feat: support CoreSpan 5090 Inference System (Blackwell / RTX 5090)

README.md

@@ -2,6 +2,8 @@
 
 A command-line tool for node provisioning (CUDA toolkit, Nvidia recommended drivers, libraries) and GPU bandwidth benchmarking via `nvbandwidth`. See https://github.com/NVIDIA/nvbandwidth.
 
+Supports all major NVIDIA GPU architectures including the **CoreSpan 5090 Inference System** (Blackwell / RTX 5090, GB202). The provisioning script automatically detects the RTX 5090, selects CUDA 12.8, and installs driver 570+.
+
 ## Building from Source
 
 ```bash
@@ -72,6 +74,28 @@ ai-studio-cli nvbandwidth test host_to_device_memcpy_sm
 
 ---
 
+## NVLink Benchmarking (CoreSpan 5090 Inference System)
+
+The CoreSpan 5090 Inference System connects multiple RTX 5090 GPUs via NVLink. Use the dedicated `nvlink` subcommand to validate peer-to-peer interconnect health and bandwidth:
+
+```bash
+ai-studio-cli nvbandwidth nvlink
+```
+
+This runs six peer-to-peer testcases covering unidirectional and bidirectional NVLink transfers as well as all-to-host / host-to-all patterns.
+
+```bash
+# JSON output (for monitoring / alerting pipelines)
+ai-studio-cli nvbandwidth nvlink --json
+
+# Larger buffer and more samples for sustained-bandwidth measurement
+ai-studio-cli nvbandwidth nvlink --buffer-size 1024 --samples 10
+```
+
+The same flags as `run` and `test` apply.
+
+---
+
 ## Usage Notes
 
 - **Sudo Password**: The tool will securely prompt for your sudo password interactively during setup.

ai-studio-cli/assets/driversInstallation.sh

@@ -70,7 +70,7 @@ detect_gpu() {
         elif echo "$GPU_UPPER" | grep -qE "A100|A6000|A30 |A10 |RTX 30|AMPERE|RTX A[0-9]"; then ARCH="ampere"
         elif echo "$GPU_UPPER" | grep -qE "RTX 40|L40|ADA|RTX 6000 ADA";                   then ARCH="ada"
         elif echo "$GPU_UPPER" | grep -qE "H100|H200|HOPPER";                               then ARCH="hopper"
-        elif echo "$GPU_UPPER" | grep -qE "RTX 50|B100|B200|BLACKWELL";                    then ARCH="blackwell"
+        elif echo "$GPU_UPPER" | grep -qE "RTX 5090|RTX 5080|RTX 5070|RTX 50|B100|B200|B300|GB202|BLACKWELL"; then ARCH="blackwell"
         fi
     fi
 

ai-studio-cli/cmd/nvlink.go

@@ -0,0 +1,57 @@
+package cmd
+
+import (
+	"github.com/corespan/ai-studio-cli/internal/nvbandwidth"
+	"github.com/spf13/cobra"
+)
+
+// nvlinkRunner shares flags with the top-level nvbwRunner but is a distinct
+// instance so the NVLink subcommand can override defaults independently.
+var nvlinkRunner = nvbandwidth.NewRunner()
+
+// nvlinkTestcases are the nvbandwidth testcases that exercise NVLink
+// peer-to-peer transfers.  These are the most relevant benchmarks for
+// multi-GPU inference systems such as the CoreSpan 5090.
+var nvlinkTestcases = []string{
+	"device_to_device_memcpy_read_sm",
+	"device_to_device_memcpy_write_sm",
+	"device_to_device_bidirectional_memcpy_read_sm",
+	"device_to_device_bidirectional_memcpy_write_sm",
+	"all_to_host_memcpy_sm",
+	"host_to_all_memcpy_sm",
+}
+
+var nvlinkCmd = &cobra.Command{
+	Use:   "nvlink",
+	Short: "Run NVLink peer-to-peer bandwidth benchmarks",
+	Long: `Runs the nvbandwidth testcases that exercise GPU-to-GPU (NVLink)
+bandwidth.  Useful for validating multi-GPU interconnect health on systems
+such as the CoreSpan 5090 Inference System.
+
+Testcases run:
+  - device_to_device_memcpy_read_sm
+  - device_to_device_memcpy_write_sm
+  - device_to_device_bidirectional_memcpy_read_sm
+  - device_to_device_bidirectional_memcpy_write_sm
+  - all_to_host_memcpy_sm
+  - host_to_all_memcpy_sm`,
+	RunE: func(cmd *cobra.Command, args []string) error {
+		for _, tc := range nvlinkTestcases {
+			if err := nvlinkRunner.Run([]string{"--testcase", tc}); err != nil {
+				return err
+			}
+		}
+		return nil
+	},
+}
+
+func init() {
+	nvlinkCmd.Flags().IntVarP(&nvlinkRunner.BufferSize, "buffer-size", "b", 512, "Memcpy buffer size in MiB")
+	nvlinkCmd.Flags().IntVarP(&nvlinkRunner.TestSamples, "samples", "i", 3, "Number of benchmark iterations")
+	nvlinkCmd.Flags().BoolVarP(&nvlinkRunner.Verbose, "verbose", "v", false, "Enable verbose output")
+	nvlinkCmd.Flags().BoolVarP(&nvlinkRunner.JSON, "json", "j", false, "Output results as JSON")
+	nvlinkCmd.Flags().BoolVarP(&nvlinkRunner.SkipVerify, "skip-verify", "s", false, "Skip data verification after copy")
+	nvlinkCmd.Flags().BoolVarP(&nvlinkRunner.UseMean, "use-mean", "m", false, "Use arithmetic mean instead of median")
+
+	nvbandwidthCmd.AddCommand(nvlinkCmd)
+}