Rig Architecture

A GeForce NOW "rig" is a Windows VM running on bare-metal NVIDIA GPU servers. Each rig hosts a single game session at a time.

Hardware Layer

Rigs run on NVIDIA's nvmetal bare-metal infrastructure. The example rig from the backup:

Property	Value
Hostname	[REDACTED_HOSTNAME]
Zone	[REDACTED_ZONE] (Amsterdam)
GPU	NVIDIA L40 (full)
Instance Type	[REDACTED_INSTANCE_TYPE]
Platform	NGN Platform v2.1
Hypervisor	Xen
OS	Windows 11 (QCOW2 image)

Virtualization

Rigs run as Xen VMs. The framework includes:

• LocalXen / RemoteXen — VM lifecycle management (start, stop, restart)
• PCI passthrough — GPU pinned directly to VM
• Xen tools — xenstore_client.exe for reading VM metadata
• Hostname sync from Xen VM name via wmic computersystem

Zone Architecture

Zones are geographic clusters of rigs:

Zone: [REDACTED_ZONE] (Amsterdam)
├── Provision Managers (PM): [REDACTED_IP], [REDACTED_IP], ...
├── Game Seat Gateway (GSG): gsg.[REDACTED_ZONE].svc.cluster.local:443
├── DNS Cache: [REDACTED_IP]
├── KMS Host: consumerkms.nvidiangn.net:1688
├── Seat Pool: [REDACTED_POOL]
└── Logging Server: [REDACTED_IP]

Zone Properties

Stored in Redis and accessible via asgard_util_zone_properties:

Section	Contents
ZoneProperties	Zone name, mode (gaming/pro), region
NetworkTopology	IP ranges, subnets, routing
GameMachine	GPU type, instance config
NATVM	NAT virtual machine settings

Machine Roles

Role	Description
awsseats	Game seat VMs
natvm	NAT/gateway VMs
pm	Provision Manager nodes
redis	State database nodes
storage	Storage servers

State Database

Every rig connects to a Redis instance (port 6399) for:

• Configuration storage
• Session state tracking
• Zone property caching
• Service registration

Seat Pool & Instance Types

The seat pool [REDACTED_POOL] determines rig capabilities:

Instance Type	GPU Fraction	Use Case
[REDACTED_INSTANCE_TYPE_PATTERN]	Full L40	Premium tier (4K, HDR, 120fps)
[REDACTED_INSTANCE_TYPE_PATTERN]	Half L40	Standard tier
[REDACTED_INSTANCE_TYPE_PATTERN]	Quarter L40	Free tier
ga10g_2.*_large	Half A10G	Standard tier
gt10_2.*_medium	Half T10	Legacy tier

Half/quarter GPUs have restrictions like:

• AV1 encoding disabled on half GPUs
• Video encoder perf check skipped
• Lower resolution caps

L1 Validation Tests

Before a rig accepts sessions, l1test.py validates:

1. Environment variables (AG_LOGS, AG_HOME)
2. Required services installed:
   • nvcloudinit
   • KioskPwdChanger
   • seatinitservice
   • NvContainerLocalSystem
3. User accounts (kiosk, xen)
4. GPU present and drivers installed
5. Network connectivity (DHCP/ping)
6. Rollback state clean

Crash Detection & Recovery

A Windows Scheduled Task (CrashDetector.xml) runs detector.exe at boot:

• Monitors for crash dumps: nvstreamer.exe, nvcontainer.exe, NvRtcStreamer
• Collects .dmp files from:
  • C:\ProgramData\NVIDIA Corporation\Crashdumps\
  • C:\Windows\MEMORY.DMP
  • C:\Windows\minidump\
  • C:\asgard\logs\AutoOnboarder\
• Reports crashes via telemetry events:
  • NGS_NvStreamerCrash
  • NGS_NvContainerCrash
  • NGS_KernelCrash
  • NGS_CTMTCrash
  • NGS_TasCrash

Windows Licensing

Rigs use KMS activation:

KMS Host: consumerkms.nvidiangn.net:1688

The rearmWindows.py script:

• Checks license grace period via slmgr.vbs
• Rearms when remaining time < 15 days
• Disables Windows Activation Technologies (WatAdminSvc)
• Triggers reboot after rearm

Performance Monitoring

GridPerf Templates

Windows Performance Monitor templates (GridPerfMonTemplate.xml) track:

• Process metrics: nvstreamer, NvRtcStreamer, NvGridSvc, apptracer
• GPU counters: Temperature, clock speeds, utilization, frame buffer usage
• System: Memory, disk, CPU, network

GPU Performance Counters

Custom NVIDIA counter manifest (nvPerfProvider.man):

Counter	Unit
Temperature	Degrees Celsius
Graphics Clock	MHz
Memory Clock	MHz
Graphics Utilization	%
Frame Buffer Utilization	%
Video Utilization	%
Fan/Cooler Level	%