Decided to post a major update on my Engineering Sample 3080Ti FE with 20GB VRAM. (My original post here).

Yes, I went back on a few things I originally said I would not do. Some of you will probably think I am crazy for modifying a rare ES card this aggressively. That is fair. But for me this card is as much a research platform as it is a collector piece.

The goal here was exploration. I learned a lot about its limitations and design in the process. And since everything I did to it worked, I figured it was worth documenting.

What Changed Since the Original Post

• Proper stack shunt mod

• Liquid metal on the die with added clamp pressure

• Upgraded from 12W/mK pads to a 20W/mK kit

• Added heatsinks to the outer shell

• Dedicated fan blowing directly across the added heatsinks

Shunt Mod
After extended benchmarking, I noticed it was performing closer to a 3080 than a 3080 Ti. Even with overclocking, it was clearly being held back by the power limit and the 320-bit memory bus. Given that AIB versions of this 20GB variant surfaced in certain regions with much higher limits, I knew the silicon itself had more headroom than the stock VBIOS allowed. Being an engineering sample, I was effectively stuck with a conservative limit.

So I researched and implemented a proper stack shunt mod. No liquid metal bridges. Clean resistor stacking.

Important context: this card uses an 18-phase core and 2-phase memory VRM layout. That is substantial. Even conservatively assuming 70A stages:

18 × 70A × ~1.08V ≈ 1360W theoretical capacity

Obviously that is not realistic sustained draw, but it makes one thing clear. The VRM is not the limiting factor here.

There are guides that suggest shunting every resistor on the board. That is unnecessary and can create instability. I only modified the shunts tied to the 12-pin main connector and PCIe slot power sense rails. I left the memory and VRM controller sense resistors alone. The shunt layout is similar to a 3090, but not identical.

The 12-pin FE connector is rated for 600W, with additional headroom available via the PCIe slot. My goal was to approach, not exceed, that combined safe envelope. I used 10 mOhm shunts on top of the 5 mOhm to effectively increase the power from reported by 1.5x.

I monitored:

• Reported power

• Connector temperature via thermocouple

• Stability under load

After the mod:

• Stock behavior reported around 320W, which is an actual pull of 480W

• With my stable OC, I saw brief peaks near 370W reported, likely actually ~555W

Connector temperatures remained reasonable under testing. Some of that draw is distributed through the PCIe slot, not solely the 12-pin. I kept constant oversight during all stress testing.

Power was no longer the bottleneck.

That led to the next limitation...


Liquid Metal
Once power limits were removed, the die began heat soaking and throttling.
Standard paste was no longer sufficient.

I applied liquid metal to the die. On first boot, the card ramped fans to 100% and lost video after POST. That indicated poor contact pressure.
I fabricated 0.5mm clamp washers and installed them on the die tension springs to increase mounting pressure. After reassembly:

• Idle temperature dropped to ~31C with fans at max

• Load thermals improved significantly

• Core throttling under heat soak was eliminated

Core thermals were solved.

Then the real design flaw showed itself...

Memory Overheating (And What It Revealed About This Card)
This card uses a 3080 Ti cooler and shroud. However, unlike a retail 3080 Ti, this PCB has 10 memory modules on the back of the board. The 3080 Ti cooler was never designed to efficiently cool fully-loaded rear memory.

The only cooling for those rear modules is:

• Thermal pads

• Backplate

• Indirect conduction into the main heatsink assembly

Under sustained load, especially ray tracing workloads, memory junction temperatures climbed into the 100C to 102C range after heat soak. I also began seeing artifacting. Clearly not sustainable.

First attempt was upgrading to 20W/mK pads. That improved heat transfer from chip to backplate, but the heat had nowhere to dissipate. The backplate itself became heat saturated. A fan alone on the backplate was not enough. The only real solution was to increase surface area.

I attached self-adhesive aluminum heatsinks directly to the outer shell and positioned a dedicated fan to push fresh air across them. Yes, this is semi-permanent. Yes, it alters the appearance. But functionally it was the correct solution short of swapping to a 3090 cooler or watercooling, which wasn't possible on this card. Plus, personally, I think the aesthetic of these heatsinks in particular works well with the style of this card.

After this modification:

• Memory junction stabilized around 94C under heavy ray tracing load

• Tested using Control, which is particularly demanding on both RT cores and VRAM

• No more artifacting

• Stable under sustained heat soak

This confirms what I suspected. The cooling solution is fundamentally mismatched to this PCB layout. For mining workloads, it may have been acceptable. For gaming loads, it absolutely was not.


Results
With:

• Power limit removed

• Core thermals stabilized

• Memory thermals controlled

The card finally started behaving like the 3080 Ti class device it always should have been. Now with thermals in check and an increased limit so it's not throttling at 390W I no longer had these limits. It was all chip stability at this point.

Performance uplift was fair compared to the earlier configuration. It no longer feels artificially constrained.

Using the same overclocking settings as before:

Speedway: 5492 (from 5403)
Steel Nomad: 5338 (from 5155)
Port Royal: 14337 (from 13947)

Now my card performs better, is more thermally safe, and should last longer.

This engineering sample continues to be one of the most interesting pieces of hardware I have worked on. Not because it is rare, but because it reveals how close this SKU was to being something very different if it had made it to retail.

Edited February 24Feb 24 by ChintzyPC

pls make this gpu cold ^^ i would love to see more