Vijay Tatkar's Blog

Sun is getting noticed in x86 market
This  interesting article in internet news points to a growing interest in Sun's x86 servers. I found it to be a good read overall, but two statements stuck in my mind, so I'm quoting them here:
"I was surprised as hell," Dan Olds, principal of GCG, told InternetNews.com. "The big thing it tells you is that real customers see plenty of differentiation in these boxes, even if the vendors are using the same chips, Ethernet cards and other components."
and finally, at the end, he notes:
"Looking ahead, and perhaps in a nod to the growing importance of virtualization and consolidation, more than half of survey respondents said they expected to purchase fewer, but larger, x86 systems in the future."
Bingo! Sun's sweet spot, I'd assume.
Posted by tatkar ( Oct 04 2007, 01:04:48 AM PDT ) Permalink Comments [0]

AMD and Intel both announce Quad Core Chips
The race is heating up for the x86 server market as Intel and AMD both rush out to get new quad-core chips into vendor systems. Intel announced Tigerton and Caneland platforms for larger than 2 CPU servers. AMD's announcement was for the long-awaited Barcelona chip in upto 8 CPU servers.

AMD's announcement is here. Related Performance data is here (AMD posted, not Sun Studio)
Intel's announcement is here. Related Performance data is here (Intel posted, not Sun Studio).

Should be an interesting period ahead!
Posted by tatkar ( Sep 12 2007, 01:22:48 PM PDT ) Permalink Comments [0]

Sun Studio (Patch) supports Barcelona
With AMD's new quad-core Barcelona just about to be released in new systems I often get asked "Does Sun Studio support Barcelona"?
I want to address this question here. Short answer: yes, of course, we do.
Long answer follows.
We introduced support for AMD/Barcelona in Patch01 (which you can get here) and includes the following things:

• Compiler option: -xtarget=barcelona is designed to set the right code selection, chip, cache and arch values  
• SSE4a intrinsics and assembly code are supported in the compiler and assembler
• an exception to this is the popcount instruction, which is currently implemented in assembler and as an inline template, but not yet exported as an intrinsic
• Quad-core FP support includes more aggressive vectorization (use -xvector=simd option)
• Instruction scheduling has been tuned to use the new (micro)architecture. This is accessible through a compiler internal option and isnt the default yet.
• Exploiting larger L2 cache and Data cache with appropriate instruction selection
• Support for Barcelona Hardware Counters in Sun Studio Performance Analyzer
• Several other specific changes still under experimentation and tuning...

Its too early to indicate here what the performance implication of these changes/improvements are. When the machines come out in the form of announced systems, I'm hoping we can quantify the uptake from this work. The announcements around the chip are indeed causing interesting news buzz around their use in HPC (High Performance Computing). It promises to be an interesting period as AMD looks to take some leadership away from Intel in the HPC market.

Posted by tatkar ( Sep 10 2007, 01:24:27 PM PDT ) Permalink Comments [0]

Sun announces Niagara2 with nice Performance numbers
Sun today announced its latest UltraSPARC T2 processor. Sun is calling it the first "system on a chip", meaning it has a lot of features integrated into it that you would normally associated with a system, like computing, networking, security, and input/output. It will be available in a  single-socket rackmount or  blade server, and is targetted primarily towards TELCO, WiMAX and Network infrastructure deployments
From my perspective, it has some other really nice features that the previous generation Niagara didnt have:

• 1 FP unit per core, 8 FP units per chip
  
• 1 Crypto unit per core, 8 per chip with 10 dedicated functions supported in HW
  
• 8 threads / core (twice of Niagara), 64 threads /chip which gave it the highest SPEC OMP2001 numbers for a single chip
  
• Highest SPECintrate 2006 and SPECfprate 2006 numbers for a single chip easily outranking every other chip
  

How competitive are these numbers? Check out these two blogs from Sun's BMSeer describing this chip and its comparison with Intel/Clovertown, AMD/Opteron and IBM/Power chips: the first blog compares SPECint and SPECfp rates on a single chip, the second blog compares SPEComp numbers.
Its a nice chip, overall, overcoming many of the common deficiencies of the earlier, but still very popular Niagara chip. I hope Sun does very well with it; it does have a lot of potential.
Of course, all reported numbers were generated with Sun Studio 12 compilers.
Required Disclosure:
All SPEC CPU metrics quoted are from full "reportable" runs, but are designated as "estimates" because they used pre-production systems. SPEC, SPECint, SPECfp reg tm of Standard Performance Evaluation Corporation.
Results from www.spec.org as of 8/6/07.
Sun Ultra SPARC T2 @1.4GHz (64 threads, 8 cores 1 chip) 78.3 est. SPECint_rate2006, 62.3 est. SPECfp_rate2006.
Posted by tatkar ( Aug 07 2007, 02:06:59 PM PDT ) Permalink Comments [2]

SPEC CPU2006 released and Sun Leads in Performance again!
SPEC has just released SPEC CPU2006; the follow on to the SPEC CPU2000 benchmark suite.
Sun, once again, leads in 3 out of the 4 major performance categories (see below), which is especially pleasing since this was a first or blind submission where none of the submitting companies knew what others were going to do.
You can find the release and descriptions here.
As SPEC says here, SPEC designed CPU2006 to provide a comparative measure of compute-intensive performance across the widest practical range of hardware using workloads developed from real user applications. Kudos to all SPEC members for getting a more realistic benchmark out into the community . Benchmarking is a constantly changing exercise, as applications emphasis shift.
You can find ALL the submitted results at this URL.
Point of joy for my team and Sun: Sun's submissions came out on top in 3 out of 4 categories: CINT2006, CFP2006 and CINTrate. Check out these results:

1. Best SPEC CPU CINT2006 (SpecINT2006) result out of 23 submissions
2. Best SPEC CPU CFP2006 (SpecFP2006) result out of 21 submissions/a>
3. Best SPEC CPU CINT2006 Rate (SpecINTrate) result out of 12 submissions for overall rate (SPARC, 144 cores as well as 2-core Opterons)
4. Also, the Best SPEC CPU CFP2006 Rate for 2-cores using AMD/Opteron chip but not the best overall, because other configurations were larger!

Required Disclosure Statements:

Sun Fire X4200 13.3 SPECINT2006 (1 core, 1 chip, Solaris 10)
Sun Fire X4200 14.7 SPECFP2006 (2 cores, 2 chips, Solaris 10)
Sun Fire E25K 904 SPECINT2006 Rate (144 cores, 72 chips, Solaris 10)
Sun Fire X4200 26.4 SPECINT2006 Rate (2 cores, 2 chips, Solaris 10)
Sun Fire X4200 25.9 SPECFP2006 Rate (2 cores, 2 chips, Solaris 10)

SPEC, SPECint, SPECfp, SPECint Rate and SPECfp Rate are Registered Trademarks of Standard Performance Evaluation Corporation. Results from www.spec.org as of August 24, 2006.

Posted by tatkar ( Aug 24 2006, 11:54:42 AM PDT ) Permalink Comments [2]