Hardware : Différence entre versions

De ClustersSophia
Aller à : navigation, rechercher
(→‎SuperMicro 4029 node)
 
(15 révisions intermédiaires par le même utilisateur non affichées)
Ligne 3 : Ligne 3 :
  
  
−
= CPU & GPU summary =
 
−
 
−
The current platform includes:
 
−
 
−
*    16 dual-Xeon SP 10 cores, 9.6GB/core RAM
 
−
*    1 quad-Xeon SP 20 cores, 12.8GB/core RAM
 
−
*    8 dual-Xeon 8 cores, 16GB/core RAM
 
−
*    16 dual-Xeon 10 cores, 9.6GB/core RAM
 
−
*    6 quad-Opteron 16 cores, 4GB/core RAM
 
−
*    6 quad-Opteron 12 cores, ~5.3GB/core RAM
 
−
*    1 dual-EPYC 32 cores, 16GB/core RAM
 
−
*    1 dual-EPYC 32 cores, 6GB/core RAM
 
−
*    2 dual-EPYC 16 cores, 8GB/core RAM
 
−
*    2 dual-EPYC 24 cores, 10GB/core RAM
 
−
*    164 Nvidia GPU
 
−
 
−
The number of CPU cores available is :
 
−
*    '''144''' Xeon SP Gold 2.6GHz cores : ~6.7 TFlops AVX-512  (9.2 TFlops peak)
 
−
*    '''320''' Xeon SP Silver 2.2GHz cores : ~8.8 TFlops AVX2 (11.2 TFlops peak) / ~6.4 TFlops AVX-512  (9.7 TFlops peak)
 
−
*    '''80''' Xeon SP Gold 2.4GHz cores : ~2.4 TFlops AVX2 (3.1 Tflops peak) / ~3.8 Tflops AVX-512 (4.9 Tflops peak)
 
−
*    '''128''' Xeon E5 v2 2.6GHz cores : ~2.4 Tflops (2.7 Tflops peak)
 
−
*    '''320''' Xeon E5 v2 2.8GHz cores : ~5.9 Tflops (7.2 Tflops peak)
 
−
*    '''384''' Opteron 2.3GHz cores : ~2.7 Tflops (3.5 Tflops peak)
 
−
*    '''288''' Opteron 2.2GHz cores : ~2 Tflops (2.5 Tflops peak)
 
−
*    '''64''' EPYC 2.9GHz cores : ~1.5 Tflops AVX2
 
−
*    '''64''' EPYC 2.5GHz cores : ~1.25 Tflops AVX2
 
−
*    '''32''' EPYC 2.8GHz cores : ~1.25 Tflops AVX2
 
−
*    '''48''' EPYC 2.65GHz cores : ~1.25 Tflops AVX2
 
−
'''The total cumulated CPU computing power is close to ~37.9 TFlops in double precision (44,5 Tflops peak)'''
 
−
 
−
 
−
The number of GPU cores available is :
 
−
*    '''625088''' Streaming Processor Cores
 
  
 
= CPU clusters =
 
= CPU clusters =
Ligne 164 : Ligne 131 :
 
= GPU nodes =
 
= GPU nodes =
  
−
== SuperMicro 4029 node ==
+
== HPE DL385 GPU nodes (2023) ==
  
−
'''SuperMicro 4029GP-TVRT''' : 1 node ('''nefgpu41''' )
+
'''HPE DL385''' dual-AMD EPYC 7313 @ 3.0GHz (16 cores) : 3 nodes ( '''nefgpu59 to nefgpu61''' )
 +
* Nvidia A40 PCIe GPUs cards
 +
**        10752 CUDA cores per card
 +
**        336 tensor cores per card
 +
**        48GB of RAM capacity per card
 +
**        Tensor performance peak : 149.6 TFlops per card
 +
**        Simple precision performance peak: 74.8 Tflops per card
 +
**        Double precision performance peak: 37.4 Tflops per card
 +
**        696 GB/s GPU memory bandwidth with error correction (ECC)
 +
*    2x gigabit network ports (one connected)
 +
*    infiniband EDR card (connected to FDR switch)
 +
*    hyperthreading active
  
−
[[Image:SYS-4028GR-TVRT.jpg|200px|right]]
+
'''HPE DL385''' dual-AMD EPYC 7513 @ 2.6GHz (32 cores) : 2 nodes ( '''nefgpu57 to nefgpu58''' )
−
 
+
* Nvidia A100 - 80GB PCIe GPUs cards
−
* 4x Nvidia Tesla V100 SXM2 GPUs cards
+
**        6912 CUDA cores per card
−
**        5120 CUDA cores per card
+
**        432 tensor cores per card
−
**        640 tensor cores per card
+
**        80GB of RAM capacity per card  
−
**        32GB of RAM capacity per card  
+
**        Tensor performance peak : 312 TFlops per card
−
**        Simple precision performance peak: 15.7 Tflops per card
+
**        FP16 Compute: 78 Tflops per card
−
**        Double precision performance peak: 7.8 Tflops per card
+
**        FP32 Compute: 19.5 Tflops per card
−
**        Tensor performance peak : 125 TFlops per card
+
**        FP64 Compute: 9.7 Tflops per card
−
**        900 GB/s GPU memory bandwidth with error correction (ECC)
+
**        1935GB/s GPU memory bandwidth with error correction (ECC)
−
*    CPU-GPU connection with a PCIe gen3 16x interface
+
*    2x gigabit network ports (one connected)
−
*    GPU-GPU connection with NVLink 2.0 (25GB/s per direction per link, 1 or 2 links per card pair)
+
*    infiniband EDR card (connected to FDR switch)
−
*    CPU : 2x Xeon SP Gold 6126 @ 2.6 GHz (24 cores)
 
−
*    RAM capacity : 384 GB
 
−
*    storage :
 
−
**        system & /tmp : 240GB SSD SATA RAID-1
 
−
**        /local/tmp scratch data : 4.8 TB SSD SATA RAID-5
 
−
**        RAID controller SAS 12Gb/s
 
−
*    2x 10Gbps ethernet ports (one connected in 1Gb/s)
 
−
*    1x infiniband FDR card
 
 
*    hyperthreading active
 
*    hyperthreading active
  
−
== Asus ESC8000 GPU node ==
+
'''HPE DL385''' dual-AMD EPYC 7662 @ 2.0GHz (64 cores) : 1 node ( '''nefgpu56''' )
−
 
+
* Nvidia A100 - 40GB PCIe GPUs cards
−
'''Asus ESC8000G4''' / Carri HighServer nodes : 1 node ('''nefgpu40''' )
+
**        6912 CUDA cores per card
−
* 8x GeForce GTX 1080 Ti GPUs cards connected with a PCIe gen3 16x interface
+
**        432 tensor cores per card
−
**        3584 CUDA cores per card
+
**        40GB of RAM capacity per card
−
**        11GB of RAM capacity per card  
+
**        Tensor performance peak : 312 TFlops per card
−
**        Simple precision performance peak: 10.6 Tflops per card
+
**        Simple precision performance peak: 156 Tflops per card
−
**        Double precision performance peak: 0.3 Tflops per card
+
**        Double precision performance peak: 19.49 Tflops per card
−
**        484 GB/s GPU memory bandwidth
+
**        1555 GB/s GPU memory bandwidth with error correction (ECC)
−
*    PCIe single-root topology (2 PCIe 96 lane switches)
+
*    2x gigabit network ports (one connected)
−
**        topology can be software modified (BIOS & reboot) to dual-root for an experiment campaign
+
*    infiniband EDR card (connected to FDR switch)
−
*    CPU : 2x Xeon SP Gold 5115 @ 2.4 GHz
 
−
*    RAM capacity : 256 GB
 
−
*    storage :
 
−
**        system & /tmp : RAID-1 2x512 GB SATA SSD
 
−
**        /local/tmp scratch data : RAID-0 4 TB SATA SSD
 
−
**        RAID controller SAS 12Gb/s
 
−
*    4x gigabit ethernet ports (one connected)
 
−
*    1x infiniband FDR card
 
 
*    hyperthreading active
 
*    hyperthreading active
 +
{| class="wikitable"
 +
|+Node details
 +
|-
 +
| '''Node name'''
 +
| '''Funding team'''
 +
| '''GPU cards'''
 +
| '''Node CPU'''
 +
| '''Node RAM'''
 +
| '''Node storage'''
 +
|-
 +
| nefgpu56
 +
| STARS
 +
| 3x A100 40GB
 +
| 2x AMD EPYC 7662
 +
| 1024 GB
 +
| system & /tmp : RAID-1 2x 480SSD<br>/local/mixed : RAID-5 5x 960 SSD
 +
|-
 +
| nefgpu57
 +
| MAASAI
 +
| 2x A100 80GB
 +
| 2x AMD EPYC 7513
 +
| 512 GB
 +
| system & /local/mixed : RAID-1 2x 480SSD<br>/local/mixed : RAID-0 6x 960 SSD
 +
|-
 +
| nefgpu58
 +
| ZENITH
 +
| 2x A100 80GB
 +
| 2x AMD EPYC 7513
 +
| 512 GB
 +
| system & /local/mixed : RAID-1 2x 480SSD<br>/local/mixed : RAID-0 6x 960 SSD
 +
|-
 +
| nefgpu59
 +
| MORPHEME
 +
| 3x A40
 +
| 2x AMD EPYC 7313
 +
| 512 GB
 +
| system & /local/mixed : RAID-1 2x 480SSD<br>/local/mixed : RAID-0 4x 960 SSD
 +
|-
 +
| nefgpu60
 +
| ACUMES
 +
| 2x A40
 +
| 2x AMD EPYC 7313
 +
| 256 GB
 +
| system & /local/mixed : RAID-1 2x 480SSD<br>/local/mixed : RAID-0 4x 960 SSD
 +
|-
 +
| nefgpu61
 +
| ACUMES
 +
| 2x A40
 +
| 2x AMD EPYC 7313
 +
| 256 GB
 +
| system & /local/mixed : RAID-1 2x 480SSD<br>/local/mixed : RAID-0 4x 960 SSD
 +
|}
  
 +
== Dell R7525 GPU nodes (2020) ==
  
−
== Dell R7525 GPU nodes ==
+
'''Dell R7525''' dual-AMD EPYC 7282 @ 2.8GHz (16 cores) : 2 nodes ( '''nefgpu52 to nefgpu53''' )<br>
−
 
+
'''Dell R7525''' dual-AMD EPYC 7413 @ 2.65GHz (24 cores) : 2 nodes ( '''nefgpu54 to nefgpu55''' )
−
'''Dell R7525''' dual-EPYC 7282 @ 2.8GHz (16 cores) : 2 nodes ( '''nefgpu52 to nefgpu53''' )<br>
 
−
'''Dell R7525''' dual-EPYC 7413 @ 2.65GHz (24 cores) : 2 nodes ( '''nefgpu54 to nefgpu55''' )
 
 
* Nvidia A40 PCIe GPUs cards
 
* Nvidia A40 PCIe GPUs cards
 
**        10752 CUDA cores per card
 
**        10752 CUDA cores per card
Ligne 227 : Ligne 240 :
 
*    infiniband EDR card (connected to FDR switch)
 
*    infiniband EDR card (connected to FDR switch)
 
*    hyperthreading active
 
*    hyperthreading active
−
*    optimal performance with AVX-256, AVX/AVX2 support
 
−
 
  
 
{| class="wikitable"
 
{| class="wikitable"
Ligne 254 : Ligne 265 :
 
| system & /tmp : RAID-1 2x 480SSD<br>/local/mixed : RAID-5 5x 480 SSD
 
| system & /tmp : RAID-1 2x 480SSD<br>/local/mixed : RAID-5 5x 480 SSD
 
|-
 
|-
−
| nefgpu53
+
| nefgpu54
 
| 3IA
 
| 3IA
 
| 3x A40
 
| 3x A40
Ligne 261 : Ligne 272 :
 
| system & /tmp : RAID-1 2x 480SSD<br>/local/mixed : RAID-5 5x 960 SSD
 
| system & /tmp : RAID-1 2x 480SSD<br>/local/mixed : RAID-5 5x 960 SSD
 
|-
 
|-
−
| nefgpu54
+
| nefgpu55
 
| 3IA
 
| 3IA
 
| 3x A40
 
| 3x A40
Ligne 269 : Ligne 280 :
 
|}
 
|}
  
−
== Dell R740 GPU nodes ==
+
 
 +
== SuperMicro 4029 node (2019) ==
 +
 
 +
'''SuperMicro 4029GP-TVRT''' : 1 node ('''nefgpu41''' )
 +
 
 +
[[Image:SYS-4028GR-TVRT.jpg|200px|right]]
 +
 
 +
* 4x Nvidia Tesla V100 SXM2 GPUs cards
 +
**        5120 CUDA cores per card
 +
**        640 tensor cores per card
 +
**        32GB of RAM capacity per card
 +
**        Simple precision performance peak: 15.7 Tflops per card
 +
**        Double precision performance peak: 7.8 Tflops per card
 +
**        Tensor performance peak : 125 TFlops per card
 +
**        900 GB/s GPU memory bandwidth with error correction (ECC)
 +
*    CPU-GPU connection with a PCIe gen3 16x interface
 +
*    GPU-GPU connection with NVLink 2.0 (25GB/s per direction per link, 1 or 2 links per card pair)
 +
*    CPU : 2x Xeon SP Gold 6126 @ 2.6 GHz (24 cores)
 +
*    RAM capacity : 384 GB
 +
*    storage :
 +
**        system & /tmp : 240GB SSD SATA RAID-1
 +
**        /local/tmp scratch data : 4.8 TB SSD SATA RAID-5
 +
**        RAID controller SAS 12Gb/s
 +
*    2x 10Gbps ethernet ports (one connected in 1Gb/s)
 +
*    1x infiniband FDR card
 +
*    hyperthreading active
 +
 
 +
== Asus ESC8000 GPU node (2018) ==
 +
 
 +
'''Asus ESC8000G4''' / Carri HighServer nodes : 1 node ('''nefgpu40''' )
 +
* 8x GeForce GTX 1080 Ti GPUs cards connected with a PCIe gen3 16x interface
 +
**        3584 CUDA cores per card
 +
**        11GB of RAM capacity per card
 +
**        Simple precision performance peak: 10.6 Tflops per card
 +
**        Double precision performance peak: 0.3 Tflops per card
 +
**        484 GB/s GPU memory bandwidth
 +
*    PCIe single-root topology (2 PCIe 96 lane switches)
 +
**        topology can be software modified (BIOS & reboot) to dual-root for an experiment campaign
 +
*    CPU : 2x Xeon SP Gold 5115 @ 2.4 GHz
 +
*    RAM capacity : 256 GB
 +
*    storage :
 +
**        system & /tmp : RAID-1 2x512 GB SATA SSD
 +
**        /local/tmp scratch data : RAID-0 4 TB SATA SSD
 +
**        RAID controller SAS 12Gb/s
 +
*    4x gigabit ethernet ports (one connected)
 +
*    1x infiniband FDR card
 +
*    hyperthreading active
 +
 
 +
== Dell R740 GPU nodes (2019) ==
  
 
'''Dell R740''' nodes: dual-Xeon Skylake or CascadeLake SP : 5 nodes ( '''nefgpu42 to nefgpu46''')  
 
'''Dell R740''' nodes: dual-Xeon Skylake or CascadeLake SP : 5 nodes ( '''nefgpu42 to nefgpu46''')  
Ligne 337 : Ligne 396 :
 
|}
 
|}
  
−
== Dell T640 GPU nodes ==
+
== Dell T640 GPU nodes (2018-2019-2020) ==
  
 
'''Dell T640''' nodes: dual-Xeon Skylake or CascadeLake SP : 21 nodes ( '''nefgpu{24-39} and nefgpu{47-51}''')  
 
'''Dell T640''' nodes: dual-Xeon Skylake or CascadeLake SP : 21 nodes ( '''nefgpu{24-39} and nefgpu{47-51}''')  
−
* Nvidia Quadro RTX 6000 GPUs cards connected with a PCIe gen3 16x interface
+
* Nvidia RTX A6000 GPUs cards connected with a PCIe gen3 16x interface (Ampere)
 +
**        10752 CUDA cores per card
 +
**        336 tensor cores per card
 +
**        48GB of RAM capacity per card
 +
**        Tensor performance peak : 309.7 TFlops per card
 +
**        Simple precision performance peak: 38.7 Tflops per card
 +
**        768 GB/s GPU memory bandwidth with error correction (ECC)
 +
* Nvidia Quadro RTX 6000 GPUs cards connected with a PCIe gen3 16x interface (Turing)
 
**        4608 CUDA cores per card
 
**        4608 CUDA cores per card
 
**        576 tensor cores per card
 
**        576 tensor cores per card
Ligne 350 : Ligne 416 :
 
**        624 GB/s GPU memory bandwidth with error correction (ECC)
 
**        624 GB/s GPU memory bandwidth with error correction (ECC)
 
**        optional NVLink 2.0 (2x25GB/s per direction) by pair of card
 
**        optional NVLink 2.0 (2x25GB/s per direction) by pair of card
−
* Nvidia Quadro RTX 8000 GPUs cards connected with a PCIe gen3 16x interface
+
* Nvidia Quadro RTX 8000 GPUs cards connected with a PCIe gen3 16x interface (Turing)
 
**        4608 CUDA cores per card
 
**        4608 CUDA cores per card
 
**        576 tensor cores per card
 
**        576 tensor cores per card
Ligne 358 : Ligne 424 :
 
**        Simple precision performance peak: 16.3 Tflops per card
 
**        Simple precision performance peak: 16.3 Tflops per card
 
**        Double precision performance peak: 0.5 Tflops per card
 
**        Double precision performance peak: 0.5 Tflops per card
−
**        624 GB/s GPU memory bandwidth with error correction (ECC)
+
**        672 GB/s GPU memory bandwidth with error correction (ECC)
 
**        optional NVLink 2.0 (2x25GB/s per direction) by pair of card
 
**        optional NVLink 2.0 (2x25GB/s per direction) by pair of card
 
* GeForce RTX 2080 Ti GPUs cards connected with a PCIe gen3 16x interface
 
* GeForce RTX 2080 Ti GPUs cards connected with a PCIe gen3 16x interface
Ligne 433 : Ligne 499 :
 
| nefgpu28
 
| nefgpu28
 
| GRAPHDECO
 
| GRAPHDECO
−
| 4x GTX 1080 Ti
+
| 4x RTX A6000
 
| 2x Xeon Silver 4110
 
| 2x Xeon Silver 4110
 
| 96 GB
 
| 96 GB
Ligne 551 : Ligne 617 :
 
|}
 
|}
  
−
== Dell T630 GPU nodes ==
+
== Dell T630 GPU nodes (2016-2017) ==
  
 
'''Dell T630''' nodes: dual-Xeon E5-26xx : 17 nodes ( '''nefgpu07 to nefgpu23''')  
 
'''Dell T630''' nodes: dual-Xeon E5-26xx : 17 nodes ( '''nefgpu07 to nefgpu23''')  
Ligne 723 : Ligne 789 :
 
|}
 
|}
  
 +
== Dell R730 GPU node (2016) ==
  
−
== Dell R730 GPU node ==
+
'''Dell R730''' nodes: dual-Xeon E5-2623v4 @ 2.6 GHz : 1 node ( '''nefgpu01''')  
−
 
 
−
'''Dell R730''' nodes: dual-Xeon E5-26xx : 1 node ( '''nefgpu01''')  
 
 
* Tesla K80 GPU cards connected with a PCIe gen3 16x interface
 
* Tesla K80 GPU cards connected with a PCIe gen3 16x interface
 
**        2x Tesla GK210 GPUs per card
 
**        2x Tesla GK210 GPUs per card

Version actuelle datée du 21 mars 2024 à 18:12



CPU clusters

Dell R7525 nodes (2020)

200px


Dell R7525 dual-EPYC 7502 @ 2.5GHz (64 cores) : 1 node ( nef059 )

  • RAM capacity : 384 GB RAM
  • storage : system 2x900 GB SATA SSD RAID-1 + local scratch data 7x7.5TB SAS HDD 44 TB RAID-5 + controller H745
  • 2x gigabit network ports (one connected)
  • infiniband EDR card (connected to FDR switch)
  • hyperthreading active


Dell R7525 dual-EPYC 7542 @ 2.9GHz (64 cores) : 1 node ( nef058 )

  • RAM capacity : 1024 GB RAM
  • storage : system 2x223 GB SATA SSD RAID-1 + local scratch data 5x445GB SAS SSD 1.8 TB RAID-5 + controller H745
  • 2x gigabit network ports (one connected)
  • infiniband EDR card (connected to FDR switch)
  • hyperthreading active

Dell C6420 cluster (2019)

Dell C6420 dual-Xeon Cascade Lake SP Gold 6240 @ 2.60GHz (36 cores) : 4 nodes ( nef054 to nef057 )

C6420.jpg


  • RAM capacity : 384 GB RAM
  • storage : system 2x600 GB SATA RAID-1 + local scratch data 960 BB SATA SSD RAID-0 + controller H330
  • 1x gigabit network port
  • 1x infiniband FDR card
  • hyperthreading active
  • optimal performance with AVX-512, AVX/AVX2 support

Dell R940 node (2017)

Dell R940 quad-Xeon SP Gold 6148 @ 2.40GHz (80 cores) : 1 node ( nef053 )

Dellr940.jpg


  • RAM capacity : 1024 GB RAM
  • storage : system 2x600 GB SATA RAID-1 + local scratch data 1.92 TB SATA SSD + controller H740P
  • 4x gigabit network ports (one connected)
  • infiniband EDR card (connected to FDR switch)
  • hyperthreading active
  • optimal performance with AVX-512, AVX/AVX2 support

Dell C6420 cluster (2017)

Dell C6420 dual-Xeon Skylake SP Silver 4114 @ 2.20GHz (20 cores) : 16 nodes ( nef037 to nef052 )

C6420.jpg


  • RAM capacity : 192 GB RAM
  • 1x600GB 10kRPM SAS HardDisk drive
  • 1x gigabit network port
  • 1x infiniband FDR card
  • hyperthreading active
  • AVX-512 support, optimal performance with AVX/AVX2

Dell C6220 cluster (2015)

Dell C6220 dual-Xeon E5-2650 v2 @ 2.60GHz (16 cores) : 8 nodes ( nef029 to nef036 )

C6220-300x191.jpg


  • RAM capacity : 256 GB RAM
  • 1x1TB SATA HardDisk drive
  • 2x gigabit network ports (one connected)
  • 2x infiniband QDR card (one connected)
  • hyperthreading not active


Dell C6220 cluster (2014)

Dell C6220 dual-Xeon E5-2680 v2 @ 2.80GHz (20 cores) : 16 nodes ( nef013 to nef028 )

C6220-300x191.jpg


  • RAM capacity : 192 GB RAM
  • 1x2TB SATA HardDisk drive
  • 2x gigabit network ports (one connected)
  • 1x infiniband FDR card (QDR used)
  • hyperthreading not active


 Dell C6145 cluster (2013)

Dell C6145 quad-Opterons 6376 @ 2.3Ghz (64 cores) : 6 nodes ( nef007 to nef012 )

Cluster-c6145.jpg


  • RAM capacity : 256 GB RAM (512GB on nef011 and nef012)
  • 1x500GB SATA HardDisk drive
  • 2x gigabit network ports (one connected)
  • 1x infiniband QDR card
  • hyperthreading not supported


Dell R815 cluster (2010)

Dell R815 quad-Opterons 6174 @ 2.2Ghz (48 cores) : 6 nodes ( nef001 to nef006 )

Cluster-dellr815.jpg


  • RAM capacity : 256 GB RAM
  • 2x600GB SAS HardDisk drive (RAID-0)
  • 4x gigabit network ports (one connected)
  • 1x infiniband QDR card
  • hyperthreading not supported

GPU nodes

HPE DL385 GPU nodes (2023)

HPE DL385 dual-AMD EPYC 7313 @ 3.0GHz (16 cores) : 3 nodes ( nefgpu59 to nefgpu61 )

  • Nvidia A40 PCIe GPUs cards
    • 10752 CUDA cores per card
    • 336 tensor cores per card
    • 48GB of RAM capacity per card
    • Tensor performance peak : 149.6 TFlops per card
    • Simple precision performance peak: 74.8 Tflops per card
    • Double precision performance peak: 37.4 Tflops per card
    • 696 GB/s GPU memory bandwidth with error correction (ECC)
  • 2x gigabit network ports (one connected)
  • infiniband EDR card (connected to FDR switch)
  • hyperthreading active

HPE DL385 dual-AMD EPYC 7513 @ 2.6GHz (32 cores) : 2 nodes ( nefgpu57 to nefgpu58 )

  • Nvidia A100 - 80GB PCIe GPUs cards
    • 6912 CUDA cores per card
    • 432 tensor cores per card
    • 80GB of RAM capacity per card
    • Tensor performance peak : 312 TFlops per card
    • FP16 Compute: 78 Tflops per card
    • FP32 Compute: 19.5 Tflops per card
    • FP64 Compute: 9.7 Tflops per card
    • 1935GB/s GPU memory bandwidth with error correction (ECC)
  • 2x gigabit network ports (one connected)
  • infiniband EDR card (connected to FDR switch)
  • hyperthreading active

HPE DL385 dual-AMD EPYC 7662 @ 2.0GHz (64 cores) : 1 node ( nefgpu56 )

  • Nvidia A100 - 40GB PCIe GPUs cards
    • 6912 CUDA cores per card
    • 432 tensor cores per card
    • 40GB of RAM capacity per card
    • Tensor performance peak : 312 TFlops per card
    • Simple precision performance peak: 156 Tflops per card
    • Double precision performance peak: 19.49 Tflops per card
    • 1555 GB/s GPU memory bandwidth with error correction (ECC)
  • 2x gigabit network ports (one connected)
  • infiniband EDR card (connected to FDR switch)
  • hyperthreading active
Node details
Node name Funding team GPU cards Node CPU Node RAM Node storage
nefgpu56 STARS 3x A100 40GB 2x AMD EPYC 7662 1024 GB system & /tmp : RAID-1 2x 480SSD
/local/mixed : RAID-5 5x 960 SSD
nefgpu57 MAASAI 2x A100 80GB 2x AMD EPYC 7513 512 GB system & /local/mixed : RAID-1 2x 480SSD
/local/mixed : RAID-0 6x 960 SSD
nefgpu58 ZENITH 2x A100 80GB 2x AMD EPYC 7513 512 GB system & /local/mixed : RAID-1 2x 480SSD
/local/mixed : RAID-0 6x 960 SSD
nefgpu59 MORPHEME 3x A40 2x AMD EPYC 7313 512 GB system & /local/mixed : RAID-1 2x 480SSD
/local/mixed : RAID-0 4x 960 SSD
nefgpu60 ACUMES 2x A40 2x AMD EPYC 7313 256 GB system & /local/mixed : RAID-1 2x 480SSD
/local/mixed : RAID-0 4x 960 SSD
nefgpu61 ACUMES 2x A40 2x AMD EPYC 7313 256 GB system & /local/mixed : RAID-1 2x 480SSD
/local/mixed : RAID-0 4x 960 SSD

Dell R7525 GPU nodes (2020)

Dell R7525 dual-AMD EPYC 7282 @ 2.8GHz (16 cores) : 2 nodes ( nefgpu52 to nefgpu53 )
Dell R7525 dual-AMD EPYC 7413 @ 2.65GHz (24 cores) : 2 nodes ( nefgpu54 to nefgpu55 )

  • Nvidia A40 PCIe GPUs cards
    • 10752 CUDA cores per card
    • 336 tensor cores per card
    • 48GB of RAM capacity per card
    • Tensor performance peak : 149.6 TFlops per card
    • Simple precision performance peak: 74.8 Tflops per card
    • Double precision performance peak: 37.4 Tflops per card
    • 696 GB/s GPU memory bandwidth with error correction (ECC)
  • 2x gigabit network ports (one connected)
  • infiniband EDR card (connected to FDR switch)
  • hyperthreading active
Node details
Node name Funding team GPU cards Node CPU Node RAM Node storage
nefgpu52 NEO 3x A40 2x AMD EPYC 7282 256 GB system & /tmp : RAID-1 2x 480SSD
/local/mixed : RAID-5 5x 480 SSD
nefgpu53 ACENTAURI 3x A40 2x AMD EPYC 7282 256 GB system & /tmp : RAID-1 2x 480SSD
/local/mixed : RAID-5 5x 480 SSD
nefgpu54 3IA 3x A40 2x AMD EPYC 7413 512 GB system & /tmp : RAID-1 2x 480SSD
/local/mixed : RAID-5 5x 960 SSD
nefgpu55 3IA 3x A40 2x AMD EPYC 7413 512 GB system & /tmp : RAID-1 2x 480SSD
/local/mixed : RAID-5 5x 960 SSD


SuperMicro 4029 node (2019)

SuperMicro 4029GP-TVRT : 1 node (nefgpu41 )

SYS-4028GR-TVRT.jpg
  • 4x Nvidia Tesla V100 SXM2 GPUs cards
    • 5120 CUDA cores per card
    • 640 tensor cores per card
    • 32GB of RAM capacity per card
    • Simple precision performance peak: 15.7 Tflops per card
    • Double precision performance peak: 7.8 Tflops per card
    • Tensor performance peak : 125 TFlops per card
    • 900 GB/s GPU memory bandwidth with error correction (ECC)
  • CPU-GPU connection with a PCIe gen3 16x interface
  • GPU-GPU connection with NVLink 2.0 (25GB/s per direction per link, 1 or 2 links per card pair)
  • CPU : 2x Xeon SP Gold 6126 @ 2.6 GHz (24 cores)
  • RAM capacity : 384 GB
  • storage :
    • system & /tmp : 240GB SSD SATA RAID-1
    • /local/tmp scratch data : 4.8 TB SSD SATA RAID-5
    • RAID controller SAS 12Gb/s
  • 2x 10Gbps ethernet ports (one connected in 1Gb/s)
  • 1x infiniband FDR card
  • hyperthreading active

Asus ESC8000 GPU node (2018)

Asus ESC8000G4 / Carri HighServer nodes : 1 node (nefgpu40 )

  • 8x GeForce GTX 1080 Ti GPUs cards connected with a PCIe gen3 16x interface
    • 3584 CUDA cores per card
    • 11GB of RAM capacity per card
    • Simple precision performance peak: 10.6 Tflops per card
    • Double precision performance peak: 0.3 Tflops per card
    • 484 GB/s GPU memory bandwidth
  • PCIe single-root topology (2 PCIe 96 lane switches)
    • topology can be software modified (BIOS & reboot) to dual-root for an experiment campaign
  • CPU : 2x Xeon SP Gold 5115 @ 2.4 GHz
  • RAM capacity : 256 GB
  • storage :
    • system & /tmp : RAID-1 2x512 GB SATA SSD
    • /local/tmp scratch data : RAID-0 4 TB SATA SSD
    • RAID controller SAS 12Gb/s
  • 4x gigabit ethernet ports (one connected)
  • 1x infiniband FDR card
  • hyperthreading active

Dell R740 GPU nodes (2019)

Dell R740 nodes: dual-Xeon Skylake or CascadeLake SP : 5 nodes ( nefgpu42 to nefgpu46)

  • Nvidia Tesla V100 PCIe GPUs cards
    • 5120 CUDA cores per card
    • 640 tensor cores per card
    • 32GB of RAM capacity per card
    • Tensor performance peak : 112 TFlops per card
    • Simple precision performance peak: 14 Tflops per card
    • Double precision performance peak: 7 Tflops per card
    • 900 GB/s GPU memory bandwidth with error correction (ECC)
  • Nvidia Tesla T4 GPUs cards connected with a PCIe gen3 16x interface
    • 2560 CUDA cores per card
    • 320 tensor cores per card
    • 16GB of RAM capacity per card
    • Simple precision performance peak: 8.1 Tflops per card
    • Double precision performance peak: 0.3 Tflops per card
    • 300 GB/s GPU memory bandwidth with error correction (ECC)
  • 2-4x ethernet ports (one connected in gigabit)
  • 1x infiniband FDR card
  • hyperthreading active
Node details
Node name Funding team GPU cards Node CPU Node RAM Node storage
nefgpu42 common 3x Tesla T4 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/tmp : RAID-0 2x 480GB SSD
nefgpu43 common 3x Tesla T4 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/tmp : RAID-0 2x 480GB SSD
nefgpu44 common 3x Tesla T4 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/tmp : RAID-0 2x 480GB SSD
nefgpu45 common 3x Tesla T4 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/tmp : RAID-0 2x 480GB SSD
nefgpu46 STARS 3x Tesla V100 2x Xeon Silver 4215 384 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-5 5x 960GB SSD

Dell T640 GPU nodes (2018-2019-2020)

Dell T640 nodes: dual-Xeon Skylake or CascadeLake SP : 21 nodes ( nefgpu{24-39} and nefgpu{47-51})

  • Nvidia RTX A6000 GPUs cards connected with a PCIe gen3 16x interface (Ampere)
    • 10752 CUDA cores per card
    • 336 tensor cores per card
    • 48GB of RAM capacity per card
    • Tensor performance peak : 309.7 TFlops per card
    • Simple precision performance peak: 38.7 Tflops per card
    • 768 GB/s GPU memory bandwidth with error correction (ECC)
  • Nvidia Quadro RTX 6000 GPUs cards connected with a PCIe gen3 16x interface (Turing)
    • 4608 CUDA cores per card
    • 576 tensor cores per card
    • 24GB of RAM capacity per card
    • Tensor performance peak : 130.5 TFlops per card
    • Half precision performance peak: 32.6 Tflops per card
    • Simple precision performance peak: 16.3 Tflops per card
    • Double precision performance peak: 0.5 Tflops per card
    • 624 GB/s GPU memory bandwidth with error correction (ECC)
    • optional NVLink 2.0 (2x25GB/s per direction) by pair of card
  • Nvidia Quadro RTX 8000 GPUs cards connected with a PCIe gen3 16x interface (Turing)
    • 4608 CUDA cores per card
    • 576 tensor cores per card
    • 48GB of RAM capacity per card
    • Tensor performance peak : 130.5 TFlops per card
    • Half precision performance peak: 32.6 Tflops per card
    • Simple precision performance peak: 16.3 Tflops per card
    • Double precision performance peak: 0.5 Tflops per card
    • 672 GB/s GPU memory bandwidth with error correction (ECC)
    • optional NVLink 2.0 (2x25GB/s per direction) by pair of card
  • GeForce RTX 2080 Ti GPUs cards connected with a PCIe gen3 16x interface
    • 4352 CUDA cores per card
    • 544 tensor cores per card
    • 11GB of RAM capacity per card
    • Half precision performance peak: 23.5 Tflops per card
    • Simple precision performance peak: 11.8 Tflops per card
    • Double precision performance peak: 0.4 Tflops per card
    • 616 GB/s GPU memory bandwidth
  • GeForce Titan RTX GPUs cards connected with a PCIe gen3 16x interface
    • 4608 CUDA cores per card
    • 576 tensor cores per card
    • 24GB of RAM capacity per card
    • Half precision performance peak: 24.9 Tflops per card
    • Simple precision performance peak: 12.4 Tflops per card
    • Double precision performance peak: 0.4 Tflops per card
    • 672 GB/s GPU memory bandwidth
    • Power limited to 210W to prevent overheat
  • GeForce GTX 1080 Ti GPUs cards connected with a PCIe gen3 16x interface
    • 3584 CUDA cores per card
    • 11GB of RAM capacity per card
    • Simple precision performance peak: 10.6 Tflops per card
    • Double precision performance peak: 0.3 Tflops per card
    • 484 GB/s GPU memory bandwidth
  • GeForce GTX Titan X GPUs cards connected with a PCIe gen3 16x interface
    • 3072 CUDA cores per card
    • 12GB of RAM capacity per card
    • Simple precision performance peak: 7.0 Tflops per card
    • Double precision performance peak: 0.2 Tflops per card
    • 336.5 GB/s GPU memory bandwidth
  • 2-4x ethernet ports (one connected in gigabit)
  • 1x infiniband FDR card
  • hyperthreading active
Node details
Node name Funding team GPU cards Node CPU Node RAM Node storage
nefgpu24 EPIONE 4x GTX 1080 Ti 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local : RAID-0 1x 960GB SSD
nefgpu25 ZENITH 4x GTX 1080 Ti 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-0 4x 480GB SSD
/local/read : RAID-0 8x 480GB SSD read intensive
nefgpu26 WIMMICS 4x GTX 1080 Ti 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local : RAID-0 4x 480GB SSD
nefgpu27 ZENITH 4x RTX 2080 Ti 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-0 2x 400GB SSD
/local/read : RAID-0 4x 480GB SSD read intensive
nefgpu28 GRAPHDECO 4x RTX A6000 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-5 5x 400GB SSD
nefgpu29 EPIONE 2x GTX Titan-X 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-5 5x 480GB SSD
nefgpu30 common 4x RTX 2080 Ti 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/tmp : RAID-0 4x 480GB SSD
nefgpu31 common 4x RTX 2080 Ti 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/tmp : RAID-0 4x 480GB SSD
nefgpu32 common 4x Titan RTX 2x Xeon Silver 4110 192 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/tmp : RAID-0 4x 480GB SSD
nefgpu33 CHORALE 4x RTX 2080 Ti 2x Xeon Silver 4110 96 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-0 1x 480GB SSD
nefgpu34 GRAPHDECO 4x RTX 6000
w/ NVLink by pair 		2x Xeon Gold 6240 384 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-5 5x 960GB SSD
nefgpu35 STARS 4x RTX 6000
w/ NVLink by pair 		2x Xeon Silver 4215 192 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-5 5x 960GB SSD
nefgpu36 STARS 4x RTX 6000
w/ NVLink by pair 		2x Xeon Silver 4215 192 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-5 5x 960GB SSD
nefgpu37 GRAPHDECO 4x RTX 8000
w/ NVLink by pair 		2x Xeon Gold 6240 384 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-5 5x 960GB SSD
nefgpu38 AYANA 2x RTX 8000
w/ NVLink by pair 		2x Xeon Silver 4214R 384 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-5 5x 960GB SSD
nefgpu39 STARS 4x RTX 8000
w/ NVLink by pair 		2x Xeon Silver 4216 384 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/mixed : RAID-5 5x 900GB SATA SSD
nefgpu47 I3S/SPARKS 4x RTX 8000
w/ NVLink by pair 		2x Xeon Gold 6230R 384 GB system & /tmp : RAID-1 2x 450GB SATA SSD
/local/mixed : RAID-5 5x 900GB SSD
nefgpu48 I3S/SPARKS 4x RTX 8000
w/ NVLink by pair 		2x Xeon Gold 6230R 384 GB system & /tmp : RAID-1 2x 450GB SATA SSD
/local/mixed : RAID-5 5x 900GB SSD
nefgpu49 I3S/SPARKS 4x RTX 8000
w/ NVLink by pair 		2x Xeon Gold 6238R 768 GB system & /tmp : RAID-1 2x 450GB SATA SSD
/local/mixed : RAID-5 5x 1.8TB SSD
nefgpu50 I3S/SPARKS 4x RTX 8000
w/ NVLink by pair 		2x Xeon Gold 6230R 384 GB system & /tmp : RAID-1 2x 450GB SATA SSD
/local/mixed : RAID-5 5x 900GB SSD
nefgpu51 I3S/SPARKS 4x RTX 8000
w/ NVLink by pair 		2x Xeon Gold 6230R 384 GB system & /tmp : RAID-1 2x 450GB SATA SSD
/local/mixed : RAID-5 5x 900GB SSD

Dell T630 GPU nodes (2016-2017)

Dell T630 nodes: dual-Xeon E5-26xx : 17 nodes ( nefgpu07 to nefgpu23)

  • GeForce GTX 1080 Ti GPUs cards connected with a PCIe gen3 16x interface
    • 3584 CUDA cores per card
    • 11GB of RAM capacity per card
    • Simple precision performance peak: 10.6 Tflops per card
    • Double precision performance peak: 0.3 Tflops per card
    • 484 GB/s GPU memory bandwidth
  • GeForce GTX 1080 GPUs cards connected with a PCIe gen3 16x interface
    • 2560 CUDA cores per card
    • 8GB of RAM capacity per card
    • Simple precision performance peak: 8.2 Tflops per card
    • Double precision performance peak: 0.3 Tflops per card
    • 320 GB/s GPU memory bandwidth
  • GeForce GTX Titan X GPUs cards connected with a PCIe gen3 16x interface
    • 3072 CUDA cores per card
    • 12GB of RAM capacity per card
    • Simple precision performance peak: 7.0 Tflops per card
    • Double precision performance peak: 0.2 Tflops per card
    • 336.5 GB/s GPU memory bandwidth
  • 4x gigabit ethernet ports (one connected)
  • 1x infiniband FDR card
Node details
Node name Funding team GPU cards Node CPU Node RAM Node storage Hyper threading active ?
nefgpu07 EPIONE 4x GTX 1080 Ti 2x E5-2620v3 128 GB system & /tmp : RAID-0 2x 1TB SATA 7.2kRPM no
nefgpu08 ZENITH 4x GTX 1080 Ti 2x E5-2630v3 64 GB system & /tmp : RAID-0 2x 300GB SAS 15kRPM
/local/mixed : 1x 800GB SSD
/local/read : RAID-0 2x 1.92TB SSD read intensive 		no
nefgpu09 GRAPHDECO 4x Titan X 2x E5-2630v4 48 GB system & /tmp : RAID-0 2x 1TB SATA 7.2kRPM
/local : 1x 400GB SSD 		no
nefgpu10 STARS 4x Titan X 2x E5-2630v4 128 GB system & /tmp : RAID-0 2x 1TB SATA 7.2kRPM
/local : 1x 1.6TB SSD 		no
nefgpu11 STARS 4x GTX 1080 2x E5-2630v4 128 GB system & /tmp : RAID-0 2x 1TB SATA 7.2kRPM
/local : 1x 1.6TB SSD 		no
nefgpu12 STARS 4x GTX 1080 2x E5-2630v4 128 GB system & /tmp : RAID-0 2x 1TB SATA 7.2kRPM
/local : 1x 1.6TB SSD 		yes
nefgpu13 GRAPHDECO 4x GTX 1080 Ti 2x E5-2650v4 64 GB system & /tmp : RAID-0 2x 1TB SATA 7.2kRPM
/local : 1x 400GB SSD 		yes
nefgpu14 STARS 4x GTX 1080 Ti 2x E5-2620v4 128 GB system & /tmp : RAID-0 2x 1TB SATA 7.2kRPM
/local : 1x 400GB SSD 		yes
nefgpu15 STARS 4x GTX 1080 Ti 2x E5-2620v4 128 GB system & /tmp : RAID-0 2x 1TB SATA 7.2kRPM
/local : 1x 400GB SSD 		yes
nefgpu16 EPIONE 4x GTX 1080 Ti 2x E5-2630v4 128 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local : 1x 1.6TB SSD 		yes
nefgpu17 ZENITH 4x GTX 1080 Ti 2x E5-2630v4 64 GB system & /tmp : RAID-0 2x 600GB SAS 10kRPM
/local/mixed : 1x 1.6TB SSD
/local/read : RAID-0 2x 1.92TB SSD read intensive 		yes
nefgpu18 common 4x GTX 1080 Ti 2x E5-2630v4 128 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/tmp : 1x 1.6TB SSD 		yes
nefgpu19 common 4x GTX 1080 Ti 2x E5-2630v4 128 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/tmp : 1x 1.6TB SSD 		yes
nefgpu20 common 4x GTX 1080 Ti 2x E5-2630v4 128 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local/tmp : 1x 1.6TB SSD 		yes
nefgpu21 STARS 4x GTX 1080 Ti 2x E5-2620v4 128 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local : 1x 480GB SSD 		yes
nefgpu22 STARS 3x GTX 1080 Ti 2x E5-2620v4 128 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local : 1x 480GB SSD 		yes
nefgpu23 TITANE-EPITOME 4x GTX 1080 Ti 2x E5-2630v4 64 GB system & /tmp : RAID-1 2x 600GB SAS 10kRPM
/local : 1x 1.6TB SSD 		yes

Dell R730 GPU node (2016)

Dell R730 nodes: dual-Xeon E5-2623v4 @ 2.6 GHz : 1 node ( nefgpu01)

  • Tesla K80 GPU cards connected with a PCIe gen3 16x interface
    • 2x Tesla GK210 GPUs per card
    • 4992 CUDA cores per card
    • 2x 12GB RAM capacity per card with error correction (ECC)
    • Simple precision performance peak: 5.61 Tflops per card
    • Double precision performance peak: 1.87 Tflops per card
    • 2x 240 GB/s GPU memory bandwidth with error correction (ECC)
  • 4x gigabit ethernet ports (one connected)
  • 1x infiniband QDR card
  • hyperthreading not active
Node details
Node name Funding team Number of GPU cards Node CPU Node RAM Node storage
nefgpu01 MATHNEURO 1x K80 2x E5-2623v4 32 GB 2x 400GB SSD

 Storage

All nodes have access to common storage :

  • common storage : /home
    • 31 TiB, available to all users, quotas per user
    • 1 Dell PowerEdge R540 server with 2 RAID-10 array 10 x 10TB SAS 7.2 kRPM disks, infiniband FDR, NFS access
  • capacity distributed and scalable common storage : /data
    • ~610TiB real capacity, 684TB theoretical capacity (06/2019)
      • permanent storage : 1TiB quota per team + teams may buy additional quota (please contact cluster administrators)
      • scratch storage : variable size (initially ~40TiB), no quota limit, for temporary storage (data may be purged)
    • BeeGFS filesystem on multiple hardware :
      • 4 Dell PowerEdge R740xd ; 1.92TB metadata : RAID-1 array 2 x 1.92TB SSD mixed use MLC disks (or RAID-10 array 4 x 960GB) ; 2x {60 or 72}TB data : 2 x RAID-6 array 8 x {10 or 12}TB SAS 7.2 kRPM disks
      • 2 Dell PowerEdge R730xd ; 800GB metadata : RAID-1 array 2 x 800 GB SSD mixed use MLC disks ; 2 x {36 or 48}TB data: 2 x RAID-6 array 8 x {6 or 8}TB SAS 7.2 kRPM disks
      • infiniband FDR

More details about quotas at: Disk space management

Récupérée de « https://wiki.inria.fr/wikis/ClustersSophia/index.php?title=Hardware&oldid=1132 »