⚠️ NOT the power-capped units flooding right now.
Multiple vendors are selling these cards locked at 110W and 1000MHz without disclosing it. These are not those cards.
What are these cards?
These ARE Instinct MI25s — full stop. AMD built these as MI25s, shipped them in MI25 cases with MI25 coolers on MI25 PCBs. Part number 102D0513300 is an OEM MI25 variant sold to enterprise clients with vendor-specific firmware that limited power and clocks for their use case. Same chip. Same PCB. Same HBM2. Just a software cap.
I've flashed every card back to standard MI25 VBIOS — restoring the clocks and power limits the hardware was physically built to run. The card already knew it was an MI25. The VBIOS just needed to agree.
Capped vs Unlocked:
What others sell (capped)This listing (unlocked)Power- 110W 220WClock- 852MHz base / 1000MHz boost 852MHz base / 1500MHz boostHBM2 Bandwidth- ~240 GB/s (power limited) 359 GB/s tested (484 GB/s spec in x16 slot)
Note on bandwidth: Benchmark shows 359 GB/s tested in PCIe x1 slot. AMD spec is 484 GB/s peak in a full x16 slot. HBM2 bandwidth scales with PCIe bandwidth — in a proper x16 slot you'll see numbers closer to spec.
Bonus: These cards support runtime overclocking via ROCm tools. Experienced users can push clocks beyond 1500MHz through software — results vary by card.
Is this safe?
Yes. This is not a cross-flash from a different GPU. The Vega10 MI25 and the OEM variant are identical silicon. Flashing to MI25 VBIOS is putting the correct firmware back on hardware designed for it from the factory. Think of it like removing a speed limiter from a truck — same engine, same drivetrain, just running like it was built to.
Card identification:
- OEM part number: 102D0513300
- Flashed to: MI25 VBIOS 016.001.001.000.009103
- Die: Vega10 (gfx900)
- Post-flash identifies as: Radeon Instinct MI25
Verified on every card before shipping:
- Boost clock: 1500MHz ✓
- HBM2 bandwidth: 359 GB/s ✓
- ROCm gfx900 compatible ✓
- Tested running live AI inference workload ✓
Important — this is a compute card only:
- No DirectX, no Vulkan gaming support
- No display output on compute VBIOS version
- Linux ROCm recommended
- Not a gaming GPU
⚠️ Passive cooling — no fans included.
These are passively cooled datacenter cards. They require adequate airflow in your system — a server chassis, open air rack with fans, or active cooling solution. Do not install in a standard desktop case without supplemental airflow directed through at the card.
Ideal for:
Home AI inference clusters, ROCm builds, llama.cpp / ik_llama.cpp RPC nodes, mixed vendor GPU stacks
Condition: Used, tested, working
Includes: Card only
