6W Intel Atom S1200 Server SoC

[octoberasian]

http://www.engadget.com/2012/12/11/intel-atom-s1200-server-data-centers/#continued

I see this as a direct response to ARM wanting to move into the server market.

The processor specs are as follows:

• 1.6GHz to 2.0 GHz speed
• Dual core, Hyperthreading enabled
• Maximum support of 8GB DDR3 RAM; ECC supported
• 8-lanes of PCI-E 2.0 support
• I/O interfaces from Intel chipsets supported on the processor
• Intel VT enabled
• 6W TDP (lowest available)
• $54 per processor when bought in bulk of 1000 processors

May actually make for a decent file server, media server, or pfsense router if the processor is available on consumer-sold Atom-based server boards.

 

[defaultluser]

So does it have QuickPath, or are OEMs just expected to work around that pathetic P4 bus when building mass clusters?

 

[RamonGTP]

Did anyone else think the "S" in the subject line was a $ sign and thought "WTF!?"

 

[octoberasian]

So does it have QuickPath, or are OEMs just expected to work around that pathetic P4 bus when building mass clusters?

That is a good question because the article doesn't say, neither does Intel.

 

[pxc]

I just looked through the Atom S1200 Volume 1 data sheet and there's not QPI on these chips. There is no external "FSB" for these SoCs either. It's one S12x0 chip = 1 node. I guess if you need to scale these out more, the 8 PCIe 2.0 lanes could offer a medium speed communication interconnect via custom fabric. Cloud servers don't even need that, even when jamming as many SoCs on one PCB as possible. It's unlikely that many S1200 series CPUs will be tied into one huge n-core computer; instead high density systems will just be a large number of dual core nodes.

Amusingly, the Atom cores on the chip still use a FSB. Atom is finally getting an overhaul in 2013, and the 22nm chips will support oooe and hopefully finally update the bus protocol since some models are getting 4 cores on a chip.

 

[pelo]

Amusingly, the Atom cores on the chip still use a FSB. Atom is finally getting an overhaul in 2013, and the 22nm chips will support oooe and hopefully finally update the bus protocol since some models are getting 4 cores on a chip.

Yea, these are still largely based on that old Bonneville architecture. The new Atoms due out late next year should be a pretty big bump in performance. Biggest bump in performance for a CPU in years, I reckon

 

[defaultluser]

One other thing I'd like to see in the 22nm Atom redesign is improved Double Precision performance.

http://vanshardware.com/tag/performance/

That's a huge falloff in multiply performance when working with doubles compared to most processors. Intel obviously spent time optimizing the SP performance, but the literal cliff you see with doubles is hard to fathom. Only the Add comes close to a sane performance drop.

 

[pxc]

One other thing I'd like to see in the 22nm Atom redesign is improved Double Precision performance.

http://vanshardware.com/tag/performance/



http://i.imgur.com/i4NiF.png[/I MG][/QUOTE]lol, good old van.

You're looking at many problems, not just fpu performance. Code optimized to "hide" latency doesn't work the same way on in-order execution processors, plus reduced memory bandwidth on stream-type benchmarks shows (atom n450 has a single channel memory interface). And aside from synthetic benchmarks and few real-world apps, x87 is largely dead. SSE3 has mostly supplanted x87. Lastly, a 2010 era version of gcc is hardly a benchmark standard for commercial software and games. TBH, that article has close to zero relevance for most users.

 

[defaultluser]

lol, good old van.

You're looking at many problems, not just fpu performance. Code optimized to "hide" latency doesn't work the same way on in-order execution processors, plus reduced memory bandwidth on stream-type benchmarks shows (atom n450 has a single channel memory interface). And aside from synthetic benchmarks and few real-world apps, x87 is largely dead. SSE3 has mostly supplanted x87. Lastly, a 2010 era version of gcc is hardly a benchmark standard for commercial software and games. TBH, that article has close to zero relevance for most users.

Nevermind, I thought he was using SSE2 (and thus the fetch bandwidth between single and double would be similar). It would be stupid to run x87 on a processor with one very wide FPU.

Well good then, the ooo processing should fix Atom's one remaining weakness (poor single-threaded performance), so I am excited for the next release.