RTX PRO 4500 Blackwell fails in TDX CC-On (0x62:0x65:2119) — which TRD release will add support?

Accelerated Computing Confidential Computing
,
1

We are testing single-GPU passthrough Confidential Computing with an Intel TDX confidential VM and two NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs.

Both physical cards reproduce the same behavior:

  • PCI ID: 10de:2c3a
  • Subsystem ID: 10de:21f4
  • VBIOS: 98.03.B1.00.01
  • Driver tested: NVIDIA Open GPU Kernel Modules 610.43.02
  • Standard KVM / CC-Off: GPU initializes successfully
  • Intel TDX / CC-On: initialization fails
  • Intel TDX / CC-DevTools: initialization also fails

Relevant log:

FSP boot command completed; ccEnabled=1
target-mask wait failed; status=0x65
FSP microcode v4.76
FSP scratch0=0x9f
FSP scratch1=0x54500003
GSP HWCFG2 / mailbox=0xbadf4100
RmInitAdapter failed! (0x62:0x65:2119)

As a control test, an RTX PRO 6000 Blackwell Server Edition initializes successfully in CC-On mode using the same host, TDX guest, and driver. Its GSP target mask is released normally.

The RTX PRO 4500 product page describes the GPU as Confidential Computing capable. However, it is not listed in the SPT CC hardware SKU list for the currently published R595 Trusted Computing Driver, so we suspect that official driver support has not yet been released.

We also tested the following on the RTX PRO 4500:

  • A second physical card with the same subsystem ID and VBIOS
  • CC-DevTools mode
  • The regular signed production GSP-FMC image instead of the CC production image

All RTX PRO 4500 CC attempts fail at the same GSP target-mask release stage. No initialization or security checks were bypassed; the open kernel module was modified only to add diagnostic logging.

Could NVIDIA please clarify:

  1. Which NVIDIA Trusted Computing Driver (TRD) release is expected to add SPT CC support for the RTX PRO 4500 Blackwell Server Edition, and when is that release expected to become publicly available?
  2. Is VBIOS 98.03.B1.00.01 already CC-capable, or will a newer VBIOS also be required?
  3. If a newer VBIOS is required, how can it be obtained for an NVIDIA subsystem board (10de:21f4) installed in a GIGABYTE/Giga Computing server?
  4. Does 0xbadf4100 indicate that the GSP remains locked because the current TRD does not support this SKU, or because the VBIOS is not provisioned for CC?

We can provide the complete instrumented logs and failed/successful FSP DMEM captures privately if needed.

2

We do not currently support that GPU for confidential compute, but keep an eye on the Secure AI Compatibility Matrix which will list required driver and VBIOS versions.