When you purchase through links on our site, we may earn an affiliate commission.Heres how it works.
An industry landscape in which component supply and demand were in lockstep feels like a lifetime ago.
But as this tide started to recede, another was rising in the form of AI.
AMD har lyckats läcka sina egna kommande processorer.
But what if those businesses were to, instead, consider using thebest CPUsinstead?
They’re more than just a viable alternative, according to Alex Babin, CEO of Zero Systems.
Can you develop your thoughts and explain how you came to this conclusion?
For instance, prediction engines are algorithmic and not driven by transformer architecture.
When deploying very large models, the ideal situation is to fit the model on a single GPU.
This is the best option with respect to performance as it eliminates the overhead of communication between GPU devices.
For some models, it is simply impossible to fit them on a single GPU due to model size.
What sort of CPU are we talking about?
x86, Arm, Risc-V?
Pure CPU vs APU/CPU + FPGA?
For other CPU types to become as pervasive as x86 it will take some time.
There are several drivers for this.
Second, new approaches are being developed to enable training with simplified LLMs, such as Llama-2.
Third, CPU providers are laser-focused on providing alternatives to GPUs with accelerators as well as memory and networking.
D-Matrix, Cerebras).
Will the CPU with its legacy architecture still have a chance against them?
However, many of the new players do not have the developer community enabled and this will delay adoption.
What can the current crop of CPUs do to compete better with GPUs at inference and training?
They need to focus on end-to-end developer experience within the existing compute footprint.
Existing providers need to engage with the ML community and have platforms that work within the existing GPU environment.
