[Success Report] RTX 3080 20GB (GA102) + Gigabyte Z890M + Windows 11 / WSL2 — FP16 103.27, BF16 100.62 TFLOPS

[xidian-zzx]

[Success Report] RTX 3080 20GB + Gigabyte Z890M + Windows 11 / WSL2

Hardware / Software

GPU: NVIDIA GeForce RTX 3080 20GB (modded GA102)
PCI ID: 10DE:2206
Subsystem ID: 146D10DE
VBIOS: 94.02.27.00.14
CPU: Intel Core Ultra 7 265K
Motherboard: Gigabyte Z890M AORUS ELITE WIFI7
Motherboard BIOS: F16
OS: Windows 11 + WSL2
NVIDIA Driver: 610.88 (WDDM)
Unlocker Version: nv-uefi-unlocker v2.1.10
Secure Boot: Disabled
VRAM: 20480 MiB

Unlock Procedure

1. Generated the GA102 firmware payload using make-usbroot.ps1 -Family 30.
2. Copied the generated files to a FAT32 USB drive.
3. Booted from the USB drive using the Gigabyte F12 boot menu.
4. First tested the 30-second timeout/abort mechanism. The loader returned to Windows normally without performing any GPU writes.
5. Rebooted from USB and pressed Enter to execute the unlock transaction.
6. The unlock transaction completed successfully and Windows 11 booted normally.
7. NVIDIA driver 610.88 initialized successfully.
8. WSL2 CUDA and PyTorch workloads ran normally.

No VBIOS flashing, NVIDIA driver replacement, motherboard BIOS update, or Windows reinstallation was required.

UEFI Unlock Results

The following results were recorded in loader.log:

TX: booter fired
TX: execution verified

SELECTOR RESULT: SS0=0x88888888 SS1=0x8

PCIe FLR triggered
FLR: pci state restored

POST-FLR selectors: ss0=0x88888888 ss1=0x8

POST-FLR: verified (selectors survive AON)

FINAL: transaction complete, all DMA released

Final UEFI handoff state:

SS0: 0x88888888
SS1: 0x8
FEAT_READOUT_1: 0x00000000
FEAT_PLM: 0xffffffff

ga/ada:ss0_unlock=true
ga/ada:ss1_unlock=true
ga/ada:plm_open=true

The SEC2 scrubber reported an ETimeout, but the subsequent booter execution, selector writes, PCIe FLR, and Windows handoff completed successfully.

WSL2 PyTorch Benchmark Results

GPU: NVIDIA GeForce RTX 3080 20GB
Matrix size: 4096 x 4096
Operation: torch.mm(a, b)
Warmup: 10 iterations
Benchmark: 30 iterations
CUDA synchronization after each operation

FP16:
Performance: 103.27 TFLOPS
Average latency: 1.33 ms

BF16:
Performance: 100.62 TFLOPS
Average latency: 1.37 ms
BF16 supported: True

Both benchmarks were performed inside WSL2 after the successful UEFI unlock procedure.

Verification Status

UEFI unlock: Confirmed
Unlock survives PCIe FLR: Confirmed
Windows 11 boot: Successful
NVIDIA driver 610.88: Working
WSL2 CUDA: Working
FP16/BF16 high throughput: Measured

Post-Windows-driver SS0/SS1 verification using ssdiag: Not yet performed
Locked vs. unlocked A/B benchmark: Not yet performed

The observed FP16/BF16 performance is consistent with the expected unlocked throughput. However, I have not yet directly verified the selector registers after Windows driver initialization or performed a controlled comparison against the locked state.

This report confirms successful UEFI unlocking and high WSL2 CUDA performance on an Intel Z890 platform with an RTX 3080 20GB.

Related issue: #109 (same GPU PCI ID, subsystem ID, and VBIOS, different motherboard).

I can provide the full loader.log and additional benchmark results if needed.

Thanks to the developer for this excellent project!

[physicsdolphin]

Thank you for the complete write-up — recorded (commit a1e0118) as the first ID-confirmed field card for the 2206 20GB-modded form, the first Z890/Arrow Lake platform datapoint, and our first WSL2 compute-stack sample in the community records and the GA102 matrix section.

Two details from your report worth calling out:

• Running the 30 s timeout abort before the real transaction is exactly the zero-write negative control we recommend — nice to see it done unprompted.
• Your card is the same ID/subsystem/VBIOS as #109 on a Z390P, where GSP devinit reset SS0/SS1 back to factory. No relock on your Z890M, which is consistent with the rule that surfaced there: on a 30-series card, the Turing GSP registry switch stays off.

The two items you flagged as unverified (no locked A/B, no post-driver ssdiag) are welcome as follow-ups but not required — the EFI read-backs plus FP16/BF16 absolutes stand on their own. If you ever capture the ssdiag verdict, a short paste here would complete the driver-side pair.