Discussion Performance per litre - Round 4

[morj]

Did you have to get rid of something for that?

Yeah, they got rid of the active SSD heatsink, so no big deal. SSD performance is not measured by PPL.

Also mad props to minisforum, they released basically an off-the-shelf solution with crazy high PPL. With such products available there's a really slim chance a build with top of the line GPU (say RTX 4090) can hit it.

Consider this: to get 2M PPL (assuming same CPU score of 12775) a build with a 4090 (which has around 33500 GPU score) must have a volume of approximately 1.6L:

log_1.5 ((12775 * 33500) / (100 * 2000000)) ~= 1.66

(12775 * 33500) / (1.66 ^ 1.5) / 100 ~= 2000984

For reference, Dell OEM 4070 GPU (smallest 4070) is around 1L in volume (with PCI slot excluded).

 

[msystems]

The new size is 0.198L (80x80x31)
The new PPL score is (12775 * 2688) / ((0.198 + 0.053) ^ 1.5) / 100 = 2,730,735

The first revision 1 of the 3D panels here:

Please show us a picture of it next to a soda can

 

[chinevo]

Congrats, impressive scoring!
How quiet is this tiny thing?

Surprisingly very quiet and completely silent when idle. I think it's due to liquid metal.
Watch this and this videos for more info about noise.

What you did is keeping the inner stock casing and replacing the top and bottom panels with thinner ones, right?

Only replaced the top and bottom panels

Stock vs Custom:

Did you have to get rid of something for that?

Yes, I got rid of the active SSD heatsink

source: https://www.techradar.com/pro/minisforum-mercury-em780-review

No DC barrel here, GaN USB-C charger fully compatible with this board.
USB-C 5A limit is respected with this low power 65W, but what about the problem with the USB Power Delivery reported earlier with such Gan chargers?
Does the board have an integrated 'USB-PD dummy trigger PCB'?
What do you guys think, @msystems, @SFFMunkee ?

Any usb-c chargers, that I tried, works well - apple, xiaomi, stock, rozetka.
But usb-c port on the front don't charge my wireless mouse or keyboard.

Please show us a picture of it next to a soda can

Yes, sure

 

[Jace]

@chinevo Have you thought about 3D printing a totally custom case and integrating the GaN charger into the EM780? Would that net you a few more points?

 

[morj]

By the way, this whole thing got me thinking about the PPL formula. Why are we measuring performance per cm^4.5?

Explanation: 1 liter is 1000 cm^3 (cubic centimeter). So the formula has Build Volume^1.5, which makes it (cm^3)^1.5 which is (cm^4.5). Performance per something that is over 4 dimensions

Moreover, for builds of, say 4L in volume we have 4^1.5 = 8, so actually 2 times the volume (of penalty) due to this exponent.

On the other hand for a build of 0.25L volume we have 0.25 ^ 1.5 = 0.125, so actually 2 times LESS the volume due to this exponent.

So basically the exponent in the formula IMMENSELY favors the < 1L builds due to the way exponents work. We are looking at like 4x advantage over a "normal" build.

EXAMPLES:


@petricor's build without the exponent: (34,565*33,304)/(5)/100 = 2302305.52

@chinevo's build without the exponent: (12775 * 2688) / ((0.198 + 0.053)) / 100 = 1368095.61753

Is this feature of the PPL formula intentional?

 

[REVOCCASES]

Good point, seems nobody expected sub 1L builds

Maybe it's time for a new round with update calculation?

 

[chinevo]

By the way, this whole thing got me thinking about the PPL formula. Why are we measuring performance per cm^4.5?

Explanation: 1 liter is 1000 cm^3 (cubic centimeter). So the formula has Build Volume^1.5, which makes it (cm^3)^1.5 which is (cm^4.5). Performance per something that is over 4 dimensions

Moreover, for builds of, say 4L in volume we have 4^1.5 = 8, so actually 2 times the volume (of penalty) due to this exponent.

On the other hand for a build of 0.25L volume we have 0.25 ^ 1.5 = 0.125, so actually 2 times LESS the volume due to this exponent.

So basically the exponent in the formula IMMENSELY favors the < 1L builds due to the way exponents work. We are looking at like 4x advantage over a "normal" build.

EXAMPLES:


@petricor's build without the exponent: (34,565*33,304)/(5)/100 = 2302305.52

@chinevo's build without the exponent: (12775 * 2688) / ((0.198 + 0.053)) / 100 = 1368095.61753

Is this feature of the PPL formula intentional?

Oh yes, the rating is unfair because @petricor is not in the first place

It's a joke. To put it seriously, the rating is really unfair cause we have performance ^ 2 (performance*performance) in the numerator and volume ^ 1.5 in the denominator. So, performance has more weight than volume. The fair formula, in my opinion, should look like <performance>^1/<volume>^1 (performance per liter). For example: (Multipier1*<GPU score> + Multipier2*<CPU score>) /<Volume>. Or separate ratings for GPU and CPU: <CPU score>/<Volume> and <GPU score>/<Volume>

For example, gamers have simple formulas
<Average FPS> / <Price> or <Performance> / <Watt>
rather than something like this:
<FPS in game1>*<FPS in game 2>/ Dollar ^ 1.5
or
<Performance1>*<Performance2>*<Performance3>/ Watt ^ 2

Good point, seems nobody expected sub 1L builds

Maybe it's time for a new round with update calculation?

"Competition ends at 25 entries (new round with new benches starts after)"

So, we need 6 more entries and go for round 5

 

[morj]

Oh yes, the rating is unfair because @petricor is not in the first place

I'm not saying the rating is unfair, I was just trying to understand how the minisforum PC beat is so easily. Based on the raw numbers, the performance is not proportional to the volume, but the rating is off the scale.

I mean, it's not fun when something you can just go and buy for 600 EUR blows everything custom (like @petricor or @guryhwa builds) out of the water.

 

[msystems]

I would assume the purpose of the exponent is to reward going smaller in general. If there are two builds which have proportionally the same performance, the exponent function causes the smaller one to receive a bonus. I find this is somewhat logical for ITX desktops because the smaller you go, the more inefficient your build is, because you have things which are standardized/ fixed in volume, requiring more extreme customization the smaller you go.

But indeed, when the total volume is less than 1, it causes a bit crazy results. It needs a clamp on the lower bound of the volume.

I'm not sure what the best mathematical approach is, but i propose that builds with a total volume under 1L should omit the exponent from their ppl calculation.

That would still give the build above a colossal 1.3M score, but not a broken one.

 

[r0ger]

Hi,

For small machines where the cooling system is shared by the CPU and the GPU, it is important to do the CPU and GPU benchmarks at the same time

In my nuc with passmark benchmark (setting duration very long), the cpu benchmark is 22,500pts, the igpu benchmark is 3,000pts, if I launch at the same time it drops to around 60%

 

[chinevo]

Hi,

For small machines where the cooling system is shared by the CPU and the GPU, it is important to do the CPU and GPU benchmarks at the same time

In my nuc with passmark benchmark (setting duration very long), the cpu benchmark is 22,500pts, the igpu benchmark is 3,000pts, if I launch at the same time it drops to around 60%

Important for what? To get a worse results? I prefer better

PS: maybe it can be helpful for somebody. Here you can download 3D model files of my custom EM780 panels

 

[r0ger]

it's sure, it suits you

but it's always interesting to know the results of the mini pc when using cpu+gpu in the same time

 

[chinevo]

it's sure, it suits you

but it's always interesting to know the results of the mini pc when using cpu+gpu in the same time

OK, I'll do the test tomorrow

 

[confusis]

ASRock DeskMini X600
Ryzen 5 8600G
2x 16GB Teamgroup Elite DDR5-5600

DeskMini X600 Dimensions - 155 x 155 x 80 mm (1.92L)
Power Brick: 120w GAN supplied with DeskMini: 65 x 98 x 23mm (0.14L)
Total volume: 2.06L

Score: ((12870 x 2372)/ (2.06 x 1.5)) /100
(30527640/ 3.09) /100
98,794.95

System pics will be in the upcoming review article

Forgot the GPU-Z instance but I'm sure this'll do.

 

[SFFMunkee]

ASRock DeskMini X600
Ryzen 5 8600G
2x 16GB Teamgroup Elite DDR5-5600

DeskMini X600 Dimensions - 155 x 155 x 80 mm (1.92L)
Power Brick: 120w GAN supplied with DeskMini: 65 x 98 x 23mm (0.14L)
Total volume: 2.06L

Score: ((12870 x 2372)/ (2.06 x 1.5)) /100
(30527640/ 3.09) /100
98,794.95

System pics will be in the upcoming review article

Forgot the GPU-Z instance but I'm sure this'll do.

View attachment 3359

CPU and GPU scores both seem low?
What's the RAM running at?

::EDIT:: Just tested my 5600G on my boring daily / stable settings - except I had to go back to BIOS and enable the iGPU and force it to be primary output.
Weirdly my CPU score is dreadful - maybe I should've waited a bit after restarting lolll

Spoiler: First run

Second run the CPU performance is still way lower than I'd expect (in my PPL test I got 8840 CPU points..?) but whatever - I'd have expected the 8600G to smash the 5600G.

Spoiler: Second Run

Ah well - we'll see what it can do when I get tinkering

 

[confusis]

CPU and GPU scores both seem low?
What's the RAM running at?

::EDIT:: Just tested my 5600G on my boring daily / stable settings - except I had to go back to BIOS and enable the iGPU and force it to be primary output.
Weirdly my CPU score is dreadful - maybe I should've waited a bit after restarting lolll

Second run the CPU performance is still way lower than I'd expect (in my PPL test I got 8840 CPU points..?) but whatever - I'd have expected the 8600G to smash the 5600G.

Ah well - we'll see what it can do when I get tinkering

5600, stock settings on this one.

 

[Steven]

Hi all,

We were talking about formulas and what would most accurately define the title "performance per liter". I've come up with a solution to this using the geometric mean (instead of the proposed arithmetic mean) of the performance numbers. Sqrt(cpu_r23*gpu_timespy)/Volume

The benefit of this is threefold: 1) there is no need for arbitrary multipliers to combine the scores 2) This ACTUALLY gives performance^1 and volume^1 as the label "PPL" claims. 3) the scores will now progress more linearly meaning numbers won't grow ridiculously large in short time, giving the formula longevity without a need to scale it.

I've run this new formula through all the available data for this round, here is how it looks:

Old #	Old Score	New PPL #	New PPL	Link	Cine R23	Timespy Graphics	Volume (with brick), L	CPU	GPU
#1	2,730,735	1	23,346	chinevo	12,775	2,688	0.251	AMD 7840U	AMD 780M
#2	1,915,262	2	10,806	guryhwa	65,000	13,000	2.69	AMD EPYC 7742	RTX 3070
#3	1,029,622	4	6,786	petricor	34,565	33,304	5	AMD 7950X3D	RTX 4090
#4	770,214	3	7,998	articnova	20,126	6,682	1.45	Intel 13900	RTX A2000
#5	504,016	6	4,748	petricor	27,554	20,451	5	AMD 5950x	RTX 3090
#6	419,894	7	4,479	AURMEND	25,367	14,966	4.35	AMD 3950x	RTX 3070
#7	374,244	8	4,362	REVOCCASES	16,140	17,653	3.87	AMD 7940HS	RTX 4070
#8	272,396	5	4,756	REVOCCASES	7,359	6,463	1.45	Intel 11700T	RTX A2000
#9	141,943	13	2,390	hereforthefeast	16,586	13,116	6.17	AMD 3900X	RTX 3070
#10	136,524	9	3,302	msystems	15,010	1,791	1.57	AMD 5700G	AMD Vega8
#11	127,017	10	3,161	dannycpham	18,590	13,386	4.99	AMD 5950X	RTX 3070
#12	114,540	11	2,782	Goatee	6,068	6,063	2.18	Intel 9600K	RTX A2000
#13	102,725	15	2,108	SFFMunkee	11,307	11,211	5.34	AMD 5600G	RTX 4060
#14	98,795	12	2,682	confusis	12,870	2,372	2.06	AMD 8600G	AMD 760M
#15	73,738	14	2,213	msystems	14,732	1,713	2.27	AMD 5700G	AMD Vega8
#16	70,696	16	1,885	Diamorif	6,359	8,761	3.96	AMD 3900	RTX 2070
#17	55,551	18	1,560	Damascus	5,053	5,246	3.3	Intel 6700	GTX 1070
#18	52,695	17	1,604	CubanLegend	6,098	7,438	4.2	Intel 7700K	GTX 1080
#19	23,754	19	903	timginter	6,797	8,679	8.51	AMD 3500X	RTX 2070
#20	14,339	20​	776	aromachi	2,985	6,486	5.67	AMD 1600	GTX 1660 Ti

 

[REVOCCASES]

Hi all,

We were talking about formulas and what would most accurately define the title "performance per liter". I've come up with a solution to this using the geometric mean (instead of the proposed arithmetic mean) of the performance numbers. Sqrt(cpu_r23*gpu_timespy)/Volume

The benefit of this is threefold: 1) there is no need for arbitrary multipliers to combine the scores 2) This ACTUALLY gives performance^1 and volume^1 as the label "PPL" claims. 3) the scores will now progress more linearly meaning numbers won't grow ridiculously large in short time, giving the formula longevity without a need to scale it.

so this new formula moves me up by 2 positions? I agree!

 

[msystems]

Hi all,

We were talking about formulas and what would most accurately define the title "performance per liter". I've come up with a solution to this using the geometric mean (instead of the proposed arithmetic mean) of the performance numbers. Sqrt(cpu_r23*gpu_timespy)/Volume

The benefit of this is threefold: 1) there is no need for arbitrary multipliers to combine the scores 2) This ACTUALLY gives performance^1 and volume^1 as the label "PPL" claims. 3) the scores will now progress more linearly meaning numbers won't grow ridiculously large in short time, giving the formula longevity without a need to scale it.

I've run this new formula through all the available data for this round, here is how it looks:

Old #	Old Score	New PPL #	New PPL	Link	Cine R23	Timespy Graphics	Volume (with brick), L	CPU	GPU
#1	2,730,735	1	23,346	chinevo	12,775	2,688	0.251	AMD 7840U	AMD 780M
#2	1,915,262	2	10,806	guryhwa	65,000	13,000	2.69	AMD EPYC 7742	RTX 3070
#3	1,029,622	4	6,786	petricor	34,565	33,304	5	AMD 7950X3D	RTX 4090
#4	770,214	3	7,998	articnova	20,126	6,682	1.45	Intel 13900	RTX A2000
#5	504,016	6	4,748	petricor	27,554	20,451	5	AMD 5950x	RTX 3090
#6	419,894	7	4,479	AURMEND	25,367	14,966	4.35	AMD 3950x	RTX 3070
#7	374,244	8	4,362	REVOCCASES	16,140	17,653	3.87	AMD 7940HS	RTX 4070
#8	272,396	5	4,756	REVOCCASES	7,359	6,463	1.45	Intel 11700T	RTX A2000
#9	141,943	13	2,390	hereforthefeast	16,586	13,116	6.17	AMD 3900X	RTX 3070
#10	136,524	9	3,302	msystems	15,010	1,791	1.57	AMD 5700G	AMD Vega8
#11	127,017	10	3,161	dannycpham	18,590	13,386	4.99	AMD 5950X	RTX 3070
#12	114,540	11	2,782	Goatee	6,068	6,063	2.18	Intel 9600K	RTX A2000
#13	102,725	15	2,108	SFFMunkee	11,307	11,211	5.34	AMD 5600G	RTX 4060
#14	98,795	12	2,682	confusis	12,870	2,372	2.06	AMD 8600G	AMD 760M
#15	73,738	14	2,213	msystems	14,732	1,713	2.27	AMD 5700G	AMD Vega8
#16	70,696	16	1,885	Diamorif	6,359	8,761	3.96	AMD 3900	RTX 2070
#17	55,551	18	1,560	Damascus	5,053	5,246	3.3	Intel 6700	GTX 1070
#18	52,695	17	1,604	CubanLegend	6,098	7,438	4.2	Intel 7700K	GTX 1080
#19	23,754	19	903	timginter	6,797	8,679	8.51	AMD 3500X	RTX 2070
#20	14,339	20​	776	aromachi	2,985	6,486	5.67	AMD 1600	GTX 1660 Ti

So this isn't better. Using the sqrt, it's compressing the higher performing builds by even worse. It's increasing the score gap between ultra small sbc and itx, when what was needed was to decrease this gap.

The original suggestion of changing the formula was to address for (1). the high relative advantage of using very small efficient ryzen mobile sbcs, and (2). under 1L it actually breaks the existing formula.

 

[REVOCCASES]

TBH, why are we not using a straight and simple calculation as it was done in PPL round one?



we are talking about Performance / Liter and not Performance / Square Liter, Squareroot Performance / Liter, or whatever...

maybe add some other benchmarks or rules for the next round but keeping it simple is often the best solution - IMHO