Advanced Micro Devices Inc (AMD)

[aleph0]

Getting back to parallel computing and that Google thread for a moment...



AMD's Hypertransport/DCArchitecure is probably a good basis for coming close to a parallel architecture "effect" - without leaving X86 ... which is was the subject-matter of that Goofle thread, which talked about extending X86 for extra paralellism.

IMO, the only mistake they made was thinking that Intel should "extend" their hyperthreading model to accomodate this, rather than looking deeper at the problem, which is effectively a problem which "firstly" has to solved at the hardware+OS level to achieve any "significant" or worthwhile effect.
i.e. AMD's Processor Interconnect Architecture would have been a good place to start IMO!

A friend and ex-employee of mine ( his 1+ Maths Doctorate concerned the theory of building a parallel machine ), actually built a parallel machine based on the MC68020 in the late 1980s. The finished machine produced a "constant" throughput which was faster than the "burst-mode" of Cray models of that era.

He designed and built the machine himself (incl. the Mobo), but needed 12 other students to take care of the large software job, i.e. creating/rewriting a high-level application programming language (APL) for the architecture.

The steps were:

1. porting a RT OS (kernal ca. 35K ) to the mainboard
2. writing/creating a "parallel version" partial-prototype of a language called APL
3. 64 mainboards each with 1MB static RAM and it's own harddisk and OS in EPROM
4. 32-bit VME-Bus as a backbone

The OS (on each Mobo) was only used for inter-CPU+Memory communication.
The architecture had effectively NO "own OS" - i.e. the "high-level language" interpreter controlled the whole machine !
In this way, the "high-level" application language (APL) could basically be nicely optimized to take advantage of the parallel architecture without letting an "OS" get in the way.


My point ! ...
<<<
I just dread to think what he might have done had AMD's technologies been available at the time !
>>>

PS:
The project started by trying to use "transputers".
They quickly gave up this idea when they discovered that the command-set of the transputer ( ca. 16 commands ) was so primitive that the overhead incurred trying to do "useful" work far outweighed the "architectural communication" advantages that might have been gained!