Tilera Taunts Intel With a 100-Core Chip
bits.blogs.nytimes.com
bits.blogs.nytimes.com
(He was interested in running x264 on it.)
The TilePRO64 Processor is programmed in ANSI standard C
You'll need to use our custom proprietary (and bad) compiler, and we probably won't give you an assembler.
32-bit VLIW processors
You'll have to spend thousands of man-hours rewriting all of your assembly code.
H.264 HD encode for 10 streams of 1080p (baseline profile)
Even when we make every last possible marketing cheat that we can possibly make with custom CPU-specific code, we can only get 10 streams going realtime, barely more than what one can do on a standard dual-Nehalem system with x264 (with similarly marketing-cheat level settings).
700mhz, 866mhz operating frequency
Unless every part of your application is embarassingly parallel (read: not x264), there's no way you'll get it to run at any reasonable speed on this thing.
iLib APIs for efficient inter-tile communication
You'll have to rewrite your code yet again to use a new library.
5.6 Mbytes of on-chip cache
Our CPU has an embarassingly small amount of cache per-core, which is why we can get so many onto the die at such lower power usage. Your program will be overwhelmed by cache misses.
19-23W @ 700mhz for typical applications
We're selling a big fancy DSP, not a real general-purpose processor. If you want to get good performance out of it, you'll probably have to rewrite a lot of code.
Although generally the posts there are just rants this one is quite good
- The instruction set used is a MIPS-derived VLIW instruction set according to http://en.wikipedia.org/wiki/TILE64
- It needs to say that 99% of developers don't really care about instruction set as they use higher-level languages.
EDIT: A new cpu architecture is always an exciting sight, though.
Don't tell me. I would _love_ to have a workstation based on one of these. While the hypervisor is a very interesting element, I would be delighted to see a kernel taking advantage of the bare-metal partitioning scheme for managing core-affinity on processes sharing a single memory space.
It's fun to imagine a thread-migration mechanism based on memory-access-hop-counting for dynamic optimization of process/thread/tile allocation.
I saw a presentation somewhere of Erlang scaling up to 64 cores on Tilera, but besides that it's very short on benchmarks.