Skip to content

Latest commit

 

History

History
115 lines (87 loc) · 6.28 KB

gap-35_gpu_pci_capability.md

File metadata and controls

115 lines (87 loc) · 6.28 KB
Raw
gap title description author status type
GAP-35
GPU/PCI capability
Golem Computing Resource Standard properties for elements of GPU/PCI specification
Stanislaw Krotki
Draft
<Experimental>

Abstract

GPU/PCI capability information in the Offer

Motivation

This GAP introduces GPU (and possibly other PCI) devices capability description. Providers will specify this information so that Requestors would be able to specify certain processing requirements in Demands.

Specification

  • golem: {
    • !exp: {
      • gap-35: {
        • v1: {
          • inf: {
            • gpu: {
              • model: string,
              • cuda: {
                • enabled: boolean,
                • cores: integer,
                • version: integer,
              • },
              • clocks: {
                • graphics.mhz: integer,
                • memory.mhz: integer,
                • sm.mhz: integer,
                • video.mhz: integer,
              • },
              • memory: {
                • bandwidth.gib: integer
                • total.gib: number
              • },
            • }
          • }
        • }
      • }
    • } }

inf-gpu

Namespace that describes GPU capabilities

Properties

Property Type Applies to Description
golem.!exp.gap-35.v1.inf.gpu.model string Offer Indicates the name of the GPU model

inf-gpu-cuda

Namespace that describes GPU CUDA capabilities

Properties

Property Type Applies to Description
golem.!exp.gap-35.v1.inf.gpu.cuda.enabled boolean Offer this is related to the image running and not the VM runtime
golem.!exp.gap-35.v1.inf.gpu.cuda.cores integer Offer This can be found from the general specification of each GPU. Otherwise it is platform specific that we would not like to work with. We should aim to focus on platform independent metrics that are easy to access in a health check
golem.!exp.gap-35.v1.inf.gpu.cuda.version string Offer A version of CUDA that is guaranteed to work. The range of compatibile versions is not known by the GPU because there is a minimum CUDA version and also it can be deprecated in a future CUDA version. It can also be different for CUDA and a given library, for example tensorflow had a different GPU compatibility requirement for building from source and using prebuild binaries. Get the info from Nvidia and/or a general guideline from StackOverflow. The property is a version string including a major version, and a minor version in #.# format.
golem.!exp.gap-35.v1.inf.gpu.cuda.compute-capability string Offer Every Nvidia GPU has Compute Capability level. It allows to identify CUDA features it supports. The property is a version string including a major version, and a minor version in #.# format.

inf-gpu-clocks

Namespace that describes information about GPU clocks

Properties

Property Type Applies to Description
golem.!exp.gap-35.v1.inf.gpu.clocks.graphics.mhz integer Offer The max rate of the graphics clock as reported by nvidia-smi
golem.!exp.gap-35.v1.inf.gpu.clocks.memory.mhz integer Offer The max rate of the memory clock as reported by nvidia-smi
golem.!exp.gap-35.v1.inf.gpu.clocks.sm.mhz integer Offer The max rate of the streaming multiprocessor clock as reported by nvidia-smi. CUDA cores are driven by this clock.
golem.!exp.gap-35.v1.inf.gpu.clocks.video.mhz integer Offer The max rate of the video clock as reported by nvidia-smi

inf-gpu-memory

Namespace that describes information about GPU memory

Properties

Property Type Applies to Description
golem.!exp.gap-35.v1.inf.gpu.memory.bandwidth.gib integer Offer Optional The theoretical maximum amount of data that the bus can handle per second
golem.!exp.gap-35.v1.inf.gpu.memory.total.gib integer Offer Indicates the amount of memory available to the GPU

Rationale

This specification is based on the wish-list provided in #157.
For now the assumption is that we only detect the first pci (possibly nvidia) gpu.

Other PCI (and similarly non-PCI) devices can be added as another sub-tree next to golem.!exp.gap-35.v1.inf.gpu.

Backwards Compatibility

If gpu node is not present in the offer then it means the Provider does not have it or does not allow using it.
If gpu node is present but the Requestor does not demand it then it should be skipped in the agreement.

Test Cases

TBD

[Optional] Reference Implementation

TBD

Security Considerations

TBD

Copyright

Copyright and related rights waived via CC0.