Controlla NVLink in Windows

Si prega di notare che è necessario Installare i driver Nvidia in Windows o Installare il toolkit CUDA in Windows prima di controllare le connessioni NVLink. Inoltre, cambia la modalità GPU in TCC.

Installa Visual Studio

Assicuriamoci che tutto funzioni correttamente eseguendo cuda-samples dal repository ufficiale. Per fare ciò, dobbiamo installare Visual Studio 2022 CE (Community Edition) in sequenza e reinstallare il toolkit CUDA per attivare i plugin di VS. Visita https://visualstudio.microsoft.com/downloads/ per scaricare Visual Studio 2022:

Esegui l'installer scaricato, seleziona Sviluppo desktop con C++, e clicca sul pulsante Installa:

Esegui test

Reinstalla il toolkit CUDA utilizzando la nostra guida passo-passo Installa il toolkit CUDA in Windows. Riavvia il server e scarica l'archivio ZIP con cuda-samples. Estrai il contenuto ed apri la sottocartella Samples\1_Utilities\bandwidthTest. Fai doppio clic su bandwidthTest_vs2022 e fai girare usando la scorciatoia da tastiera Ctrl + F5:

[CUDA Bandwidth Test] - Starting...
  Running on...
  
   Device 0: NVIDIA RTX A6000
   Quick Mode
  
   Host to Device Bandwidth, 1 Device(s)
   PINNED Memory Transfers
     Transfer Size (Bytes)        Bandwidth(GB/s)
     32000000                     6.0
  
   Device to Host Bandwidth, 1 Device(s)
   PINNED Memory Transfers
     Transfer Size (Bytes)        Bandwidth(GB/s)
     32000000                     6.6
  
   Device to Device Bandwidth, 1 Device(s)
   PINNED Memory Transfers
     Transfer Size (Bytes)        Bandwidth(GB/s)
     32000000                     637.2
  
  Result = PASS
  
  NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.

Puoi eseguire qualsiasi esempio. Prova Samples\5_Domain_Specific\p2pBandwidthLatencyTest per vedere la tua topologia e matrice di connettività:

[P2P (Peer-to-Peer) GPU Bandwidth Latency Test]
  Device: 0, NVIDIA RTX A6000, pciBusID: 3, pciDeviceID: 0, pciDomainID:0
  Device: 1, NVIDIA RTX A6000, pciBusID: 4, pciDeviceID: 0, pciDomainID:0
  Device=0 CAN Access Peer Device=1
  Device=1 CAN Access Peer Device=0
  
  ***NOTE: In case a device doesn't have P2P access to other one, it falls back to normal memcopy procedure.
  So you can see lesser Bandwidth (GB/s) and unstable Latency (us) in those cases.
  
  P2P Connectivity Matrix
       D\D     0     1
       0       1     1
       1       1     1
  Unidirectional P2P=Disabled Bandwidth Matrix (GB/s)
     D\D     0      1
       0 671.38   6.06
       1   6.06 671.47
  Unidirectional P2P=Enabled Bandwidth (P2P Writes) Matrix (GB/s)
     D\D     0      1
       0 631.31  52.73
       1  52.83 673.00
  Bidirectional P2P=Disabled Bandwidth Matrix (GB/s)
     D\D     0      1
       0 645.00   8.19
       1   8.11 677.87
  Bidirectional P2P=Enabled Bandwidth Matrix (GB/s)
     D\D     0      1
       0 655.96 101.78
       1 101.70 677.92
  P2P=Disabled Latency Matrix (us)
     GPU     0      1
       0   2.20  49.07
       1  10.33   2.20
  
     CPU     0      1
       0   3.55   7.01
       1   6.79   3.39
  P2P=Enabled Latency (P2P Writes) Matrix (us)
     GPU     0      1
       0   2.19   1.33
       1   1.26   2.22
  
     CPU     0      1
       0   6.80   4.86
       1   2.09   3.02
  
  NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.