InfiniBand-on-die MIA in Oracle's new 'Sonoma' Sparc S7 processor Oracle's Sparc S7 processor codenamed Sonoma will not feature on-chip InfiniBand interfaces as expected. The CPU, designed for scale-out systems and revealed in detail by The Register in August, was due to sport an integrated InfiniBand controller capable of shoveling 28GBit/s directly between the processor and other nodes and …
The infiniband bits should not really take up much die area in comparison to cache etc. What isn't clear to me is whether the package has "Infiniband pins" that are not connected - hopefully that is the case so they can follow up on their promise to make Infiniband work without forcing a board redesign...
"Therefore, verifying that Sonoma's InfiniBand controller works with all the various InfiniBand adapters on the market was deemed a distraction."
If that were the case I would have expected them to ship it with the pins connected but disabled by firmware so they can verify it *after* the shipping date. On the other hand shipping with the pads disconnected would make sense if they didn't actually implement the interfaces or discovered some basic design flaw...
Interesting option if it's priced as competitively with x86. Solaris has improved greatly since the S10 days and the Linux v. Solaris argument is moot for us. Our Linux guys picked it up quickly. Others may see it differently. We've been playing with in memory and DAX on the M7 v. x86. Large data sets and cube manipulations benefit from DAX to the point we are testing running analytics on OLTP systems directly. Could be interesting if costs come down.
"....Oracle reckons the S7 is on a parity with commodity x86 in terms of price-performance...." Snoreacle massively under-prices SPARC hardware as a means of locking customers into licensing and support contracts for their software, then uses unrealistic benchmarks when comparing performance. SUN used to do the same and we know how that panned out.
The product is called SPARC, not Sparc. And this is the Internet, not the internet. Lazy key-shiftless, can't be bothered with the spacebar, elreg. Learn to type properly. We have the shift key for a reason. Please spend US$10 and get some new keyboards that have this special pair of keys. Thanks!
"The UNIX" (UNIX is a registered trademark of The Open Group) vs "an Unix"! (Actually the latter is verboten, one has to write " A UNIX® operating system")
"The Internet" vs "an internet"!
"The Cloud" vs "a cloud"!
It's enough to make one go brexit!
You missed out the space! It should be UNIX ®, at least according to the Open Group trademark usage guidelines - "The Trade Marks must always be used with white space surrounding them".
See Section 2.1.
UNIX ® is a registered trademark of The Open Group
Really, you want to play grammar police with all the errors in your statement?
To quell the capitalization of 'internet'.
It's capitalized if referring to the name of the world wide web encompassing internet (proper noun).
It's not capitalized if referring to a connection to a group of computers or when using the word in general.
SPARC is clearly still viable as a business proposition; Oracle must still be making money out of the range if they're willing to develop new variants. So, no demise in sight yet.
Wrt to Intel vs ARM; the fatal blow in that fight was struck by ARM a long time ago. That blow was ARM's launch of their licensing model for a cheap and effective low power 32 bit CPU core, and carrying that over to 64 bit designs suitable for use in server applications.
I say fatal; it's not killed Intel yet, but it's going to kill them one day and there's nothing they can do about it. They can survive by building their own Arm CPUs, but then ARM would have won.
More competition for Intel's dominance in the data center is good, looking forward to benchmarks as with that clock speed they should be good. Solaris vs Linux is a valid point, comparing Solaris 11 with RHEL 7 there is a lot to sway us back to Solaris. It does looks like Oracle have done a lot with SPARC and Solaris over over the past few years so planning to give the platform another look, the only worry with Oracle is pricing...
"More competition for Intel's dominance in the data center is good, looking forward to benchmarks as with that clock speed they should be good."
I agree, competition would be good and I really want the left-field stuff to thrive (variety is good !) but I doubt the S7 will be anywhere near Intel chips single-threaded performance. The M7's SPEC_rate figures look competitive against Intel v3s, but Oracle won't publish non-rate figures so you can't compare apples with apples for single threaded performance. Dividing the _rate figures by the # of physical cores on both Xeon v3's & M7 doesn't paint a pretty picture for the M7, so I wouldn't expect the S7 to turn the tables on single-thread performance either.
I want to be proven wrong by Oracle, but they are very reluctant to do so. ;)
FYI. You'll find all the SPARC S7 benchmarks here with lots of comparisons to Xeon. With one quarter of the cores per chip, more integration at chip level and higher clock, the SPARC S7 will provide faster performance/core and per thread than the SPARC M7. https://blogs.oracle.com/BestPerf/
And what I find amazing is the bisection bandwidth of SPARC S7 that outperforms the latest E5-2699 v4 with less than half the cores! https://blogs.oracle.com/BestPerf/entry/20160629_stream_sparc_s7_2
But as usual, Oracle does not bother to publish cpu2006 benchmarks.
Which means, that most likely, S7 is much slower than Xeon.
Oracle still treats benchmarks as just a marketing gimmick.
Until they publish I will advice my customers that performance per core is 50% that of Xeon.
I spec and advice in millions per year in servers and storage, ans depend on industry benchmarks for capacity panning. Cherry pick benchmerks at your own peril.
Taking some liberty with that phrase I think as I know many people who were fired for buying Intel solutions to replace their enterprise solutions from HP, IBM and Oracle (SPARC).
Also, Oracle talks out of both sides of their mouth. For years they have touted performance based on the total sum of all cores in a socket as they always had weak cores but alot of them. Now they cut a S7 by 75% improving thermal dynamics and less plumbing contention that the 32c chip has and now its all about core strength. Interesting given that Intel is now delivering 22 cores (EP) & 24 cores (EX) respectively so having a 8 core chip seems perplexing if they are competing against them.
Things Change. Go back far enough and you'll find people were touting slide rules Hollerith cards. Oracle was touting RAC on x86/Linux because they didn't have an option. The Sun acquisition gave Larry the vehicle to blend s/w and h/w engineering and do some really interesting things. Regardless of what anyone thinks of Oracle, they do build some interesting stuff. I'd put the latest SPARC kit in that category.
You also cannot compare SPARC core count with that of an Intel CPU. Very different beasts. Single threaded computation goes to Intel. Low memory latency, in memory analytics goes to SPARC. We've done the tests. SPARC was so good we thought we screwed up the test.
My point is simply that all this stuff has a place. All SPARC isn't the answer, but nor is all Intel or Power or Itanium (ahem...). Workload matters. The trick is matching workload to platform efficiently.
I see your point but why does this matter? Oracle has already proven with 20+ benchmarks that SPARC M7 is the leader in throughput and in many benchmarks like Database, with the HammerDB benchmark using the TPC-C workload, demonstrated better performance/core than the Power8 in the S824. A majority of those benchmarks also demonstrate up to 50% faster/core than the latest E5-2699 v4. Now they've come out with a lower core count SPARC S7, that by the way, is more cores/CPU than 80% of the Power8's shipping (Power8 12-core only ships in one system, the E880 and remaining systems are predominantly 3-core, 4-core, 5-core, 6-core DCM variants) and clearly its strength is performance/core, price/core and performance/license where majority of software it will be running is licensed per core.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
