IBM Cloud Docs
Accelerated profile family - Gen 3

Accelerated profile family - Gen 3

The accelerated family of profiles provides on-demand, cost-effective access to NVIDIA GPUs. GPUs help to accelerate the processing time that is required for compute intensive workloads such as AI, machine learning, inferencing, and more.

NVIDIA Hopper HGX instance profiles

The Hopper-based Accelerated virtual server profiles are built atop NVIDIA H100 accelerators. These accelerators are tuned for AI workloads, including inferencing, fine-tuning, and large-scale training. The solution is paired with the 4th Generation Intel® Xeon® Scalable processors.

Select availability This solution also runs with IBM Cloud® cluster networks. The cluster network implementation for the Hopper generation of accelerators runs atop eight accelerated NICs, providing a total aggregate cluster throughput of 3.2 Tbps. The solution also provides RoCEv2 to support RDMA-based workloads. For more information, see About cluster networks.

Operating systems

  • Linux

Processor generation

  • Intel 8474C - 4th Generation Xeon® Scalable processor

Accelerator

  • NVIDIA H100 SXM5 (80 GB)

Availability

Status: Select Availability

Table 1. Supported regions and zones
Region Universal zone Cluster network
Dallas (us-south) us-south-dal10-a No
Washington DC (us-east) us-east-wdc07-a Yes
Toronto (ca-tor) ca-tor-tor05-a No
Sao Paulo (br-sao) br-sao-sao01-a No
Frankfurt (eu-de) eu-de-fra04-a Yes
London (eu-gb) eu-gb-lon05-a No
Madrid (eu-es) eu-es-mad05-a No
Sydney (au-syd) au-syd-syd04-a No
Tokyo (jp-tok) jp-tok-tok05-a No
Osaka (jp-osa) Not Available No

For more information about regions and universal zones, see Regions.

Capabilities

  • Core type: Dedicated
  • Dedicated host: No
  • Hyperthreading: Yes (SMT-2)
  • Secure boot: No
  • Confidential computing: No
  • Live migration: No
  • Instance storage: Yes
  • NVLink: Yes (900 GB/s)
  • NVIDIA GPUDirect Capable: Yes
  • Cluster network capable: Yes (limited regions)
    • Bandwidth: 3.2 Tbps (8x 400 Gbps)
    • Type: Dedicated

VM configuration

  • Hardware type: q35
  • Cloud networking: virtio
  • Cluster networking: SR-IOV
    • Type: NVIDIA CX-7 – Virtual Function
    • Quantity: 8x Dedicated 400 Gbps Physical NICs
  • Block boot volume: virtio
  • Block data volumes: virtio
  • Instance storage: NVMe

Instance profiles

Accelerated NVIDIA Hopper HGX profile options
Instance profile vCPU / Cores Memory (GiB) Bandwidth cap (Gbps) Dedicated cluster network bandwidth Accelerators Instance storage (GB)
gx3d-160x1792x8h100 160 / 80 1792 200 3.2 Tbps (8x 400 Gbps Dedicated NVIDIA CX-7) 8x NVIDIA H100 (80 GB) 8 x 7.68 TB

Limits

An instance has a limit for the number of volumes and virtual network interfaces that can be attached. This limit is based on the size of the instance.

Accelerated NVIDIA Hopper HGX limits for vCPU, maximum volumes, and maximum network interfaces
Number of vCPUs Max volumes Max vNICs
2-16 15 5
17-48 15 10
49+ 15 15

If you configure an RDMA-enabled cluster network, you must have either 8, 16 or 32 cluster network interfaces available. Having the correct number of cluster network interfaces available helps ensure proper distribution of the network interfaces across the underlying physical infrastructure. Most users typically use only 8. The cluster network interfaces can be configured only when the instance is powered off.

NVIDIA L4 instance profiles

The virtual server profiles are built atop NVIDIA L4 accelerators. These accelerators are tuned for graphics workloads. The solution is paired with the 4th Generation Intel® Xeon® Scalable processors.

Operating systems

  • Windows
  • Linux

Processor generation

  • Intel 8474C - 4th Generation Xeon® Scalable processor

Accelerator

  • NVIDIA L4 GPU (24 GB)

Availability

Status: Generally Available

Regions:

  • Americas
    • Sao Paulo (br-sao)
    • Toronto (ca-tor)
    • Dallas (us-south)
    • Washington DC (us-east)
  • Europe
    • Frankfurt (eu-de)
    • London (eu-gb)
    • Madrid (eu-es)
  • Asia Pacific
    • Sydney (au-syd)
    • Tokyo (jp-tok)

Capabilities

  • Core type: Dedicated
  • Dedicated host: Yes
  • Hyperthreading: Yes (SMT-2)
  • Secure boot: No
  • Confidential computing: No
  • Live migration: No
  • Instance storage: No
  • NVLink: No

VM configuration

  • Hardware type: i440fx
  • Cloud networking: virtio
  • Block boot volume: virtio
    • Exception: vscsi for Windows-based instances
  • Block data volumes: virtio

Instance profiles

Accelerated l4 profile options
vCPUs / Cores Memory (GiB) Bandwidth Cap (Gbps) Accelerators
gx3-16x80x1l4 16 / 8 80 32 1x NVIDIA L4 (24 GB)
gx3-32x160x2l4 32 / 16 160 64 2x NVIDIA L4 (24 GB)
gx3-64x320x4l4 64 / 32 320 128 4x NVIDIA L4 (24 GB)

Limits

An instance has a limit for the number of volumes and virtual network interfaces that can be attached. This limit is based on the size of the instance.

Accelerated L4 limits for vCPU, maximum volumes, and maximum network interfaces
Number of vCPUs Max volumes Max vNICs
2-16 15 5
17-48 15 10
49+ 15 15

NVIDIA L40S instance profiles

The L40s profiles are built atop NVIDIA L40s accelerators. These accelerators are tuned for graphics and inferencing workloads. The solution is paired with the 4th Generation Intel® Xeon® Scalable processors.

Operating systems

  • Windows
  • Linux

Processor generation

  • Intel 8474C - 4th Generation Xeon® Scalable processor

Accelerator

  • NVIDIA L40s GPU (48 GB)

Availability

Status: Generally Available

Regions:

  • Americas
    • Sao Paulo (br-sao)
    • Toronto (ca-tor)
    • Dallas (us-south)
    • Washington DC (us-east)
  • Europe
    • Frankfurt (eu-de)
    • London (eu-gb)
    • Madrid (eu-es)
  • Asia Pacific
    • Sydney (au-syd)
    • Tokyo (jp-tok)

Capabilities

  • Core type: Dedicated
  • Dedicated host: Yes
  • Hyperthreading: Yes (SMT-2)
  • Secure boot: No
  • Confidential computing: No
  • Live migration: No
  • Instance storage: No
  • NVLink: No

VM configuration

  • Hardware type: i440fx
  • Cloud networking: virtio
  • Block boot volume: virtio
    • Exception: vscsi for Windows-based instances
  • Block data volumes: virtio

Instance profiles

Accelerated L40s profile options
Instance profile vCPUs / Cores Memory (GiB) Bandwidth Cap (Gbps) Accelerators
gx3-24x120x1l40s 24 / 12 120 48 1x NVIDIA L40s (48 GB)
gx3-48x240x-2l40s 48 / 24 240 96 2x NVIDIA L40s (48 GB)

Limits

An instance has a limit for the number of volumes and virtual network interfaces that can be attached. This limit is based on the size of the instance.

Accelerated L40s limits for vCPU, maximum volumes, and maximum network interfaces
Number of vCPUs Max volumes Max vNICs
2-16 15 5
17-48 15 10
49+ 15 15