c0t0d0s0.org

Rick Hetherington was in Munich today ... he gave a good presentation into the future of CMT SPARC and what we will see in the near and the middle term. Many people have the opinion, that Sparc will see it´s end soon, especially after the announcement of TI to stop investments in new process technologies. The direct opposite is true, more than ever. I was quite impressed. Unfortunately it´s not up to me to disclose things.

We fastly ramping up UltraSPARC T2. Here you see: 12 E10000k, only a little bit short of a quarter Terabit per second networkbandwith, 0,75 Terabyte per second memory bandwith, 96 cores, 192 integer pipelines, 96 crypro accelerators, 96 floating point units, 768 threads, 768 virtual processor. Imagine this processors in some nifty boxes, with a nice silver finish, 1 or 2 rack units high. Consuming only 1200 Watts for the processors. Imagine the load this babies can handle. 5 years ago, you had to fill a complete datacenter for a similar performance. Now it´s a small tray of CPUs (respective 12 or 24 rack units). Simply mind-boggling. But it´s even getting better: In the near future you will have this sheer power in only three machines.

Ich muss ja ganz offen gestehen, das ich die c´t nicht mehr lese. Die Zeitschrift hat vor vielen Jahren irgendwie den letzten Reiz verloren und in Zeiten der weltumspannenden, drahtgestützten Pornoverteilung gibt es als Nebenprodukt viele weitere Webseiten, die viel schneller Informationen liefern. Einer der wenigen Lichtblicke der c´t waren zuweilen die Kolumne "Prozessorgeflüster" von Andreas Stiller, auch wenn dieser sich selten positiv ueber das Haus Sun geäussert hat. Das scheint mit Von Wasserfällen und Ökowellen endlich ein Ende zu haben. Ich zitiere mal meinen Lieblingssatz:

All das ist jetzt Geschichte, denn nun hat Sun mit dem UltraSPARC T2 (Codename Niagara2) wieder ein ganz heißes selbstgestricktes Eisen im solaren Feuer, das zumindest als Einzelchip in der Multithreaded-Welt alle anderen in die Finsternis schickt - wobei man Intels „geschummelten“ Quad-Core mit zwei Dice in einem Gehäuse ruhig mit hinzurechnen darf.

Die PC Games berichtet über Niagara2 in Schnellste CPU: Nicht von Intel oder AMD:

Die Werte des professionellen SPEC-rate-Benchmarks, welcher auch die Skalierung mit mehreren Kernen berücksichtigt, liefern einen großen Vorsprung gegenüber der Konkurrenz aus den Häusern Intel, AMD und auch IBMs Power6.

Hätte ja nicht gedacht, das mal aus der Richtung Lob kommt.

Lawrence Spracklen gives a detailed overview about the crypto performance of the Niagara II chip. Alexandre Chartre decribes in his blog the usage of Linux on an UltraSPARC T2 processor.

What did we announce yesterday? In my opinion, the most important chip for Sun. It´s even more important than Rock. When you look around, the loads for single threaded performance decrease, and as the complete industry turns in the direction of "more cores" (even IBM talks about this in the Power7 timeframe). Rock will be important for the highest end. But the future of general purpose computing begins right here, right now.

So, why is Niagara 2 so important. Niagara 1 had it´s weak points for general purpose computing. We decided to design a processor with such weaknesses, as we designed it with a certain workload in mind: Internet, a niche according to IBM and HP, but hey ... it´s a really big niche. Niagara 2 was designed as a general purpose CPU. The problem of the single FPU? Addressed ... the N2 has 8 of them? The issue of SSL-centric Crypto-circuits. Addressed ... they were subsituted by full-fledged crypto accelerators. You get 64 threads, you get two 10 Gigabit Ethernet interfaces with chipsets specifically designed for multi core processing, as they implement the processing tasks in a similar multithreaded way. You get eight lanes of PCI Express. 2 Gigabyte/s in and 2 Gigabyte/s out. Solely for storage attach, as you have already two really big fat pipes of Ethernet directly connected to the inner crossbar of the chip. 60 GB/s worth of memory bandwith.

The point of beeing only capable to run in single socket systems will be solved soon by Victoria Falls. Imagine a system of two or four this processors. Running on an operating system really capable to handle such large amounts of computing threads because four processor with 64 threads each means scheduling you processors on 256 hardware processor. Not an easy task.

Now many people will say: Each thread will only run on 1.4 Gigahertz. This is half as fast as a modern x86 CPU. But is this really a factor? Maybe you can do more cycles in a second, but what´s the gain, when you wait most cycles for memory. A T1/T2 swiches simply to a non-memory starved thread. This is one the reason why a 1.4 GHz CPU can be faster than a 4.7 GHz CPU in SPECfp rate and SPECint rate . It´s like my sister and me at playing racing games on her PC. I´ve lost every time in spite of having the faster car. The problem: I´ve lost traction in every curve of the race track and had to reaccelerate while my sister was able to take the curve without no problems. And: 1.4 Ghz is just the beginning.

IBM want´s to tell to plays it´s virtualisation card by saying: "But we have better virtualisation, we can do more than 64 LDOMS". But the decision for limit was a sentient one: When ever you want to switch to a virtual machine, you have to save the actual register sets and restore the stored register contents of the next vm into the register sets. Takes a vast amount of clock cycles. Now: A 64 threads processor has 64 register sets. By limiting the number of domains to the number of threads - i hope you´ve already got the point - you get an important advantage: Switch from a VM to the next in a clock cycle. Neat, isn´t it ?At the end, it´s not important to be able to partition the cpu into smallest fragments. It´s important that the performance isn´t evaporated by the VM layer. Or didn´t you asked your self, why there´s no benchmark result for a virtualized system in the large benchmark portfolio of IBM or why the VMware licence prohibits benchmarks? LDOMs virtualisation is virtualisation done right.

Now, as the N2 hasn´t the problem of the single FPU, that prevented N1 of beeing a general purpose CPU, the game changes again. 18 month after we did it with N1 the first time. With Niagara II you get a processor suitable for almost all tasks of computing. And as Jonathan said: You´ve ain´t seen nothing yet. But you see the future of computing today. And really soon in your datacenter.


PS: Maybe some statements are really bold ones. But when you´ve read, what i´ve read in the past, when you know what i know, when you saw what i saw you would be so enthusiastic as well.

SearchDatacenter cites Nathan Brookwood in Sun Microsystems unveils Niagara UltraSparc T2 chip:

Nathan Brookwood, the founder and principal analyst for research firm Insight 64, a consulting firm based in Saratoga, Calif., said he's impressed with the upgrade.[..]All in all, it's a very impressive chip in terms of its absolute performance and its performance per watt."

And I´m really interested about a discussion at osnews.com. This portal was pro-Sun in the last few years.

Jonathan in Sun Enters the Commodity Silicon Business:

To add fuel to the fire, the blueprints for our UltraSPARC T2 (I personally like the moniker, "Niagara 2" - named after Niagara Falls, btw, and the great volumes of water that pass over them), the core design files and test suites, will be available to the open source community, via its most popular license: the GPL. Making Niagara 2 the only commodity silicon whose core designs are available to the open source community - whose strength, and market power, only grows by the day.

BTW: Now i know why there were several articles about N2, while i bite myself on my tongue all day long:

In the interim, on Monday a reporter accidentally violated our news embargo, which set off a flurry of press coverage.

And Rick Hetherington and Hal Stern state in Innovation@Sun: A System on a Chip – The Sequel:

And then being the leader in power efficiency, we paid special attention to how we can manage power internally, so we do have features like clock-gating, extensive clock-gating throughout the core. So if one is not running a floating point application, the clocks are not being submitted to that floating point unit.

David Yen in an interview with ZDnet:

"At the same time, we will [also] roll out a developer beta program in which we will provide access to Verilog RTL design file and test suites, so that we can get early feedback to refine our open source process and eventually complete the open-sourcing [of UltraSparc T2] by the summer of 2008," he explained.

And even a enterprise systems affine news sites like Engadget write about the announcement.
Timothy Prickett Morgan writes Sun Polishes Up Sparc T2 Multithreaded Chips:

Sun is not saying much about performance when it comes to the T2 chips, but the T2 is expected to have about 2.5 times the throughput performance of the T1, and Victoria Falls is expected to nearly double that again, with a 65 percent performance boost over the T2's at the system level.

The usual suspect at the Register wrote a nice piece about the Niagara II: Sun releases world's fastest chip - at 1.4GHz. Ashlee, you´ve ain´t seen nothing yet. Other benchmarks are even more impressing than SPECint rate or it´s floating point counterpart.

The support for logical domains made it into the 2.6.23 linux kernel. Wow, this was really fast ...

The Register reports (speculates) about the upcoming releases of Niagara II and later Rock:Sun preps 2048-thread monster. It´s hard to keep thinks secret, when you write and open-source the code to support your new gear. I won´t comment the speculations of the new systems. But they will be available on time

The Unix Guardian is in heavy speculation mode regarding UltraSPARC RK. I won´t comment what´s right and what´s not, but read it yourself

You find four good articles how to use logical domains in Octaves Weblog:

• Introduction to LDOM's
• LDOM Installation
• Configuring the Control Domain
• Time to Build a Guest Domain!

There is a slight but extremely important change in the licensing modell of Oracle. The Multicore FAQ states:

Oracle Database Standard Edition can only be licensed on servers that have a maximum capacity of 4 sockets. When using RAC to cluster this limitation is mirrored in the cluster. It may be licensed on a single cluster of servers supporting up to a maximum of four sockets.

What does this means: Our workhorse server V490 is now usable with Standard Edition. So you can save a vast amount of money. Instead of 3 Licenses (each core is counted as 0.75) Enterprise Edition you need only three licenses Standard.

Or you can build an incredible powerful OLTP RAC-Cluster with 4 T2000 servers. Although they have 32 threads there is only one single socket per Server. 128 parallel threads = 4 UltraSPARC T1 = 4 Sockets = within the limitation Standard Edition. Standard Edition = RAC included.

And now the kicker: Niagara 2 equals 64 threads per socket. Means 256 threads per 4-node cluster. But: Still only 4 sockets. And you surely know it: Still Standard Edition.

At OpenSparc .net you will find a in-depth presentation of the upcoming Niagara II processor: "An 8-core, 64-thread, 64-bit, power efficient SPARC SoC (Niagara2)". The presentation was held at the International Solid State Circuits Conference 2007.