Well, to Jensen's point: GPUs haven't been about graphics for a good while now. It's good he is reaminding the audience they're just general purpose accelerators (glorified SIMDs on a PCB of sorts) for embarrasingly parallel tasks. Float point monsters is another good name, I guess? Heh. Even using "GPGPU" is kind of wrong nowadays.
What's ironic about saying this
now is that it's gotten less and less true, over the past decade. If we go back about 10 years ago, you would be spot-on that, aside from video codecs, display engines, ROPs, TMUs, and perhaps Tessellation engines, GPUs were just straight-forward floating point SIMD + SMT engines.
In the meantime, along came Tensor cores, integer arithmetic, and RT cores. Nvidia has even added hardware for performing optical flow analysis! They're still
mostly arrays of SIMD processing engines, but not as much as before (and not only floating point).
let's call them "general purpose accelerators" (GPAs)
That's a little too non-specific. CPUs are the ultimate generalists.
Intel has promoted the following classification:
They define them as follows:
- Scalar architecture typically refers to the type of workloads that are optimal on a CPU, where one stream of instructions operates at a given rate typically driven by CPU clock cycles. From system boot and productivity applications to advanced workloads like cryptography and AI, scalar-based CPUs work across a wide range of topographies with consistent, predictable performance.
- Vector architecture is optimal for workloads, which can be decomposed into vectors of instructions or vectors of data elements. GPUs and VPUs deliver vector-based parallel processing to accelerate graphics rendering for gaming, rich media, analytics and deep learning training and inference. By scaling vector architectures from client, data center, and the edge, we can take parallel processing performance from gigaFLOPS to teraFLOPS, petaFLOPS, and exaFLOPS.
- Matrix architecture derives its name from a common operation typically performed for AI workloads (matrix multiplication). While other architectures can execute matrix multiply code, ASICs have traditionally achieved the highest performance implementing the type of operations typically needed for AI inferencing and training, including matrix multiplication.
- Spatial architecture is a special architecture usually associated with an FPGA. Here, the data flows through the chip, and the computing operation performed on the data element is based on the physical location of the data in the device. The specific data transformation algorithm that has been programmed into the FPGA.
Source: https://www.intel.com/content/www/us/en/silicon-innovations/6-pillars/architecture.html
Using that taxonomy, we might call GPU-like compute devices something like "Vector Compute Accelerators". However, there have long been
Vector Processors in the super computing sector, and I'm sure purists would turn up their nose at trying to lump GPUs in with some of those elegant machines.
Facebook
Twitter
Instagram