For this release, the software versions are as follows:

• CUDA Toolkit 12: 12.4

  Note that starting with CUDA 11, individual components of the toolkit are versioned independently. For a full list of the individual versioned components (for example, nvcc, CUDA libraries, and so on), see the CUDA Toolkit Release Notes.

• NVIDIA Data Center GPU Driver: 550.90.07 (Linux) / 552.55 (Windows)

• Fabric Manager: 550.90.07 (Use nv-fabricmanager -v)

• NVFlash: 5.791

For more information on getting started with the NVIDIA Fabric Manager on NVSwitch-based systems (for example, NVIDIA HGX A100), refer to the Fabric Manager User Guide.

• On Intel Emerald Rapids chipset ID 0x1b81, PCIe P2P sometimes did not work even though the GPU driver reports P2P as supported in nvidia-smi topo -p2p rw command. This issue is resolved. 4448950
• Running nvidia-smi caused a kernel panic, which was due to Null pointer dereferencing in the code path. This was fixed by making it a valid pointer. 4644086
• The crashcat buffer was not large enough to contain all crash information, resulting in truncated reports. The buffer size has been increased. 4607635
• The gRPC event poller is now proactively triggered by invoking waitForConnected() on the channel with an empirically shorter deadline, which is sufficient to get the channel to READY state. 4605545
• Fixed a null pointer dereference caused by an unchecked return status from the API that allocates the channel notifier action table. 4603505
• RC errors were observed during the channel allocation path, leading to accessing the notifier action table which has not been allocated yet. This issue is resolved. 4589964
• Invalid objects are no longer accessed when a process is killed while profiling with MIG enabled. 4584716
• On some versions of the GPU driver, the nvidia-smi topo -m command showed GPU connections as SYS, implying that they are connected across the SMP interconnect between NUMA nodes when they are actually part of the same NUMA node. This issue has been resolved and these connections now correctly show NODE instead of SYS. 4576789
• Resolved an issue that caused driver compilation warnings whenever a .run package was installed on systems with kernels built with Clang or newer GCC. 4569283
• Fixed a bug in the compiler's register allocator, where it was previously unable to correctly spill/refill and thereby allocate register for the barrier register class for test cases involving loops with breaks. 4546457
• XID error messages for DRAM and LTCs are no longer limited to 1 per boot. XID error messages for DBEs are now printed when error containment is enabled. 4478617
• The GPU Max Operating Temp now correctly displays the Slowdown temperature. 4507899
• Changed row remapping histogram to only account for uncorrectable row remappings. 4451237
• NVENC session data processing tasks are now scheduled into work-items so processing time won't be propagated up. NVENC NVENC session data are processed as CPU's work-items when no high priority tasks occupy the CPU. 4437306
• Improved the ability of nvidia-modprobe to detect whether kernel modules are already loaded. This corrects an issue that prevented nvidia-persistenced from setting persistence mode on some systems. 4435513
• The drain and reset flag is now set when uncorrectable row remapping occurs. 4432594
• Fixed an issue with particular types of host memory allocations, for example CMA, where registering that memory buffer on the GPU (via cuMemHostRegister) first would later prevent to register the same memory buffer on the RDMA NIC (via ibv_reg_mr). 4429264
• Fixed the regression issue where the new hwloc version 2.9.2 (integrated by NVML library) failed to detect two GPUs under a common NUMA node, resulting in the nvidia-smi topo -m command showing SYS instead of NODE. 4412560
• When running MPS where the client and server are in different containers, it is possible that they may have access to different versions of the glibc library. There were observed hangs in some combinations of this mismatch due to differing ABI for the glibc semaphore implementation. This version includes a fix for this issue by making the synchronization used by MPS agnostic to the GLIBC implementation. 4382578
• A clear signal is needed on OOB when the GPU exceeds the SRAM UCE > 4 criteria for RMA. The SRAM Threshold Exceeded flag from nvidia-smi is now reported through OOB. 4315988
• Improved error containment by forwarding poisoned data to target engine instead of handling at GPU granularity. 4060421

• 4755732 - Fabric Manager won’t start on Version 550.90.07.
  • Newer 550 branch will fix this issue (target 1.6.0 firmware time frame)
  • Workaround: Remove After=multi-user.target from the service unit file (for example, /lib/systemd/system/nvidia-fabricmanager.service)

• 4687401 - Some packages may install a different Linux kernel flavor when upgrading the Open-source driver from older branches to 550 and newer.

  To work around this, end-users on openSUSE and SLES should upgrade their NVIDIA Open GPU kernel modules using the following:

  sudo zypper -v install $(zypper search -s nvidia-open-driver-G06-kmp-<flavor> | sed 's| ||g' | awk -F '|' '/<branch>/ {print $2"="$4}')

  sudo zypper -v install cuda-drivers-<branch>

  For example:

  sudo zypper -v install $(zypper search -s nvidia-open-driver-G06-kmp-default | sed 's| ||g' | awk -F '|' '/550/ {print $2"="$4}')

  sudo zypper -v install cuda-drivers-550