Oracle Dusts Off SPARC, Announces a Big SPARC CPU Development

[btarunr]

Oracle, which inherited the SPARC CPU machine architecture from its big Sun Microsystems acquisition from way back in 2010, made its biggest SPARC-related announcement ever since. The company is developing a new multi-core SPARC processor, codenamed "M7." This chip, according to Oracle, will not only be in the same league as today's contemporary enterprise CPU architectures, but will also feature Oracle's on-chip optimizations, such as special on-chip firmware, that improves performance and security of applications.

A large, on-chip micro-code would work to ensure security of in-memory databases (the chip would feature an integrated memory controller), and hardware-accelerated compression/decompression of very large databases for near real-time analytics. The M7 is the first SPARC-based processor designed from the ground-up by Oracle, and it will go on sale later this week, as part of the company's new T-series and M-series servers. It will also be offered in upgrade kits for Oracle Supercluster, a high-density machine designed specifically for Oracle.



View at TechPowerUp Main Site

 

[RejZoR]

I like the Liam Neeson thing going on with the photo above

 

[FordGT90Concept]

Doesn't sound like it offers anything over modern x86-64 processors.

 

[Steevo]

We could say that general purpose ARM computing with JIT compilers may be the generic way of the future, but high performance native code will always kick its ass, and the industry has kinda spoken on the X86-64 system, perhaps the next revolution in computing will make this obsolete, but for now X86-64 with all of its baubles is the highest performing hardware and software as a set we have seen in its field.

 

[Easy Rhino]

Doesn't sound like it offers anything over modern x86-64 processors.

This isn't going to replace x86-64. This is going to be a chip Oracle puts in their exadata machines and will be better suited for in-memory plsql queries that scale massively.

 

[EarthDog]

This isn't going to replace x86-64. This is going to be a chip Oracle puts in their exadata machines and will be better suited for in-memory plsql queries that scale massively.

we have already heard their salesmen talk about them.

We own.. Exadata(2), exalogic(2), big data, and zero data loss machines already.

 

[t_ski]

...but will also feature Oracle's on-chip optimizations, such as special on-chip firmware, that improves performance and security of applications.

Security?!?! From Sun, the maker of Java?!?!?

 

[R-T-B]

Doesn't sound like it offers anything over modern x86-64 processors.

For data centers, it does. SPARC has had the best barrel processors (basically hyperthreading only times 8) for years. This is only really helpful on highly threaded SQL or java apps however. Guess what Oracle makes?

Security?!?! From Sun, the maker of Java?!?!?

There's nothing inherently insecure about java anymore than theres anything "insecure" about an exe. Java web plugin is the issue and it should've been put to rest along with flash years ago.

 

[FordGT90Concept]

For data centers, it does. SPARC has had the best barrel processors (basically hyperthreading only times 8) for years. This is only really helpful on highly threaded SQL or java apps however. Guess what Oracle makes.

That's what 22-core Xeons and Intel NVMe cards are for.

 

[Aquinus]

That's what 22-core Xeons and Intel NVMe cards are for.

The SPARC T5 is an 8 core CPU that can do 8 threads per core, forget the support for 8 CPUs per server. Intel's CPUs are fast but, for really heavily-threaded workloads, I bet the SPARC T5 is the winner, hands down. Granted, it's tough to compare the two as software has to be specially written to take advantage of so many threads. As @Easy Rhino said earlier, the real benefit here comes from some special server workloads such as optimized in-memory queries as well as being able to accommodate large numbers of queries at once.

The point is that Oracle's intent is to scale horizontally, which is incredibly important if you're going to be working with huge data sets that are being used a lot and often.

 

[FordGT90Concept]

Pretty sure 22 physical cores with SMT is going to beat 8 cores with SMT. I presume SPARC has instructions that are specific to SQL queries (stating the obvious) but I still have serious doubts that it would compensate for almost three times the execution power in a competing Xeon. SPARC may be more attractive though because I bet it is priced significantly cheaper than 22 core Xeons. Oracle can cut a deal on the hardware because they can cash in on the software.

 

[Aquinus]

Pretty sure 22 physical cores with SMT is going to beat 8 cores with SMT. I presume SPARC has instructions that are specific to SQL queries (stating the obvious) but I still have serious doubts that it would compensate for almost three times the execution power in a competing Xeon. SPARC may be more attractive though because I bet it is priced significantly cheaper than 22 core Xeons. Oracle can cut a deal on the hardware because they can cash in on the software.

I would agree if we were talking about CPUs with the same instruction set but, we're not. SPARCs are a very different kind of CPU versus traditional x86 CPUs.

I also stand corrected, the T4 was 8c/64t. The T5 is 16c/128t.

Oracle seems to tout when and how the T5 can haul ass.
https://blogs.oracle.com/JeffV/entry/more_sparc_t5_performance_results
http://www.oracle.com/us/solutions/performance-scalability/sparc-t5-4-tpc-h-record-2065520.html

Now Oracle is clearly going to be biased towards themselves but, even outside Oracle, the T5 is being compared to the E7 series Xeons which we all know pushes the core count envelope by supporting (like the T5,) up to 8 CPUs in a single server.
http://www.enterprisetech.com/2014/02/21/stacking-xeon-e7-v2-chips-competition/

Simple point is that the T5 stacks up pretty well for what it is and that if a "T6" that contains more cores and performance improvements very well could be a powerful option for the server market. Not to say that it would be cheap but, it's not like E7s are cheap either. E5s can't compare because you're limited to 4 CPUs per server max and that's very dependent on which Xeons you have as some can only 2s and others can only do 1s.

Simply put, Oracle is a company that has resources to influence the market. People shouldn't forget that in the software market, Oracle is second only to Microsoft.

 

[FordGT90Concept]

At those links, those are Nehalem/Westmere-based (aka really old). They're actually EOL now and were two years old when the test was conducted (more than a lifetime in terms of hardware).

This is the best Intel has now:
http://ark.intel.com/products/84685/Intel-Xeon-Processor-E7-8890-v3-45M-Cache-2_50-GHz

18-core, 36 threads, 2.5 Ghz, 8-way, $7174.00 a pop.


The third link provides a much better comparison and, unless T6 really jumps ahead by leaps and bounds, it's going to have difficulty stealing business from Intel.

 

[Aquinus]

At those links, those are Nehalem/Westmere-based (aka really old).

The non-oracle link is talking about E7 v2s. As I said, Oracle is biased towards themselves.

Now Oracle is clearly going to be biased towards themselves but, even outside Oracle, the T5 is being compared to the E7 series Xeons which we all know pushes the core count envelope by supporting (like the T5,) up to 8 CPUs in a single server.
http://www.enterprisetech.com/2014/02/21/stacking-xeon-e7-v2-chips-competition/

Edit: Also when the Oracle links were posted, the E7 v2 wasn't out yet. T5 has been around longer.

 

[FordGT90Concept]

SPARC T5 debuted 2013.

You're right that v2 didn't launch until Q1'14. v1 was the standard at the time so SPARC T5 was technically on top of a little while.


The thread title is quite misleading. Two years for a new architecture isn't bad. Actually, it's pretty good.

 

[Aquinus]

SPARC T5 debuted 2013.

You're right that v2 didn't launch until Q1'14. v1 was the standard at the time so SPARC T5 was technically on top of a little while.


The thread title is quite misleading. Two years for a new architecture isn't bad. Actually, it's pretty good.

It's also important to remember that Oracle has the ability to make hardware and software jive nicely together by controlling everything from top to bottom. It's kind of like how Apple restricts its platform and maintains just their lineup. There are big advantages to optimizing hardware for the kinds of workloads you'll be putting on them. In that respect I think Oracle and Sun have done a great job. SPARCs are most definitely not a general purpose CPU as it is incredibly gimped when it comes to floating point performance as there is only a single floating point unit per core despite the huge number of integer execution units but, that's mainly a result of the kind of things databases are doing because most of it is utilizing integer math (think utilizing indexes, hash map joins, etc.)

With respect to the title, it's has been some time since SPARC was revised and given how often new SPARCs used to get released, this was (in my opinion,) overdue. It has a little dust on it, sure, but it's not like we're hauling it out of the attic.

 

[FordGT90Concept]

Wasn't Haswell launched in 2013? Yup. Haswell (22nm) to Skylake (14nm)--not a big jump. T5 was 28nm (TSMC?)...I wonder what T6 will be (16nm?).

 

[Aquinus]

Wasn't Haswell launched in 2013? Yup. Haswell (22nm) to Skylake (14nm)--not a big jump. T5 was 28nm (TSMC?)...I wonder what T6 will be (16nm?).

Haswell-EX chips on the E7 lineup came out 2nd quarter of this year. Just FYI.

 

[FordGT90Concept]

Yeah, LGA2011 is quite a bit behind the consumer products. I hate that.

 

[Easy Rhino]

I presume SPARC has instructions that are specific to SQL queries (stating the obvious) but I still have serious doubts that it would compensate for almost three times the execution power in a competing Xeon.

You would actually be quite surprised how well on chip plsql optimization works due to the nature of plsql compilation and its path of execution. It's like putting SSE instructions on an intel chip. Running SSE instructions was so much faster on the slower chips with SSE than the super fast chips without it.

 

[Waternub]

codenamed "M7."

cant wait for m8....