Skip to content
View in the app

A better way to browse. Learn more.

HWBOT Community Forums

A full-screen app on your home screen with push notifications, badges and more.

To install this app on iOS and iPadOS
  1. Tap the Share icon in Safari
  2. Scroll the menu and tap Add to Home Screen.
  3. Tap Add in the top-right corner.
To install this app on Android
  1. Tap the 3-dot menu (⋮) in the top-right corner of the browser.
  2. Tap Add to Home screen or Install app.
  3. Confirm by tapping Install.

RTX 3080 Ti 20GB Engineering Sample PCB analysis and bench results

Featured Replies

(TL;DR at bottom)
I recently got my hands on a 20GB variant of the RTX 3080 Ti that appears to be an unreleased engineering sample. This post is a mix of teardown notes, benching results, and general observations while figuring out what exactly this thing is and how it behaves.

STORY TIME
(skip to TEARDOWN if you just want the technical details)

Here's some unboxing photos: Link


A friend of mine came across a local listing for a “3080 Ti 20GB Engineering Sample” for $200. Obviously that set off alarms. Retail 3080 Ti is 12GB, so it sounded fake, a typo, or something weird. The listing repeated 20GB multiple times and explicitly said Engineering Sample, so he decided to take the gamble.

He bought two. The seller vanished immediately after the transaction which did not inspire confidence.

On first install the card only produced basic display output, almost like pre-driver behavior. GPU-Z showed a 3080 Ti with 20480MB VRAM. That was enough to justify opening it up. The memory layout matched what you would expect for 20GB, so at that point it was clearly not a simple mod or spoof.

It turns out the card requires a patched driver to actually function properly:
https://github.com/dartraiden/NVIDIA-patcher

Once patched, it behaved normally in games and basic benchmarks.

From what I can piece together, this looks like one of the scrapped higher-VRAM Ampere variants meant for purely mining that never made it to market. Likely explored late in the mining boom when VRAM capacity mattered more than bandwidth. Production seems to have been halted and these remained as engineering samples with partners. How they ended up in the wild is anyone’s guess.

The original owner opened it which means the paste was no longer good. Being several years old anyway, it needed fresh pads and paste. I offered to repaste/pad it in exchange for getting to spend time with it, but he ended up selling it to me for $700. For something this odd and fully functional, I was not going to pass on it.


TEARDOWN

The PCB is where things get interesting. Here’s some photos of the teardown: Link

And here's some up close photos of the chip and board: Link

It is not a standard 3080 Ti or 3090 board at all. The layout is basically a hybrid between a 3080 Ti and a 3090. Instead of twelve memory modules on the front like a normal 3080 Ti, this has ten on the front and ten on the back. There are four empty pads total. It looks very much like a reduced 3090 board that was adapted for 20GB instead of 24GB, and without NVLink. It’s as if they just chopped off the NVLink portion altogether seeing as this made it fit nicely in the 3080Ti chassis and mining likely wouldn’t benefit from this anyhow.

The memory bus is 320-bit instead of the 384-bit bus used on retail 3080 Ti. That lines up with ten active memory controllers.

The GPU die is GA102-250, which is typically associated with the 3090. Performance does not match a full 3090 though. The chip is clearly configured in a more limited way. My assumption is that Nvidia used harvested GA102 dies for compatibility with the memory layout rather than creating a completely new variant.

So what this ends up being is a strange mix of parts.
GA102-250 die
20GB GDDR6X
320-bit bus
3090-like memory layout
3080 Ti class configuration
No NVLink
Engineering firmware

It feels very much like a product that existed briefly for a specific purpose and then got cancelled.


BENCHING

Here's some photos of the repad process: Link

Since the card had already been opened and is several years old, I replaced the paste and pads before doing anything else. I used Kryonaut Extreme and Thermalright Odyssey 12W/mK pads. All pads ended up being 1.5mm. A 3090 repad guide got me most of the way there.

Test system:
Ryzen 9 5950X @ 4.675GHz all core
48GB DDR4 3266
Crosshair VIII Hero
EVGA P2 1000W
Slimmed Windows 11 install

Stock behavior:
Boost around 1980MHz
Memory 9500MHz
Max voltage 1.081V
Power 350W
Load temp 56C with fans at 100%

Bench Results
Speed Way: 5042
Steel Nomad: 4863
Port Royal: 13105

In gaming-oriented synthetics it consistently lands closer to a stock 3080 than a 3080 Ti. That makes sense given the 320-bit bus and the 350W limit. It just does not have the bandwidth of a normal 3080 Ti.

Where things got frustrating was compute. Any heavier compute workload or rendering test would start and then crash the system within seconds. It seems like a driver or firmware limitation tied to this being an engineering sample. So the one area where 20GB VRAM would actually matter is the one area I cannot properly test.


OVERCLOCKING

This is where the engineering sample nature really shows.

Flashing a different VBIOS is not something I am willing to try. The board is too unique and there is no dual BIOS safety net.

Voltage control is locked. The slider is disabled. Only power and frequency offsets work.

I did consider a shunt mod (not LM, actual shunt piggyback) since the card hits the power limit quickly and I am very familiar with the process, but modifying hardware on something this rare did not feel worth the risk.

I also wanted to try using Liquid Metal for expanded cooling but Founders Edition cards have very rigid mounting with little compliance, and Z-height contact is risky without significant modification.

Power limit: +14% for 396W max
Core offset: +150
Memory offset: +920

Speed Way: 5403
Steel Nomad: 5155
Port Royal: 13947

Scaling is pretty normal for Ampere. Memory overclocking headroom is decent which surprised me a bit, but the physical memory layout is more robust than a typical 3080 Ti even with the narrower bus.

I do have a friend who has an LN2 setup I've used in the past (Flagship Labs, the namesake of my OC team on here) so I am considering sending it to him. Doing so without a shunt mod makes me wonder if it would work though. Also would this site allow any points for recognition?


CONCLUSION

This has turned into one of the more interesting GPUs I have owned. It behaves like a slightly faster 3080 with a lot of VRAM and a lot of quirks. Gaming performance is below a real 3080 Ti, but the novelty and engineering aspect make up for it.

Cooling options are limited due to the unique PCB. There is no compatible full-cover waterblock since there are too many memory spots for a 3080Ti block, and a 3090 wouldn't fit with this NVLink-less PCB. For now it is running on air and working fine.

I am keeping it mainly because it is such an odd piece of hardware history. It feels like a snapshot of a product that almost existed but never did.

If anyone else has encountered one of these or has more information about similar late-Ampere engineering boards, I would be interested in comparing notes.

Here's a photo of it in my rig currently: Link

TL;DR:
Unreleased RTX 3080 Ti 20GB engineering sample with GA102-250 on a hybrid 3090-style PCB. 320-bit bus, requires patched driver. Gaming performance sits around a 3080. Compute workloads crash. Overclocks normally but firmware is heavily locked down. Very strange card and a fun one to dig into.

Edited by ChintzyPC

3 hours ago, ChintzyPC said:

I recently got my hands on a 20GB variant of the RTX 3080 Ti that appears to be an unreleased engineering sample. This post is a mix of teardown notes, benching results, and general observations while figuring out what exactly this thing is and how it behaves.

You will be surprised , if i tell you how many of these "engineering samples" exist in the local Chinese markets ?

In most cases , these are not real Nvidia engineering samples , but more like , hybrid local constructions from people or small workshops that have the means and skills , trying to make a buck.
They have access to pcb's , and stock of vga cores and ram chips ... they can produce almost anything.

At first it was the mining industry ... now it's the AI.

Anyway , you got a cheap , nice performing vga.

If you could implement a proper bios for it , that would be really great.

Enjoy it.

  • Author
Just now, TASOS said:

In most cases , these are not real Nvidia engineering samples , but more like , hybrid local constructions from people or small workshops that have the means and skills , trying to make a buck.

That’s interesting. I’ve heard about some of those hybrid builds coming out of the Chinese secondary market, especially from the mining/AI era.

I’d be curious how to distinguish one of those from a genuine NVIDIA engineering sample though.

On this card the serials match the box, labeling and PCB screen printing look consistent with NVIDIA factory style, chips are all consistent, and the assembly is extremely clean.

I also haven’t been able to find this exact FE version except for one, which wasn't very well documented publicly, except only a few 20 GB AIB variants floating around in RU/CN markets.

Are there specific tells you look for when identifying a workshop hybrid vs a true NVIDIA ES? Memory rework signs, BIOS structure, strap tables, etc?

4 hours ago, ChintzyPC said:

That’s interesting. I’ve heard about some of those hybrid builds coming out of the Chinese secondary market, especially from the mining/AI era.

I’d be curious how to distinguish one of those from a genuine NVIDIA engineering sample though.

On this card the serials match the box, labeling and PCB screen printing look consistent with NVIDIA factory style, chips are all consistent, and the assembly is extremely clean.

I also haven’t been able to find this exact FE version except for one, which wasn't very well documented publicly, except only a few 20 GB AIB variants floating around in RU/CN markets.

Are there specific tells you look for when identifying a workshop hybrid vs a true NVIDIA ES? Memory rework signs, BIOS structure, strap tables, etc?

Cant really help you on this one.

Engineering Samples don't usually come in full boxes with matching P/N's , S/N's , UPC's

They also dont usually need patched drivers to work ... as for example a workcenter RTX A4500 (that has a similar hardware configuration like yours) does.

They usually have somekind of indication or remark in the bios header

What does gpu=z , "reads" for your card ?

Can you post the graphic card tab and the advanced tabs ?

Here are a couple of pictures from a recent Chinese add

O1CN01x2d6QB23A4K4PA129_!!46116860184273

O1CN01WYwy4O23A4K3j8xMd_!!46116860184273

O1CN01kmuhjK23A4KdQM8QT_!!46116860184273

O1CN01HJHDof23A4KbYVMwx_!!46116860184273

GPU-Z from other adds

O1CN01UPzQUq1Cv5IrScry8_!!46116860184273

O1CN01ve3wDo1dWLYnd3gam_!!46116860184273

They seem to use driver version 576.40 , 576.52 , 581.08 (probably modded).

  • Author
3 hours ago, TASOS said:

They seem to use driver version 576.40 , 576.52 , 581.08 (probably modded).

Mine does require a patched driver, but it's only because it's a mining card which doesn't natively support normal vga output. Once it's patched it's fine. Uses 581.94 at the latest (anything later and the patch doesn't support it).

The vbios version also matches up with TechPowerUp's already uploaded and confirmed version for the ES too.

The reasons I’m leaning ES rather than a later workshop mod are mostly consistency details:

- FE-style PCB and construction all look factory, not reworked
- memory placement/routing looks native for a 20-chip config rather than added later
- serials and box all match and look NVIDIA-typical
- device IDs/BIOS strings line up more with pre-release GA102 configs than patched retail firmware
- no visible signs of reballing or memory swap work

Also, most of the modded cards I’ve seen are 3090 boards adapted into something else. This one appears to be a purpose-built 3080 Ti-class FE layout with 20 GB rather than a conversion, especially with that NVLink being completely missing.

Screenshot 2026-02-09 155123.png

Screenshot 2026-02-09 155242.png

Edited by ChintzyPC

  • Author

As an update, I did a super crazy thing and did a shunt mod and used LM. I know I said in my original post that I wasn't going to, but after seeing that baseline scores were already uploaded by others I figured what the hell. I've shunt modded and used LM on a wide variety of cards before and figured I knew how to do it without borking this card. Plus I was super bummed by the low power cap. I wanted to see what improvement I could get.

Now it boosts like crazy. Used to be limited to 1985Mhz even with the slider up due to eventual throttling. Now it's steady around 2085. Memory didn't change. Scores did improve quite a bit though. Now it seems to be pulling upwards of 500W without hitting power limits! That power cap was brutal.

Going to be going through the gauntlet and uploading scores soon. They're already significantly higher than before.

Now... LN2? hmmm...

Edited by ChintzyPC

3 hours ago, ChintzyPC said:

As an update, I did a super crazy thing and did a shunt mod and used LM. I know I said in my original post that I wasn't going to, but after seeing that baseline scores were already uploaded by others I figured what the hell. I've shunt modded and used LM on a wide variety of cards before and figured I knew how to do it without borking this card. Plus I was super bummed by the low power cap. I wanted to see what improvement I could get.

Now it boosts like crazy. Used to be limited to 1985Mhz even with the slider up due to eventual throttling. Now it's steady around 2085. Memory didn't change. Scores did improve quite a bit though. Now it seems to be pulling upwards of 500W without hitting power limits! That power cap was brutal.

Going to be going through the gauntlet and uploading scores soon. They're already significantly higher than before.

Now... LN2? hmmm...

Now you only have to worry about your vrm endurance.
Can it handle the extra load and temps ?

Is it designed for such kind of wattage ?

This sample is definitively an ES sample, it have a GA102-250 core (10DE 2205 device ID) which is a known ES core, it's not something local shop are making. If it was hand made it would have been using a GA102-225 core (10DE 2208 device ID)/regular 3080 Ti core with a bios that disable a channel (not hard to find) and no driver limitation (cause it's a retail core).

You're right when subbing the card in the 3080Ti 320bit category cause it correspond to this exact card.

Though be careful with the card/any serial number, I heard that Nvidia doesn't like you to own this ES card.

  • Author
On 2/16/2026 at 4:31 AM, TheQuentincc said:

Though be careful with the card/any serial number, I heard that Nvidia doesn't like you to own this ES card.

Thanks for the heads up. All photos should have the SN blurred and I know to not tell anyone about it.

On 2/13/2026 at 2:17 AM, TASOS said:

Now you only have to worry about your vrm endurance.
Can it handle the extra load and temps ?

Is it designed for such kind of wattage ?

I've replaced all the thermal pads with 20W/mK and am blowing an additional fan on the back so thermals should be fine. And after testing I haven't had any degradation in performance after heatsoak.
I've also made sure to keep it around 550W overall (after calculating based on what the actual wattage is with the shunts vs reported). So it's underclocked to meet that.

Edited by ChintzyPC

On 2/9/2026 at 4:39 PM, ChintzyPC said:

Mine does require a patched driver, but it's only because it's a mining card which doesn't natively support normal vga output. Once it's patched it's fine. Uses 581.94 at the latest (anything later and the patch doesn't support it).

The vbios version also matches up with TechPowerUp's already uploaded and confirmed version for the ES too.

The reasons I’m leaning ES rather than a later workshop mod are mostly consistency details:

- FE-style PCB and construction all look factory, not reworked
- memory placement/routing looks native for a 20-chip config rather than added later
- serials and box all match and look NVIDIA-typical
- device IDs/BIOS strings line up more with pre-release GA102 configs than patched retail firmware
- no visible signs of reballing or memory swap work

Also, most of the modded cards I’ve seen are 3090 boards adapted into something else. This one appears to be a purpose-built 3080 Ti-class FE layout with 20 GB rather than a conversion, especially with that NVLink being completely missing.

Screenshot 2026-02-09 155123.png

Screenshot 2026-02-09 155242.png

On 2/9/2026 at 4:39 PM, ChintzyPC said:

Mine does require a patched driver, but it's only because it's a mining card which doesn't natively support normal vga output. Once it's patched it's fine. Uses 581.94 at the latest (anything later and the patch doesn't support it).

The vbios version also matches up with TechPowerUp's already uploaded and confirmed version for the ES too.

The reasons I’m leaning ES rather than a later workshop mod are mostly consistency details:

- FE-style PCB and construction all look factory, not reworked
- memory placement/routing looks native for a 20-chip config rather than added later
- serials and box all match and look NVIDIA-typical
- device IDs/BIOS strings line up more with pre-release GA102 configs than patched retail firmware
- no visible signs of reballing or memory swap work

Also, most of the modded cards I’ve seen are 3090 boards adapted into something else. This one appears to be a purpose-built 3080 Ti-class FE layout with 20 GB rather than a conversion, especially with that NVLink being completely missing.

Screenshot 2026-02-09 155123.png

Screenshot 2026-02-09 155242.png

From what Ive seen in HPC servers.. i would guess these were used for Ai workflow in multiples, inside an HPC chassis 4u-6u [theres zero reason to mine anything with 20GB of gddr]. Bandwidth is the highest single GPU solution in the power range and core speeds are much easier to sustain vs 3090. The FW version likely disables certain instruction sets to make the cores 'purpose orientated' without adding code to define constraints at the hw level.

On 2/16/2026 at 12:31 PM, TheQuentincc said:

This sample is definitively an ES sample, it have a GA102-250 core (10DE 2205 device ID) which is a known ES core, it's not something local shop are making. If it was hand made it would have been using a GA102-225 core (10DE 2208 device ID)/regular 3080 Ti core with a bios that disable a channel (not hard to find) and no driver limitation (cause it's a retail core).

You're right when subbing the card in the 3080Ti 320bit category cause it correspond to this exact card.

Though be careful with the card/any serial number, I heard that Nvidia doesn't like you to own this ES card.

Don't take for granted that core GA102-250 is ES only.

There are many cards like these in local Chinese market , with different bios and subventors.

Perhaps it was a limited Asia only version (common practice).

O1CN01EuJuS91RVBBnK7zWj_!!46116860184273

O1CN01Nxtr1m1RVBBnK6mYl_!!46116860184273

O1CN01tTX74N27SANtGrjcG_!!46116860184273

*Whoever wants the pics i am posting , save them local , cause these are from current sale adds , and they will be deleted short after each card sales.

5 hours ago, ChintzyPC said:

Thanks for the heads up. All photos should have the SN blurred and I know to not tell anyone about it.

I've replaced all the thermal pads with 20W/mK and am blowing an additional fan on the back so thermals should be fine. And after testing I haven't had any degradation in performance after heatsoak.
I've also made sure to keep it around 550W overall (after calculating based on what the actual wattage is with the shunts vs reported). So it's underclocked to meet that.

It's not only about cooling properly the vrm area.
You (or an electronic expert) should calculate according to the hardware parts used and number of phases , how much wattage they can support.

Your card's pcb is the "short" one.

1 hour ago, TASOS said:

Don't take for granted that core GA102-250 is ES only.

There are many cards like these in local Chinese market , with different bios and subventors.

Perhaps it was a limited Asia only version (common practice).

O1CN01EuJuS91RVBBnK7zWj_!!46116860184273

O1CN01Nxtr1m1RVBBnK6mYl_!!46116860184273

O1CN01tTX74N27SANtGrjcG_!!46116860184273

*Whoever wants the pics i am posting , save them local , cause these are from current sale adds , and they will be deleted short after each card sales.

GA102-250 is ES only.

It's not because there is more than one core ever that's it's retail core. The need for patched driver is the biggest clue here. Just have a look at all ES Intel CPU sold the past few years in Asia, they all are still ES despite hundreds/thousands of SKU sold.

2 hours ago, TheQuentincc said:

GA102-250 is ES only.

It's not because there is more than one core ever that's it's retail core. The need for patched driver is the biggest clue here. Just have a look at all ES Intel CPU sold the past few years in Asia, they all are still ES despite hundreds/thousands of SKU sold.

Maybe actual facts show something different ?
Branded companies (like Colorful , Palit , Gainward) would circulate a product with S/N and P/N using ES vga core ?

*edit

My personal point of view is , that cause of the "last minute" SKU cancel of this specific model ... companies that already had a limited stock of these "products" , just fabricated them and "throw them" to the market.

Edited by TASOS

  • Author
18 hours ago, TASOS said:

It's not only about cooling properly the vrm area.
You (or an electronic expert) should calculate according to the hardware parts used and number of phases , how much wattage they can support.

Your card's pcb is the "short" one.

Which is why I included the detail about 550W. It uses a 18 phase core and 2 phase memory layout which is a huge number of phases for a 3080Ti. This means the voltage system is capable of at least 1100W (conservative) of power. (18 x 70A x 1.081v = 1362.06W)

So if thermals are under control for the VRM then it will be able to handle the wattage easily. But VRM isn't the concern. It's the connector. It uses a 2-plug pigtail 12-pin connector. Theoretical max of 600-650W before the plug starts to overheat. I've tested the plug temp with my thermocouple and it's fine at 550W, which is where I set the limit with my overclock.

Edited by ChintzyPC

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
Reply to this topic...

Account

Navigation

Search

Search

Configure browser push notifications

Chrome (Android)
  1. Tap the lock icon next to the address bar.
  2. Tap Permissions → Notifications.
  3. Adjust your preference.
Chrome (Desktop)
  1. Click the padlock icon in the address bar.
  2. Select Site settings.
  3. Find Notifications and adjust your preference.