	89.8	228	198	206	160	144	129	59.3	75.8	127	109	195	146	154	131	124	68.5	164	93.3
Twister	2015	562	1,083	421	457	619	238	260	1,953	1,593	1,550	552	293	729	297	378	945	625	487	1,184	161	695	631	513	605	255
ARM Cortex A53	2016	169	298	179	295	227	122	90.0	474	304	305	148	94.8	122	68.8	110	164	153	223	259	144	129	198	117	213	98.0
Hurricane + Zephyr	2016	669	1,433	850	500	682	256	313	2,180	2,032	1,663	846	317	802	310	431	1,286	678	491	1,363	162	748	834	664	674	347
Zen1	2017	377	1,119	401	517	432	218	249	1,398	607	671	295	233	345	129	193	500	320	451	782	199	253	780	290	403	147
Kaby Lake (Skylake variant)	2017	348	1,059	382	381	311	192	250	1,222	605	636	302	200	396	146	181	547	297	360	740	171	207	808	274	322	131
Coffee Lake (Skylake variant)	2017	459	1,338	484	498	412	257	323	1,584	779	824	387	260	509	190	242	736	393	479	944	235	302	1,077	353	432	186
Monsoon + Mistral	2017	1,080	1,622	1,142	914	1,063	510	714	2,705	2,526	2,317	1,099	531	1,659	471	630	1,803	1,130	1,083	1,849	422	1,205	1,184	965	1,354	660
Coffee Lake refresh (Skylake variant)	2018	277	875	276	298	269	145	194	962	479	505	241	154	324	122	141	451	192	284	584	134	164	646	217	255	140
Vortex + Tempest	2018	1,415	2,185	1,180	1,096	1,350	601	833	3,678	3,105	3,222	1,583	665	2,362	594	867	2,677	1,676	1,315	2,527	513	1,686	1,553	1,406	1,664	981
Zen2 "Matisse"	2019	504	1,637	546	670	568	257	368	1,900	865	951	340	304	622	158	275	634	418	507	1,097	242	333	1,155	386	579	168
ARM Cortex A72	2019	441	1,122	580	548	453	241	332	1,274	1,213	861	399	275	446	119	227	456	398	433	634	249	443	888	332	455	179
Lightning + Thunder	2019	1,192	1,765	1,099	907	1,156	566	661	3,341	2,607	2,680	1,258	593	1,544	471	695	2,155	1,350	1,263	2,194	453	1,546	1,302	1,387	1,420	694
Zen3 "Vermeer"	2020	548	1,575	526	671	582	327	386	2,083	963	979	426	312	620	210	318	643	427	630	1,208	309	364	1,215	418	602	166
Comet Lake (Skylake variant)	2020	194	582	209	218	199	111	141	645	316	338	143	110	196	78.9	103	303	166	203	404	95.0	118	444	147	181	102
Firestorm + Icestorm	2020	912	1,729	1,164	602	767	364	380	3,279	2,664	2,781	1,157	410	1,261	410	645	1,896	940	618	2,087	208	981	1,121	1,168	787	466
Firestorm + Icestorm	2020	1,496	2,600	1,591	1,006	1,260	635	722	4,618	3,897	3,485	1,675	676	2,001	624	984	3,117	1,465	1,347	3,154	463	2,089	1,884	1,667	1,594	797
Rocket Lake / Cypress Cove / Sunny Cove	2021	314	914	334	325	316	170	212	1,045	520	541	294	152	314	154	179	531	289	301	694	136	171	731	250	285	170
Alder Lake / Golden Cove	2021	325	1,003	361	420	343	155	206	1,172	564	631	294	193	327	162	164	442	256	330	721	136	145	902	252	329	143
Avalanche + Blizzard	2021	1,556	2,642	1,596	1,112	1,367	682	793	4,810	3,836	3,427	1,652	685	2,210	653	972	3,289	1,566	1,491	3,377	541	2,106	1,927	1,615	1,621	774
Zen4 "Raphael"	2022	692	2,246	650	850	719	412	470	2,429	1,168	1,168	560	406	845	269	391	801	578	771	1,382	347	400	1,438	545	720	326
Raptor Lake / Golden Cove	2022	324	884	332	396	326	155	211	1,218	605	616	312	178	330	160	172	420	275	321	681	137	152	954	273	342	131
Avalanche + Blizzard	2022	582	1,134	755	394	483	229	247	2,065	1,691	1,781	756	265	764	273	404	1,191	610	426	1,340	119	580	703	730	541	297
Everest + Sawtooth	2022	2,052	3,664	2,094	1,287	1,756	803	946	6,699	5,430	5,313	2,246	855	2,748	804	1,305	4,566	2,158	1,789	4,723	594	2,571	2,592	2,642	2,271	1,118
Raptor Lake Refresh / Golden Cove	2023	429	1,386	461	549	457	213	294	1,599	760	824	425	211	423	238	233	551	347	432	883	176	191	1,228	349	419	165
Macbook Pro M3 (2023)	2023	1,074	1,905	1,230	742	975	466	475	3,512	2,750	3,050	1,224	476	1,341	517	700	2,800	1,085	823	2,505	258	1,211	1,187	1,364	1,017	546
ARM Cortex A76	2023	511	1,628	469	634	516	223	354	2,174	1,108	1,075	448	250	507	165	269	667	391	412	1,102	254	527	1,208	386	495	210
Zen5 "Granite Ridge"	2024	561	1,530	494	657	587	278	391	2,051	960	1,098	537	320	587	219	336	618	531	575	1,106	255	271	1,171	502	639	348
(?)	2024	1,440	2,380	1,577	929	1,494	593	657	4,714	3,465	4,641	1,712	651	1,886	688	896	3,839	1,500	1,029	2,950	335	1,584	1,456	1,912	1,800	688
Apple M5	2025	1,985	2,919	1,904	1,399	2,155	914	1,035	5,867	4,196	5,830	2,298	876	2,460	918	1,188	4,826	2,099	1,518	4,484	545	2,730	1,921	2,543	2,306	986

Top 15 most efficient machines measured

#◂	CPU◂	Family◂	TDP W◂	Score MT◂	MT / Watt◂
1	Samsung Galaxy S21 / Qualcomm Snapdragon 888		5.00	11,691	2,338
2	 Apple A16 / iPhone 14 Pro	Everest + Sawtooth	6.00	12,314	2,052
3	Intel(R) Core(TM) Ultra 7 258V		17.0	33,948	1,997
4	 Apple M5 / iPad Pro	Apple M5	15.0	29,775	1,985
5	 Apple M4 / iPad Pro	(?)	15.0	27,554	1,837
6	Samsung Galaxy S9 / Exynos 9810		5.00	7,986	1,597
7	Samsung Galaxy S26 / Qualcomm SM8850-AC		15.0	23,911	1,594
8	 Apple A15 / iPhone 13 Pro	Avalanche + Blizzard	6.00	9,333	1,556
9	NVIDIA Grace		200	311,082	1,555
10	 Apple A14 / iPhone 12	Firestorm + Icestorm	6.00	8,974	1,496
11	 Apple A12 / iPhone XS	Vortex + Tempest	5.00	7,074	1,415
12	 Apple M1 / iPad Pro	Firestorm + Icestorm	15.0	20,350	1,357
13	 Apple M3 / iPad Air	Macbook Pro M3 (2023)	15.0	18,976	1,265
14	Samsung Galaxy S23 / Qualcomm SM8550-AC		12.0	14,654	1,221
15	 Apple A13 / iPhone 11	Lightning + Thunder	6.00	7,150	1,192

Multi-core scaling efficiency

MT/ST speedup divided by logical core count (4+ core machines). 100% = perfect linear scaling. SMT/Hyper-Threading counts as cores here, which is why 8C/16T desktops settle near 50–60%: the second thread of a core only adds a fraction of a core. Memory-bandwidth-bound kernels also cap scaling.

Best scaling

CPU	Cores	MT/ST	Eff.
NVIDIA Grace	72	x72.9	101%
Ampere Altra Q80-30	80	x80.0	100%
AMD Phenom(tm) II X4 955 Processor	4	x4.00	100%
Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz	4	x4.00	100%
Raspberry Pi 2 / BCM2836 / Cortex A7 @ 900 MHz	4	x3.96	99%
Intel(R) Core(TM) i3-9100 CPU @ 3.60GHz	4	x3.81	95%
Raspberry Pi 3 / BCM2837 / Cortex A53 @ 1.20GHz	4	x3.79	95%
Intel(R) Core(TM) i5-4570 CPU @ 3.20GHz	4	x3.78	95%

Worst scaling

CPU	Cores	MT/ST	Eff.
Intel(R) Core(TM) i5-1235U	12	x3.63	30%
Intel(R) Core(TM) i9-13900	32	x10.4	32%
Samsung Galaxy S24 / Exynos 2400	10	x3.36	34%
Samsung Galaxy S25 / Qualcomm SM8750-AC	8	x2.77	35%
AMD Ryzen 5 5600U with Radeon Graphics	12	x4.41	37%
 Apple A11 / iPhone X	6	x2.29	38%
Samsung Galaxy S21 / Exynos 2100	8	x3.09	39%
 Apple M3 / iPad Air	8	x3.11	39%