Patrick Mochel wrote:
>> Has anyone worked with LinuxBIOS or even seen detailed specs. yet for
>> the Crusoe from Transmeta?
>
> Yes. I started to look into it some time ago, but have been swamped with
> various other things. I have a copy of the BIOS Writer's guide, though
> how much of it can go out is apparently still undefined.
>
>> You'd think info would be easier to get a hold of with Linus working
>> there.
>
>
> We do have several Linux people here, but that doesn't mean that
> _everything_ here is open source ;)
Who do I have to know at Transmeta to get specs? We've requested data
sheets about 10 times. We considered using the Crusoe for much denser
clusters than RLX when Transmeta first began sales but we never heard
even a single response. We'd still like to work on a Crusoe solution.
>
>> RLX finally released their servers based on cards that use a Crusoe and
>> up to two 2.5" hard drives on a single 3U high card (they are calling a
>> Blade??) and can fit 24 cards into a 3U enclosure. Crusoes aren't
>> anything to write home about as far as number crunching performance but
>> they do produce much less heat than a K7. They also don't have much for
>> internetworking but they could be useful for some clustering applications.
>
>
> I'm not about to try and sell these things, but as far as MIPs/Watt and
> MIPs/cu. ft. go, it's attractive. Also, though I'm no clustering expert,
> how much more than 3 10/100 NICs is necessary for a decent cluster?
>
> -pat
>
1,000,000 MIPs in 18 cu. ft. single 19" rack (56K Mips/cu. ft) is what
we're designing now with the XScale if you want fixed point performance.
1-TFLOP in the same space using mobile K-7's and ULV P-III's for
floating point. There are also some cousins to the Mips Emotion Engine
that have even better floating point performance that any of the x86. 3
Fast Ethernet nics per node is nice for flat networks if you have nice
wide and very fast switches.
Bari
