Support NeoGAF

[Lacix]

Company	Console	CPU description	GPU description	CPU flops	GPU flops	CPU + GPU flops
SONY	PS2	Emotion Engine (299MHz)	Graphics Synthesizer (147MHz)	6,1 GFlops	4,7 GFlops	10,8 GFlops
	PS3	CELL (IBM) (3.2GHz)	RSX (NVidia) (500, 550 MHz)	204,8 GFlops	172,0 GFlops	376,8 GFlops
	PS4	AMD Jaguar (1.6 GHz)	AMD GPGPU 18 CU (800 MHz)	102,4 GFlops	1 843,2 GFlops	1 945,6 GFlops
	PS4 PRO	AMD Jaguar (2.1 GHz)	AMD GPGPU 36 CU (911 MHz)	134,4 GFlops	4 197,9 GFlops	4 332,3 GFlops
	PS5	AMD Zen 2 (3.5 GHz)	AMD RDNA2 36 CU (2230 MHz)	448,0 GFlops	10 275,8 GFlops	10 723,8 GFlops
Microsoft	XBOX	Pentium 3 (733 MHz)	Nvidia Geforce 3 (233 MHz)	1,5 GFlops	5,8 GFlops	7,3 GFlops
	XBOX 360	Xenon (IBM) (3.2GHz)	Xenos (AMD) (500MHz)	115,0 GFlops	240,0 GFlops	355,0 GFlops
	XBOX ONE	AMD Jaguar (1.75 GHz)	AMD GPGPU 12 CU (853 MHz)	112,0 GFlops	1 310,2 GFlops	1 422,2 GFlops
	XBOX ONE S	AMD Jaguar (1.75 GHz)	AMD GPGPU 12 CU (914 MHz)	112,0 GFlops	1 404,0 GFlops	1 516,0 GFlops
	XBOX ONE X	AMD Jaguar Evolved (2.3 GHz)	AMD GPGPU 40 CU (1172 MHz)	147,2 GFlops	6 000,6 GFlops	6 147,8 GFlops
	XBOX SERIES X	AMD Zen 2 (3.8 GHz)	AMD RDNA2 52 CU (1825 MHz)	486,4 GFlops	12 147,2 GFlops	12 633,6 GFlops
Nintendo	Wii	Broadway (IBM) (729 MHz)	ATI HollyWood (243 MHz)	8,8 GFlops	12,2 GFlops	20,9 GFlops
	Wii U	Espresso (IBM) (1,243 GHz)	ATI Latte (550 MHz)	14,9 GFlops	352,0 GFlops	366,9 GFlops
	Nintendo Switch	ARM Cortex-A57 (1,78 GHz 3 Core)	Nvidia Tegra X1 (921 MHz)	21,4 GFlops	471,5 GFlops	492,9 GFlops

Calculations:

Year	Console	Single precision floating point instructions per seconds	Notes
CPUs
2000.03.04. JP	Playstation 2	Emotion Engine = (VPU0 ( 2 * 4 FMAC Flops + 1/7 * FDIV Flops ) + VPU1 ( 2 * 5 FMAC Flops + 2/7 * FDIV Flops ) + FPU (1 Flop + 1 * FDIV Flops)) * 299 MHz = 6,128 GFLOPs	FMACs - 1 addition and 1 multiplication in 1 cycle FDIVs - division in 7 cycles
2001.11.15. NA	XBOX	Pentium 3 = 2 Flops * 733 MHz = 1,466 GFlops
2005.11.22. US	XBOX 360	Xenon (IBM) = (3 PPE *(4 FPU Flops + 2 * 4 VMX128 Flops) * 3.2 GHz) = 115 GFlops
2006.11.11. JP	PS3	CELL (IBM) = (1 PPE * (4 FPU Flops + 4 VMX Flops) * 3.2 Ghz) + (7 SPE * 8 Flops * 3.2 Ghz) = 204,8 GFlops	(8 SPEs, 1 SPE disabled)
2006.11.19. US	Wii	Boradway (IBM) = (3 cores * 4 FPU Flops) * 729 MHz = 8,75 GFlops
2012.11.18. US	Wii U	Espresso (IBM) = (3 cores * 4 FPU Flops) * 1,243 Ghz = 14,916 GFlops
2013.11.13. NA	PS4	Jaguar (AMD) = (8 FLOPs * 8 cores) * 1.6 GHz = 102,4 GFlops
2013.11.22.	XBOX ONE	Jaguar (AMD) = (8 FLOPs * 8 cores) * 1.75 GHz = 112 GFlops
2016.11.10.	PS4 Pro	Jaguar (AMD) = (8 FLOPs * 8 cores) * 2.1 GHz = 134,4 GFlops
2017.03.03.	Nintendo Switch	Cortex-A57 (ARM) = (3 cores * 4 flops ) * 1.78 GHz = 21,4 GFlops
2017.11.10.	XBOX ONE X	Jaguar Evolved(AMD) = (8 FLOPs * 8 cores) * 2.3 GHz = 147,2 GFlops
	PS5	Zen 2(AMD) = (16 FLOPs * 8 cores) * 3.5 GHz = 448 GFlops
	XBOX SERIES X	Zen 2(AMD) = (16 FLOPs * 8 cores) * 3.8 GHz = 486,4 GFlops
GPUs
	Playstation 2	Graphics Synthetiser = 16 pixel pipeline * 2 flops * 147 MHz = 4,704 GFlops
	XBOX	Nvidia Geforce 3 = (4 pixel pipeline * 5 Flops + 1 vertex * 5 Flops) * 233 MHz = 5,825 GFlops
	XBOX 360	Xenos (AMD) = ( 48 shader processors * 10 Flops ) * 500 MHz = 240 GFlops	64 shaders, 16 shaders disabled
	PS3	RSX (NVIDIA) = ( 24 parallel pixel-shader * 10 Flops ) * 550 MHz + (8 parallel vertex-shader pipelines * 10 Flops) * 500 MHz = 172 GFlops	28 pixel shaders, 4 pixel shader disabled
	Wii	ATI Hollywood (AMD) = (4 Pixel shaders * 10 Flops + 1 Vertex shader * 10 Flops) * 243 MHz = 12,15 GFlops
	Wii U	ATI Latte (AMD) = (5 CU * 64 cores * 2flops) * 550 Mhz = 352 GFlops
	PS4	GPGPU (AMD) = (18 CU * 64 cores * 2 flops) * 800 MHz = 1843,2 GFlops
	XBOX ONE	GPGPU (AMD) = (12 CU * 64 cores * 2 flops) * 853 MHz = 1310,208 GFlops
	PS4 Pro	GPGPU (AMD) = (36 CU * 64 cores * 2 flops) * 911 MHz = 4197,9 GFlops
	Nintendo Switch	Tegra X1 (NVIDIA) = (1 FP 32 * 2 FMA * 256 CUDA) * 921 MHz = 471,5 GFlops
	XBOX ONE X	GPGPU (AMD) = (40 CU * 64 cores * 2 flops) * 1172 MHz = 6000,64 GFlops
	PS5	GPGPU (AMD) = (36 CU * 64 cores * 2 flops) * 2230 MHz = 10275,84 GFlops
	XBOX SERIES X	GPGPU (AMD) = (52 CU * 64 cores * 2 flops) * 1825 MHz = 12147,2 GFlops

Sheets about console performances in flops starting from PS2.
Feel free to comment to fix possible errors in calculations.

 

[IbizaPocholo]

Console Power Comparison Chart

The Xbox Series X is packing some serious power.

thegamingsetup.com

 

[Radical_3d]

Something is wrong with your OG numbers. In the graph it aligns with the difference versus the PS2, but in the numbers is even weaker and that’s not possible.

 

[DESTROYA]

Who cares about flops when all you need is good quality games.

 

[deriks]

Console Power Comparison Chart

The Xbox Series X is packing some serious power.

thegamingsetup.com

 

[Tamy]

You should take into account that PS5 is using a variable frequency, so that peak TFLOP number can not be reached at all times.

 

[Bladed Thesis]

You should take into account that PS5 is using a variable frequency, so that peak TFLOP number can not be reached at all times.

Wow, I've never heard this before. Huge if true. Massive. Bigly.

 

[Panajev2001a]

I am not sure the GS should be counted in floating point operations... pretty sure it worked in fixed point ints, have not read the HW docs in sooooo long.

 

[FranXico]

You should take into account that PS5 is using a variable frequency, so that peak TFLOP number can not be reached at all times.

I suggest you look up what "peak theoretical performance" really means.

 

[sonomamashine]

... care to explain ?

 

[Kev Kev]

i said ah-flip, flop, ah-flippy to the floppy, the flip flip a flop, and you don't stop, a rock it

 

[sonomamashine]

i said ah-flip, flop, ah-flippy to the floppy, the flip flip a flop, and you don't stop, a rock it

Nah it's S to the D, SS to the D, and you don't stop, a 5,5GB speed.

 

[Red_Driver85]

IIRC, but for Zen2 is 32flops, not 16. Doubled fpu, no?

 

[thicc_girls_are_teh_best]

Holy shit the CPUs in XBO in PS4 were mega trash . Like, I always knew it from 2013 but actually seeing the numbers next to 360 and PS3's CPUs it's kind of startling in a way.

That said, XBO and PS4 CPUs did benefit from a more refined multi-core approach and developers being more skilled in multi-core CPU development by the time they launched, let alone the years to follow, so even if the CPUs were worst on paper, in practice they were probably used at least as well as the 360 and PS3 CPUs, if not even more efficiently by a fair margin.

...just...not enough to negate the need for GPU asynchronous compute utilization the past gen, though.

 

[D.Final]

Console Power Comparison Chart

The Xbox Series X is packing some serious power.

thegamingsetup.com

Interesting

 

[Matreios Raven]

Was the PS3 RSX GPU actually really weak compared to the X360 Xenon GPU? I know they didn‘t share the same architecture so I don’t know if the two GF numbers are directly comparable.

 

[JLB]

this is the most lisp-friendly title ive ever saw.

 

[Fafalada]

I am not sure the GS should be counted in floating point operations... pretty sure it worked in fixed point ints, have not read the HW docs in sooooo long.

None of that gen console had Floating point on pixel operations. But it works in context of comparing compute throughput... sort of. GC geometry being fixed pipeline is cheating a bit.

 

[Bogroll]

12.155 TF for Xbox if you're counting the 10.28 TF for PS5.

 

[Bogroll]

I suggest you look up what "peak theoretical performance" really means.

Can you or anyone tell me why it was mentioned by Cerny boosted to 10.28TF, why not just say 10.28 TF ?

 

[FranXico]

Can you or anyone tell me why it was mentioned by Cerny boosted to 10.28TF, why not just say 10.28 TF ?

Cerny said they boosted the frequency up to 2.23GHz, therefore making the theoretical maximum 10.28TF.

It doesn't matter if the clock frequency is boosted temporarily or fixed, something theoretical is always theoretical. Read that as a throughput that could in theory under ideal conditions be reached, but in practice never, or hardly ever, will.

Sure, it can be used to compare and predict relative performances, but again, it's theoretical.

 

[Panajev2001a]

None of that gen console had Floating point on pixel operations. But it works in context of comparing compute throughput... sort of. GC geometry being fixed pipeline is cheating a bit.

Yeah, well this makes it look like PS2 was a lot faster than Xbox overall (by 3 GFLOPS) and just a bit less powerful GPU alone. Not really fair to NV2A programmable VS and PS pipelines IMHO .

 

[UnNamed]

IIRC before the PS360 Flops were intended as FP16 single precision and FP32 double precision. So those number has to be divided by 2.

 

[pawel86ck]

Xbox classic had modified geforce 3 with 2 vertex shaders, so in your calculation it should be 6.9 GFLOPs instead of 5.8 GFLOPs.

 

[Bogroll]

Cerny said they boosted the frequency up to 2.23GHz, therefore making the theoretical maximum 10.28TF.

It doesn't matter if the clock frequency is boosted temporarily or fixed, something theoretical is always theoretical. Read that as a throughput that could in theory under ideal conditions be reached, but in practice never, or hardly ever, will.

Sure, it can be used to compare and predict relative performances, but again, it's theoretical.

So it comes into play when games are being pushed. Eg DF videos when those games are being hitting 50fps to 60fps when it can be brute forced to to hit 60fps more often when cpu/bandwidth isn't a problem or indeed when it is a problem it can help out in those areas.

 

[rofif]

See this is shit metric. ps4 pro vs ps4 is not 100% faster. ps4 is 2k flops and ps4 pro over 4k flops.... yeah right.
ALso - 360 having less flops than ps3 but all multiplatform games looking better and running better on 360

 

[Bogroll]

See this is shit metric. ps4 pro vs ps4 is not 100% faster. ps4 is 2k flops and ps4 pro over 4k flops.... yeah right.
ALso - 360 having less flops than ps3 but all multiplatform games looking better and running better on 360

It can be used as a guide if they're the same architecture.
The Tf thing-Having a 150hp VW Golf doesn't make it twice as fast as a 75hp Golf.

 

[Omega Supreme Holopsicon]

Not really fair to NV2A programmable VS and PS pipelines IMHO

The PS are an advantage of the original xbox, but the vector units of the ps2 were said to be more advanced than the VS in the xbox.

 

[deriks]

... care to explain ?

This kind of meme, man

 

[Fafalada]

Yeah, well this makes it look like PS2 was a lot faster than Xbox overall (by 3 GFLOPS) and just a bit less powerful GPU alone. Not really fair to NV2A programmable VS and PS pipelines IMHO .

Well to begin with - the chart rates the Celeron in XB at half its actual Flops throughput. That and the GS number is inflated (only 8 pixel pipelines can do the full arithmetic ops with texture input).
It 'is' however a fact, that NV2A Flop rating (Vertex processing only) is in in fact lower than that of EE VUs - they have equivalent per-clock throughput but it's 300mhz vs 233mhz.
However we shouldn't forget that fixed-function stuff has a lot of impact on performance that programmable-compute chart cannot account for - that's why we end up with 'NV' vs 'AMD' flops nonsense.

See this is shit metric. ps4 pro vs ps4 is not 100% faster.

Yes it's more than 100% in GPU limited workloads - but it's close enough as an approximation.

 

[Fafalada]

Feel free to comment to fix possible errors in calculations.

Playstation 2

Graphics Synthetiser = 16 pixel pipeline * 2 flops * 147 MHz = 4,704 GFlops

Only 8 pixel pipelines can work with a texture input(making that into full featured MADD operations).
When operating with all 16, you are limited to essentially 'constant' input (vertex interpolated color) which really limits usefulness of the programmable arithmetic - not sure how you'd want to represent that in a peak-compute chart though.

XBOX	Pentium 3 = 2 Flops * 733 MHz = 1,466 GFlops

XBox CPU had a SIMD rated at 4 Flops/cycle - the correct peak rating would be 2.8GFlops.

PS3	CELL (IBM) = (1 PPE * (4 FPU Flops + 4 VMX Flops) * 3.2 Ghz) + (7 SPE * 8 Flops * 3.2 Ghz) = 204,8 GFlops

VMX had FMADDs - it's 8Flop/cycle throughput - so this would be 217 total.
Though I'm not a fan of counting PPE FP flops in addition to SIMD (goes for 360 CPU as well) as that was really not a practical use-case for these CPUs.

 

[Panajev2001a]

The PS are an advantage of the original xbox, but the vector units of the ps2 were said to be more advanced than the VS in the xbox.

Oh... they are, it is lovely to code for them (such fun on the low level PS2 Linux kit) . Then I am not a PS2 guru, I let Fafalada take the deserved mantle !

 

[carsar]

Would be interesting to see the real effectiveness of xbox 1/ps2 vs xbox 360/ps3 vs xbox one\ps4 flops.
You can find results of GeForce4 Ti 4200(similar to xbox) and 7800 gtx(similar to ps3) in same tests, 7800 vs 8800gtx, 8800(250gts) vs 7770(xbox 1)/7850(ps4).

 

[Journey]

Something is wrong with your OG numbers. In the graph it aligns with the difference versus the PS2, but in the numbers is even weaker and that’s not possible.

Because it's wrong lol

 

[Journey]

See this is shit metric. ps4 pro vs ps4 is not 100% faster. ps4 is 2k flops and ps4 pro over 4k flops.... yeah right.
ALso - 360 having less flops than ps3 but all multiplatform games looking better and running better on 360

Because it's wrong. This chart is more accurate

 

[nowhat]

I'm not really commenting on the thread - wanking over specs is fine and all, but not going to get into that debate.

But I must congratulate the OP, because I was really about to get a serious eye twitch about mismatching parentheses in the title, but upon closer inspection, they all match. So kudos OP.

 

[Pagusas]

This kind of meme, man

still dont get it

 

[Arkam]

Because it's wrong. This chart is more accurate

Yea but whoever lazily did the bar graph just eyeballed (or intentionally distorted). Measure the PS5 ans XSX bars and do the math.... pretty far off. Crazy that the graph even has hatch marks they could have used if just eyeballing. Luckily the numbers for the two are accurate at least.

 

[sonomamashine]

Yea but whoever lazily did the bar graph just eyeballed (or intentionally distorted). Measure the PS5 ans XSX bars and do the math.... pretty far off. Crazy that the graph even has hatch marks they could have used if just eyeballing. Luckily the numbers for the two are accurate at least.

Lol you measured it didn't you , this one is more accurate (10,3 V 12,1)

 

[Arkam]

Lol you measured it didn't you , this one is more accurate (10,3 V 12,1)

.....maybe lol It just looked off so I had to grab the tape measure and crunch the numbers.

 

[Journey]

Lol you measured it didn't you , this one is more accurate (10,3 V 12,1)

Hmm, why is the 12.1 below the 12 line? LMAO

 

[Journey]

.....maybe lol It just looked off so I had to grab the tape measure and crunch the numbers.

This one needs a tape measure as well, just look at 12.1 below the 12 line

 

[sonomamashine]

Hmm, why is the 12.1 below the 12 line? LMAO

This one needs a tape measure as well, just look at 12.1 below the 12 line

OH SNAP ! trying to stay humble i guess...