Spyre
An AI accelerator chip introduced by IBM. Through the torch-spyre project it becomes a PyTorch-native device, supporting device scheduling, memory management and asynchronous execution.
Timeline (last 90 days)
As the compilation layer tends toward unification, the competitive focus of AI chip hardware vendors shifts to chip architecture, interconnect bandwidth, memory bandwidth, and the parts of the software stack that are truly unique (such as Spyre's specific dataflow scheduling).
The IBM Spyre AI accelerator switched its torch.compile backend from the in-house sendnn compilation stack to the open-source community-maintained Inductor.
sendnn is a completely independently built compilation stack from scratch that the IBM Spyre accelerator software stack previously relied on.
In its adaptation for Spyre, IBM did Spyre-specific scheduling and memory hierarchy adaptation on top of Inductor, needing only to map the intermediate representation output by Inductor to the Spyre instruction set.
As a general-purpose backend, in workloads that are extremely latency-sensitive or highly customized at the operator level, Inductor's general passes may be inferior to specialized compilers, and IBM needs to retain a handwritten optimization channel for Spyre's key operators.
The official PyTorch blog published a technical article introducing how IBM's AI accelerator Spyre was made into a native PyTorch device.
After registration is complete, users can directly specify the hardware with a string such as torch.device("spyre"), and PyTorch maps it internally to PrivateUse1.
Spyre implements PyTorch's allocator interface and takes over the allocation and release of tensor memory, which determines whether the memory reuse strategy and device memory management can work in concert with the framework's caching mechanism, and is the foundation of performance.
torch-spyre connects PyTorch's computation graph to Spyre's runtime and firmware, so that operators can be correctly compiled and dispatched to the hardware for execution.
Torch Spyre is a PyTorch backend built by IBM for its Spyre accelerator.
All Facts (10)
As the compilation layer tends toward unification, the competitive focus of AI chip hardware vendors shifts to chip architecture, interconnect bandwidth, memory bandwidth, and the parts of the software stack that are truly unique (such as Spyre's specific dataflow scheduling).
50%UnverifiedThe IBM Spyre AI accelerator switched its torch.compile backend from the in-house sendnn compilation stack to the open-source community-maintained Inductor.
50%Unverifiedsendnn is a completely independently built compilation stack from scratch that the IBM Spyre accelerator software stack previously relied on.
50%UnverifiedIn its adaptation for Spyre, IBM did Spyre-specific scheduling and memory hierarchy adaptation on top of Inductor, needing only to map the intermediate representation output by Inductor to the Spyre instruction set.
50%UnverifiedAs a general-purpose backend, in workloads that are extremely latency-sensitive or highly customized at the operator level, Inductor's general passes may be inferior to specialized compilers, and IBM needs to retain a handwritten optimization channel for Spyre's key operators.
50%UnverifiedAfter registration is complete, users can directly specify the hardware with a string such as torch.device("spyre"), and PyTorch maps it internally to PrivateUse1.
50%UnverifiedSpyre implements PyTorch's allocator interface and takes over the allocation and release of tensor memory, which determines whether the memory reuse strategy and device memory management can work in concert with the framework's caching mechanism, and is the foundation of performance.
50%Unverifiedtorch-spyre connects PyTorch's computation graph to Spyre's runtime and firmware, so that operators can be correctly compiled and dispatched to the hardware for execution.
50%UnverifiedTorch Spyre is a PyTorch backend built by IBM for its Spyre accelerator.
50%UnverifiedThe official PyTorch blog published a technical article introducing how IBM's AI accelerator Spyre was made into a native PyTorch device.
50%