|
|
|
|
|
by jasonwatkinspdx
4613 days ago
|
|
The Cell design was interesting, but it was obvious that chip multiprocessors would win. The argument is similar to why x86 continues to prevail vs RISC and VLIW: supporting x86 is only a modest constant factor in transistor count but preserves access to the existing software ecosystem. The latter trumps the pennies per part the x86 decoders add. Likewise with chip multiprocessors vs systolic arrays: making a core fully general and with proper cache coherant memory access isn't that expensive in transistors, and allows you to leverage existing multi-threaded code. Systolic arrays still have a future in true embeded devices: radar processing is a good example. Oh, and before you point to the AMD architecture as being Cell-like, it's not at all. Unified virtual memory instead of Cell's complex 2 level store. Cell requires the CPU to manage DMA's to the APU's, on Fusion you can just pass a pointer. CPU and APU instruction sets are entirely decoupled on Fusion. APU supports thread pre-emption. In any case, the lesson here is obvious: design hardware in opposition to the existing software ecosystem in your industry, and you'll make something that is academically interesting but struggles to deliver value to end products. |
|
It'll be interesting to see how Intel's x86 SoC development pans out. With the Atom they weren't really aiming for the SoC market and only recently have they actually bothered to try to do so. Even with their first generation stuff they've produced pretty competitive, to ARM, results, given Intel's research muscle and their FAB capacity it wouldn't be surprising to see a lot more x86-SoC based smartphones and tablets hitting the market in the coming years.