How PCBS Scores Computers

A little bird told me the formulas that PCBS uses to calculate the Time Spy scores, among other things. In this guide, I will do my best to explain the formulas, to hopefully help you overclock and predict performance.

Please note that, just like PCBS itself, this guide isn't accurate to real life, and there is a lot of nuance in real-world computers that simply isn't modelled in PCBS.

As of March 2025, this guide is a WIP. If you're reading this in ten years, I guess I never finished it, but y'all can probably derive the constants yourself if they're not stated here.

The CPU Formula

PCBS calculates a computer's CPU score by considering the clock speed, the memory speed, and the number of memory channels filled. If it's overheating, that will affect the score, but we're going to assume it isn't. Nothing else affects CPU performance. Notice that the maximum number of memory channels is a big factor here - for most CPUs it's two, for Threadrippers and LGA-2066 CPUs it's four. Putting four sticks in a dual-channel CPU will still only give 2 channel performance.

In real life, CPU performance is also affected by a number of other things, including RAM timings, clock floating, CCD/ring latency, threading... just go read an overclocking guide. CPUs are complicated, and no real-world test can perfectly measure all components and predict performance perfectly. PCBS approximates 3DMark's Time Spy, and Time Spy approximates real-world games.

First, the game looks up the CPU in a large database. It's looking for four values; The "Core Clock Multipler", the "Memory Clock Multipler", the "Memory Channel Multipler", and the "Adjustment". The full-ish tables can (probably?) be found at the bottom of this guide. For example, the values for an i9-10900K are:

CPU	Core Mult.	Memory Mult.	Memory Channel Mult.	Adjustment
i9-10900K	0.005652	11.57727	0.004724	-22.0717

The score is then calculated using the following formula;

CPU_score = (CPU_freq * core_mult  +  memory_freq * memory_mult  +  memory_channel_count * memory_channel_multiplier  +  adjustment) * 298 

...where all frequencies are in MHz (not GHz! Remember, 1000MHz = 1GHz)

For example, a 10900K running at 5.3GHz with 2 channels of 4800MHz RAM would look like;

(5300 * 0.005652  +  4800 * 0.004724  +  2 * 11.57727  +  -22.0717) * 298 = 13604.63334

...note that PCBS always rounds down.

You can see this score in the "CPU Score" part of the 3DMark result page.

The GPU Formula

The GPU formula is similar but a little more complicated. The GPU test also works by looking up the GPU in a big database and finding some values, BUT;

1: The GPU tests have seperate coefficients, so there are two tables
2: There are separate databases depending on whether you have one or two GPUs installed.

Otherwise, the formula is almost identical to the CPU formula. We're going to call the first test "T1" and the second test "T2", and I'll seperate it into dual-GPU and single-GPU columns. The values for dual-GPU ARN'T exactly double the single-GPU values. The different values are *usually* the same between different GPUs with the same cores, but I have noticed that they're *sometimes* different. I think they're all constant for the 3090s and the 6900XTs, though. You should check before relying on them by setting two GPUs to the same frequency and seeing if they get the same scores.

In real life, SLI/Crossfire rarely double performance. Unfortunately, rendering half the screen with each GPU still leads to stutters and hitches, and the GPUs need to communicate with each other, which slows things down a lot. Time Spy is unusual because dual-GPU improves scores greatly, while most actual games barely improve or simply don't work. PCBS does a good job of approximating Time Spy here. Personally, I strongly recommend that you, if you're considering SLI/Crossfire IRL, simply use one big GPU instead. My 4080 Super can beat dual 3090s in almost all workloads.

Here are the values for the 3090, for example. I'll show you the calculation below;

GPU	RTX 3090
T1 Core Mult.	0.025237
T1 Memory Mult.	0.035803
T1 Adjustment	32.83273
T2 Core Mult.	0.022912
T2 Memory Mult.	0.028006
T2 Adjustment	28.98821
T1 Core Mult +SLI.	0.049293
T1 Memory Mult.	0.101481
T1 Adjustment	-1.17649
T2 Core Mult.	0.063254
T2 Memory Mult.	0.082598
T2 Adjustment	-30.8004

Similar to the CPU, the GPU formula for each test is;

 testscore = core_freq * core_multiplier  +  mem_freq * mem_multiplier  +  adjustment

...but we're not done yet. We need to do this for both tests (T1 and T2) and then use THIS formula to bring them together;

328 / ( 1 / test_1  +  1 / test_2)

Yes, 328 divided by the sum of the two reciprocals. It's weird, I know. For example, for the above a 3090 at 1860MHz with 1219MHz memory:

328 / (
   1 / (1860 * 0.025237  +  1219 * 0.035803 + 32.83273) +
   1 / (1860 * 0.022912  +  1219 * 0.028006 + 28.98821) 
)

The Full 3DMark Formula

Now that we've calculated the scores for the CPU and the GPU (you can see both of those scores on the 3DMark result screen in-game), we just bring them together using;

1 / (0.85 / GPU_score  +  0.15 / CPU_score)

...which will give us our final score. The game rounds it down before giving you the number.

Have fun! I hear that some people are able to hit >41,000 using the 3960X + 6800XT.

CPU Data Tables

This part is a WIP. The full CPU data table is too large for me to manually edit into Steam, so I'm just going to post the most interesting CPUs here for now.

Multipliers	5950X	5900X	5800X	3960X	3990X	i9-10980XE	i9-11900KF
Core Mult.	0.001457	0.001072	0.002264	0.003832	0.004492	0.006234	0.004559
Channel Mult.	10.6221	7.714274	5.554834	2.259872	3.459974	3.990303	12.91
RAM Mult.	0.00598	0.004702	0.003082	0.008126	0.006017	0.005258	0.003981
Adjustment	-0.87348	8.196178	5.95047	-11.9601	-16.4578	-18.0401	-20.72

GPU Data Tables

This part is a WIP. The full GPU data table is too large for me to manually edit into Steam, so I'm just going to post the most interesting GPUs here for now.

Multipliers (1 GPU)	RTX 3090	RTX 3080Ti	RTX 3080	RX6900XT	RX6800XT
T1 Core Mult	0.025237	0.026233	0.029755	0.026054	0.026364
T1 VRAM Mult	0.035803	0.045869	0.000901	0.011659	0.000178
T1 Adjustment	32.83273	20.12464	44.22204	38.30114	54.67308
T2 Core Mult	0.022912	0.02305	0.025457	0.022238	0.021609
T2 VRAM Mult	0.028006	0.034954	0.000666	0.009601	0.000143
T2 Adjustment	28.98821	21.30461	39.45169	30.47505	46.61601
Multipliers (2 GPU)
T1 Core Mult	0.049293	0	0	0.15178	0.212636
T1 VRAM Mult	0.101481	0	0	0.00324	0.157119
T1 Adjustment	-1.17649	0	0	-161.062	-609.238
T2 Core Mult	0.063254	0	0	0.125149	0.2395
T2 VRAM Mult	0.082598	0	0	0.002499	0.03177
T2 Adjustment	-30.8004	0	0	-124.335	-440.674

Spreadsheet Formulas

These data tables are very handy when you combine them with spreadsheets, so here's a basic template. You're going to have to copy it over manually, sadly, since Steam doesn't let you copy tables.

Make a spreadsheet with this structure;

FILL IN THESE	Example 1
CPU Name	i9-10900K
CPU Freq.	5300
RAM Freq.	4800
# RAM Chan.	2
.
Core Mult.	0.005652
Channel Mult.	11.57727
RAM Mult.	0.004724
Adjustment	-22.0717
.
GPU Name	EVGA 3090
GPU Freq.	1800
VRAM Freq.	1219
.
T1 Core Mult	0.025237
T1 VRAM Mult	0.035803
T1 Adjustment	32.83273
T2 Core Mult	0.022912
T2 VRAM Mult	0.028006
T2 Adjustment	28.98821
.
DON'T EDIT THIS
CPU Score	<COPY IN HERE>
GPU Score	<COPY IN HERE>
Final Score	<COPY IN HERE>

...This should be in the top left of your spreadsheet, or the formulas won't work! The A1 square should say "FILL IN THESE".

Remember to pick the right table if you're using two GPUs!

Now replace those final score boxes with;

=298*(B3*B7+B4*B9+B5*B8+B10)
=328/(1/(B16*B13+B17*B14+B18)+1/(B19*B13+B20*B14+B21))
=0.85*B25+0.15*B24

More Info

There are some more tips I noticed while finding and testing these formulas;

- Temperature has no effect on stability, unless the CPU is overheating, unlike real life.
- The wattage of a CPU seems to depend on the "silicon lottery" a lot. I'll try to find that formula next, so stay tuned.
- The multipliers are constant for all GPUs using the same core, EXCEPT FOR SOME EXCEPTIONS. I'm 100% sure that they're all the same for all air-cooled cards in the above data table, though. I'm not totally sure about water cooled cards, but I think they are too.
- AIOs have the same cooling power as an air cooler of the same CFM rating. Radiators are 1.5x more effective.
- PSUs don't have any effect besides their power rating.
- Cases have permanent cooling values, but I'm not really sure how it works, and it's 0 for most of them.
- Cables, coolants, pipes, and SSDs have no effect.
- Fans seem to have a "pressure" factor, but I don't know if it affects airflow, I'm still looking at that.
- The scaling of the "silicon lottery" seem to be constant for all components. I'm still figuring out that part, but it seems like it's a multiplier applied to the normal clock speed of the component.
- The silicon lottery percentage is "5", but I don't know if that's a hard limit or a standard deviation (seems more likely).
- Air coolers and AIOs have no effect other than their normal CFM parameter.
- There seems to be a lot of cut content, including FM2/+/AM3+ CPUs, DDR3, more real-world games, Intel NUCs, etc. I bet the game could be modded to feature AM5/LGA-1851/etc.
- Room temperature is 21C, no matter what.
- You can't spend more than $4000 per day, and you can't overdraft more than $3000.
- Absolute maximum CPU and RAM voltages are 2.5V and 2.0V respectively.
- Components that overheat will ignore the 3DMark score calculations and return a fixed value.
- Overclocking components won't change the "Will It Run?" rating, even if you overclock them to be better than the minimum spec. Customers won't accept it.