Fixing the Performance of my custom RTX 3080 Ti Laptop (+ Sequre HT140 hot tweezers in action)
In my previous videos I have replaced the soldered GPUs in two of my laptops with higher speced models. This one got a 3080 Mobile inside, and this one a 3080 Ti Mobile, both with 16 GB video memory. However, I found that the 3080 Ti laptop does not perform too well compared to the 3080 laptop in classic rasterized benchmarks like 3Dmark TimeSpy. Especially at the lower power target range. My main suspicion was that the bigger GA103 die is just not made for low power targets below 150W and can’t really spread it’s wings in any laptop. As someone stated in the comments, it was
originally made to be the RTX 3080 desktop GPU die, but got stockpiled and cut down to make an appearance as a mobile GPU instead. ... the 3080 Ti Mobile and RTX A5500 Laptop GPU. Nevertheless, I intended to find a fix for the performance issues. Or at least improve the performance a little. And I got some things in mind for today’s video. Step A: I want to do a capacitor mod on the GPU voltage rails, to help the VRMs to maintain a stable voltage and current supply. Because the 3080 Ti draws a lot more current at the same TGP, than the 3070 I had in this laptop originally.
And … Step B: Try out another memory strap setting on the GPU, to potentially get rid of the weird memory speed hopping behavior. Step C: Use another vBIOS and see if there are some major differences between different brands. And finally … Step D: Investigate if the reported TGPs, so power telemetry, is even correct at all. Plus, potentially fixing it. To get all done I mainly used one tool, which made things so much easier, faster and a lot less stressful for the components on the laptop’s mainboard. My new Sequre HT140 hot tweezers. Now, previously I was using 938D hot tweezers, but as you might have seen in my videos, they struggle to grab on small SMD components. What makes things worse is their tips are too big for my use and are not exchangeable. So, if their coating is damaged you can just throw them away.
Plus, they are kind-of wobbly and the tips can become loose during soldering. Compared to them, the HT140 feel like a real upgrade. They are more sturdy and got JBC style C210 exchangeable tips, solving the e-waste problem of my old hot tweezers.
Plus, the C210 style tips allow them to heat up to soldering temperature in just 3 seconds. They might be quite small, but powerful with up to 140W heating power. Compared to my old hot tweezers the HT140 are much better suited for SMD rework and should be a perfect match for swapping and adding some SMD capacitors in this video. Besides the hot tweezers I will use some flux, regular tweezers and a tablet screen repair preheater to soak the board with some heat, which will make the soldering on the larger caps much easier. To understand the purpose of the capacitors in
the GPU voltage rails, I think it might help to visualize my laptop’s GPU VRMs, first. Basically, those are the important bits and pieces of the two essential GPU voltage rails. The 19V of the power supply charge the input filtering capacitors and the power stages chop those 19V into small 19V spikes using PWM. Finally, the inductors and output capacitors filter those spikes to a steady 0.7 to 1V voltage. Which voltage exactly is determined by the GPU to hit a certain power target. Every time the GPU switches to a new computational
task the load on the VRM changes, though, which introduces voltage over- and undershoots. A capacitor mod has the goal to reduce those voltage instabilities. In the original configuration the laptop comes with quite some bulk capacitance on the GPU core rail already. Which you can think of as a big, but quite slow type of capacitor. The laptop’s mainboard uses polymer
caps with 330uF and 9mOhm ESR quite heavily. On the GPU core voltage rail it got 9 of them at the VRM output and two under the GPU core. To improve things, I made use of the three empty capacitor pads at the VRM output and soldered in three 470uF polymer caps with 6mOhm ESR. Additionally I have replaced the two polymer caps under the GPU with the highest end ones I was able to find. With 560uF and just 3mOhm ESR. All in all I increased the capacitance by 47% at the VRM output and by 70% under the GPU, including a reduction of the ESR. Of course, as I said that’s just bulk capacitance. In theory that should give
me no real benefit in case of sudden load spikes. But, that’s what the small ceramic caps are for. They might not have a lot of capacitance, but they can handle high frequencies much better. In the original configuration there were 18x 22uF at the VRM output, and now there is another 6x 47uF caps added on top of them. Under the GPU there were 27x 22uF caps and now there are another 27x 47uF caps added on top, or next to them. For the fast ceramic caps that’s 71% more capacitance at the VRM output and a whopping 200% more capacitance under the GPU.
Honestly, I think that’s just overkill and would not have been necessary. But it is a good measure to see if it does have an effect at all. If there are not much gains you will know a cap mod is not worth it for sure. Not to mention the high costs of those caps. The high-end Panasonic GX polymer caps under the GPU core cost more than a dollar each. And the small ceramic caps roughly 15 cent each. But of course, you need a lot of them. To round things up I replaced the 19V input bulk caps with much better ones and added one more similar ceramic cap to the VRM powerstage inputs.
A small addition, but it should help to keep the input side more stable as well. To measure the effect of the cap mod I hooked up my oscilloscope to the same two measuring points before and after the modification. One of them is under the GPU and the other one under the VRAM chip, which is the furthest away from the VRM. Unfortunately, my oscilloscope is a quite cheap one, which is unable to record readings continuously. So, I had to record a video of the screen
which showed the values of interest and wrote them down later by hand. Running Furmark 2 at locked 687mV resulted in up to 160W TGP. Measured using the same oscilloscope settings as buildzoid usually uses in his cap-mod videos, namely a 5ms per division time frame and a 20MHz bandwidth limit, I got 148mV peak-to-peak before the cap mod and just 116mV after. So,
in total a 21% improvement over stock. Compared to his experiments with desktop GPUs that’s a quite good result. I ran some other tests using different settings, but what I think might be the most valuable is a comparison of a Time Spy benchmark run at the same timestamp, before and after. Please note that I have used a much bigger 200ms time frame, which increased the peak-to-peak numbers quite a lot. The results are noticeable better with the cap mod. I got 291mV peak-to-peak before the cap mod and just 213mV after, which is a reduction by roughly 27%.
The results sound quite nice, but unfortunately, they did not improve the stock performance. However as seen with other people’s cap mods it improved overclocking to some extend. Before the cap mod I could run the Time Spy benchmark with +255 on the core most of the time. Now, with the cap mod +255 was rock solid, and even +270 finished most of the time. Rarely even +285 was able to finish the benchmark,
but with some visual glitches. From an overclocker’s perspective the gains are a “mildly” good result, but for real use it doesn’t really make a difference. If +210 on the core was stable for daily use before, now +225 could be. Which comes down to an underwhelming improvement of at maybe 1% in games, in the best case. So, let’s stop wasting time and go on with the next idea.
Coming to the idea that there might be a better memory strap setting, which does not change clocks rapidly, as seen with the 3080 Ti Mobile. I was planning to change the memory straps to the config the 3080 Mobile has, because it behaves quite well. And according to GPU-Z the strap tables of both GPUs are the same. So, I used the same strap resistor configuration as seen with the 3080 Mobile. To keep things short however, it doesn’t make a difference. Despite Nvidias’ own diagnostic tool NVMT show the
changes to the strap configuration as well. And because I saw no performance differences either, it looks like the 3080 Ti Mobile vBIOSes use at least roughly the same memory settings for both strap configs. So, let’s explore another idea. Speaking about different vBIOSes I can make things short again and say that none of them made a noticeable difference in performance outside of tolerances. After those two very short summarized dead ends, I can finally tell you that there is an explanation for the relatively low performance. And even better, it is fixable with some work on the laptop’s mainboard. In simple words: the telemetry is broken.
Not just for the 3080 Ti Mobile, but also for the 3080 Mobile. For example, when you set 115W TGP (and yes, GPU-Z reports 115W under load as well) then the 3080 Ti Mobile is drawing only 105W, while the 3080 Mobile draws 109W. At 150W on the other hand the 3080 Ti Mobile draws roughly 142W, while the 3080 Mobile turns into a little liar, drawing 156W. Which is an overall delta of 14 Watts!
Responsible for the power monitoring in most laptops and also my one is a chip called NCP45495. It does read the voltages and current at two shunt resistors and reports the readings directly to the Nvidia GPU. Which then is able to limit its power draw according to the specifications defined in the vBIOS. Those readings are the very same values that show up in GPU-Z. So, I learned I couldn’t trust the original telemetry. To get hopefully more correct TGP readings I had to hook up an Elmor Labs EVC2X to both of my laptops as shown in some earlier videos. Using the TGP readings directly from the VRMs show that the 3080 Ti Mobile could improve it’s standing in comparison to the 3080 Mobile.
In the Time Spy benchmark results you can see the 3080 Ti Mobile is still losing to the 3080 Mobile at the lower end, but is able to outperform it earlier at the top end compared to my previous results represented with a dashed line. Plus, the curve of the 3080 Ti Mobile is steeper and it basically looks like it would keep going with increasing TGP, while the 3080 Mobile’s curve seems to stagnate at the top end. The new results are not groundbreaking, but are more in line with what to expect from those two GPUs. Oh, and please note that I couldn’t retest the 3070 Mobile results, because I actually desoldered both 3070 Mobile GPUs from my two laptops and replaced them with the better performing ones. Oops. So please keep in mind. In conclusion, the performance graph comparing
the 3080 Mobile with the 3080 Ti Mobile shown in my previous video was probably not reflecting the reality. But bear in mind that I can’t tell if the VRM readings are more accurate than the ones reported by the dedicated power monitoring IC. At least I can say the VRM readings were consistent from run to run and the resulting performance data seems to make more sense this way. But I could be completely wrong about this. After all it’s very likely both power reading methods are skewed one way or the other. However, if the new readings - directly from the VRM - are closer to reality, I wonder why the GPU’s own power monitoring is wrong and IF it is just this laptop model, or a more common behavior.
To fix the performance of the 3080 Ti Mobile I tested two promising modifications. Both of them are shunt mods, BUT not in the traditional way. To better understand what I am about to show you I’ve created a very simplified simulation of a typical power monitoring circuit used in Laptops with Nvidia Mobile GPUs. At the top, there are 20V coming in, usually provided by the laptop’s PSU. It is used to power the two biggest GPU VRMs. One for the GPU core and one for video memory. In the case of RTX 3000 Laptops the GPU core VRM comes right after the primary shunt resistor. And the Video Memory VRM comes right after the secondary one.
In summary the primary shunt does measure the total “Board Power Draw”, as it is called in GPU-Z. The secondary shunt does represent the “PWR_SRC Power Draw”. And to get the “GPU Chip Power Draw” the GPU simply subtracts the secondary from the first reading. As you may know, a shunt mod usually does refer to replacing the current measurement shunt resistor with a smaller one to trick the GPU into thinking it does draw less current and power than it actually does. The GPU is programmed to hit a power limit though,
so it fills the power gap by bumping up the voltage and clocks, until it hits the – now modified - power limit again. In this example, I chose a 4 milliOhm shunt resistor over the regular 5 milliOhm one, and I get 5/4 times, so 1.25 times the power draw. Instead of 115W the GPU runs with roughly 140W, while still pretending to consume 115W in GPU-Z. However, this method is very coarse and does not allow fine adjustments to the power draw. Luckily, there are more ways to manipulate the power readings.
As you have seen before, the 3080 Ti Mobile in my laptop does not hit the power draw it is supposed to have. On average it was roughly 5% behind the set TGP. To fix that I made use of that little input filter which you can find near the power monitoring IC. In this example it is set up with 220 Ohm resistors and a 22nF filtering cap to smooth out the voltage and current readings for the GPU. To bump up the power draw of the Nvidia GPU by 5% I add a little resistor in parallel to the capacitor to drain it’s charge and reduce the voltage differential. This way the power monitoring IC submits 5% less current draw to the GPU. Which basically means the GPU will have 5% more headroom to hit the power limit. The other method I was thinking about is a little stupid, but it caught my curiosity, because I thought it could have been a potential fix for the weird memory speed hopping of the 3080 Ti Mobile as shown in my previous video. The idea is to simply cut off the secondary
shunt from the output of the primary one and connect it to the power source as well. This way any power deviations from frequent memory speed switching, as this GPU does regularly, is excluded from the total “board power draw”. The effect I hoped for are steadier GPU core and memory clocks. And this shunt mod was the first one I tried out, simply because it’s effects sounded quite promising.
To pull it off I had to cut off the big trace connecting the input side of the second shunt leading to the video memory … And route a cable directly from the power source to the input of the second shunt. Later, after verifying it’s use, I added a little 2-way switch to the board to quickly switch between the classic shunt arrangement and the “detached-shunt mod”. Of course it does look a bit janky, but let me tell you, it makes life way easier when testing both setups back and forth. To my surprise the performance actually did improve a little by up to 4% in the best case. BUT only in some TGP ranges. After further investigation it became clear that the performance gain was the highest for TGP settings where the video memory speed was changing quite a lot during the benchmark. The more you move the TGP slider to the top end the more the GPU preferred the highest memory speed. At the lowest TGP settings it was basically the same,
but of course the GPU preferred the lowest memory speeds instead. So it looks like the more the 3080 Ti Mobile tends to hop between different memory speeds, the more you gain with the detached-shunt mod. Oh and of course, I took the power draw data directly from the VRMs again. Otherwise, these graphs would make no sense at all. And finally, it was about time to address the 5% TGP gap. The fix made use of the earlier
discussed input filter mod, to bump up and fine tune the shunt resistor readings. As mentioned, it is as easy as adding a resistor in parallel to the filter capacitor. The hard part definitely is the micro soldering. The components you see me working on in the footage are 0402 package parts measuring 1 by 0.5 Millimeter. As you can see in the table the power metering
adjustments work as expected. The TGP measured directly at the VRMs increased by roughly 5%. Of course this method is usable for much higher changes to the TGP as well. When I would swap the 8.2k Ohm resistor with a 2k Ohm one I would get a 20% higher power draw, instead of just 5%. But please note that the input filter is not standardized. As far as I know most laptops with Ampere GPUs came with 0 Ohm resistors instead of 220 Ohm. In this case you would have to swap them as well.
After all the effort the 3080 Ti Mobile is still trailing behind the 3080 Mobile in rasterization at lower TGPs. But at least it is faster in raytracing benchmarks, but not by crazy numbers. Let’s be honest, it is not getting any better, without pumping in more power.
Even Jarrod’s Tech showed in his 4090 Mobile review impressively, that the 3080 Ti Mobile just doesn’t scale too well. And his results are pretty much in line with mine. Personally, I must admit that I have invested way too much time into benchmarking this GPU, to find an explanation. However, I got one future project in mind that could benefit this GPU as well. I will let you know, when it is done. And finally, nothing more to say than: thanks for watching!
2025-05-16 14:29
