Sparc T4 chips: Oracle makes stealthy spec change
Hmm <beard stroke> - what could this mean?
Oracle is putting its money where its mouth is on the single-threaded performance of the future Sparc T4 processors. Or more precisely, it is putting Sparc customers' money where its mouth is.
The Sparc T4 processors went into beta with "select customers" back in July and are due sometime before the end of the year. These new eight-core 64-thread Sparc chips, arguably the most important ones to come out of Sun Microsystems or Oracle in a decade, will very likely be announced at Oracle's OpenWorld 2011 customer and partner event in San Francisco next month. Shipments will no doubt follow soon after with the kind of gradual rollout that we saw when the 16-core Sparc T3 chips were launched in September 2010.
Oracle gave out a bunch of details on the impending Sparc T4 processors at the Hot Chips 23 conference in Stanford University in August. The eight-core chips are based on a new core design called the S3, formerly known as the "VT" core when it was designed by Sun Microsystems a few years back, and are expected to run between 2.5GHz and 3GHz.
While the three generations of many-cored many-threaded T series of Sparc chips have been good with workloads that like lots of threads and don't necessarily care about relatively low clock speeds, like database processing or Java application crunching, the prior Sparc T1, T2, and T3 chips did not do very well on jobs where single-threaded performance was important.
The prior Sparc T processors had hard-coded threading mechanisms and algorithms, but the Sparc T4s have dynamic threading, enabled through something Oracle calls the critical thread API, that allows a high-priority application to grab one thread on a core and hog all of its resources, thus significantly boosting that single-thread performance. Each S3 core in the Sparc T4 processor has a dual-issue 16-stage integer pipeline that supports out-of-order execution and delivers around five times the performance of the S2 core on the SPECint2006 benchmark. Tests carried out by Oracle show that the eight-core Sparc T4 processor has nearly the same performance on the TPC-C online benchmark test as the 16-core Sparc T3, which suggests that the clock speeds on the T4 chip will be roughly double that of the T3 chip.
So what does this have to do with software pricing at Oracle?
Well, Oracle charges per-core for its database software, and adjusts the pricing based on what it calls a core scaling factor. When the Sparc T3 was announced last year, the T3s were given a core scaling factor of 0.25, which means you add up the number of cores, multiply by 0.25, round up, and that is the number of licences you need to buy for Oracle 10g or 11g.
That scaling factor on an IBM Power7 server is 1.0 per core, and on the Sparc64-VII-based Sparc Enterprise M machines it is 0.75 per core. (Earlier this year, on the new Sparc64-VII+ chips, Oracle set the scaling factor to 0.5 while at the same time bumping up the new Itanium 9300s to 1.0, cutting its own pricing and boosting Hewlett-Packard's.) When you do the sums, the licensing costs for midrange machines of roughly equivalent performance have the same Oracle database pricing – well, at list price, anyway – if you compare the software price on a four-socket Sparc T3 machine to an IBM Power 750 or Oracle/Fujitsu Sparc Enterprise M5000 with the same million-transactions-per-minute performance.
Late last week, Oracle quietly updated that core scaling factor table: with the forthcoming Sparc T4 chips, Oracle is doubling the core scaling factor to 0.5.
In one way of looking at it, the database performance on OLTP workloads will be essentially the same at the server socket level, whether you choose a Sparc T3 or T4 chip. In another way of looking at it, Oracle believes in the T4 chip enough not to have to keep the core scaling factor at 0.25, a move it made with the T3 after trying to use a 0.5 scaling factor on the Sparc T2+ chip that preceded it. Oracle had to cut software prices to sweeten Sparc T3 deals last year. And this year, Oracle is holding pat, at least on a per-socket basis, to keep the Sparc T servers competitive at a system level. It will be interesting to see how aggressive Oracle is with the hardware pricing on the T4 machines. ®
Re: What Sun give'th, Oracle Take'th away
Oh because I'm sure reading open source processor instruction sets and open sourced material is to riveting, right? I'm sure OpenSPARC.net and SPARC.org were just slammed with ever increasing page hits because they were so compelling like LOTR - T1 - The fellowship of the threads, T2 - The two Towers...errr....I mean threads times four. T4 - The Return of Single Threaded Computing. I personally prefer all of Intel's, AMD's, POWER's, Itanium's open source processor documentation right before some Oliver Twist before I go to bed myself.
The Niagara chips are not built for threaded performance, but they are built for massive throughput, as everybody knows. Niagara servers have higher throughput than any other architecture. If you are not building a high throughput application, then Niagara is not for you.
I suggest you try to load the Niagara massively with lots of throughput and you will see it can cope with more load than any other cpu. If you have only average load, then any other cpu will be faster.
spec change != price hike
Let me see ... where exactly in the article was a spec change mentioned ?
Well, I guess if you're considering Oracle's pricing a feature, you may have a point. Go buy more Oracle stock.