Slashdot Mirror
Sneak Peek At Sun's SPARC Server Roadmap
The folks at The Register have gotten their hands on Sun's confidential roadmap from June, which outlines the company's plans for SPARC product lines. The chart has some basic technical details for the UltraSPARC T-series and the SPARC64 line. The long-anticipated "Rock" line is not mentioned. "We can expect a goosed SPARC64-VII+ chip any day now, which will run at 2.88 GHz and which will be a four-core, eight-threaded chip like its 'Jupiter' predecessor. This Jupiter+ chip is implemented in the same 65 nanometer process as the Jupiter chip was, and it is made by Fujitsu, a company that is in the process of outsourcing its chip manufacturing to Taiwan Semiconductor Manufacturing Corp. ... not only has Sun cut back on the threads with [the 2010 UltraSPARC model, codenamed Rainbow Falls], it has also cut back on the socket count, keeping it at the same four sockets used by the T5440 server. And instead of hitting something close to 2 GHz as it should be able to do as it shifts from a 65 nanometer to a 45 nanometer process in the middle of 2010, Sun is only telling customers that it can boost clock speeds to 1.67 GHz with Rainbow Falls."
The only things on Sun's roadmap now are signs to the effect of "Road Closed 1000 feet".
If it's confidential then the Reg shouldn't publish the details. Unless they want to give Sun's competitors a leg-up. I'm sure Sun's competitors marketing teams are happy to have this. [sigh]
Cory Doctorow talking about cloud computing makes as much sense as George W Bush talking about electrical engineering.
I know this is a dark age for literacy, but s/b peek -ya know? like PEEK and POKE???
I'm just sayin'
Sun have been providing theses details to their Partners at the Sun Partner Advantage Summits, I got this info last month.
Plus Sun Partners just have to contact their Sun Sales managers and just ask for a Roadmap Session(Under Signed NDA)
The Register are just publishing what already is pretty common knowledge amongst most people working with Sun/SPARC hardware already, it won't give their competitors a huge advantage at all, the fact that Sun are already revealing this stuff to their wide partner network means that the development of it is well and truly in its final stages, and if their competitors are finding this out through The Register, then they really are not doing their jobs properly.
Google goes for the lowest watt per processing, the actual hardware cost is probably negligible compared to the cost of years of electricity for powering the systems and cooling the surroundings.
Sun had a 486i workstation roadmap, too. They never built a single one. Marketing dreams on a PowerPoint slide doesn't mean squat.
By contrast, though Sun Microsystems often boasted that it has -- actually, had -- the largest microprocessor team after Intel, the team could not design a chip that sold to hundreds of millions of customers. They numbered only several hundred thousand. That sell rate could not pay the salaries of the 2nd largest microprocessor team in the world.
Here is another example of Sun arrogance. In the 1990s, Sun could have easily built their company on the unglamorous ARM RISC processor, but Sun management wanted to exhibit the "pride" (and arrogance) of homegrown technology. So, the management spent billions of dollars on designing one of the worst microprocessors in the industry. The UltraSPARC III was overbudget, was late (to market), and underperformed its peers.
The Sun has set. Good riddance to arrogance.
P. S.
Yes. The UltraSPARC I and II were okay. However, they were not stellar. What helped their sales was mating them to the server box, E10000, that Sun luckily acquired from SGI/Cray. However, starting from the UltraSPARC III (an utter fiasco), the processors were so horribly underperforming that even an outstanding server design could not compensate for the mediocre performance of its processor.
Multi-threading per core helps with video encoding. I saw benchmarks just today at http://www.anandtech.com/weblog/showpost.aspx?i=642 showing the results of the same processors run against the same tasks with and without HT enabled. How many thousand more examples do you need to see?
"It's the height of ridiculousness to say for those 9 lines you get hundreds of millions."
On the first day I worked for Sun, an x86 implemented emulator of a SPARC ran at 2-5 % the speed of the SPARC of the day. Within 4 years, the same emulator on x86 of the day ran at about 50% the speed of SPARC of the day. By now, it may well be 1.2 times SPARC of the day. That is the nature of commodity electronics.
The story isn't SPARC vs x86; it is low-run specialized electronics vs huge commodity production. There is no craftsman-like advantage to the low-run specialty; this isn't furniture, it is etchings on polished sand.
Sun, if this is the best you can do -- 4 cores, 8 threads, arriving at 45nm just as everyone else is getting to 32nm
Sun's performance as a chip vendor is far better that your performance as a Slashdot troll. According to Sun's roadmap, a 16 core times 8 threads processor (128 threads just to be clear) at 40 nanometers arrives in 2010. That would be four sockets per blade, 48 blades per chassis for a respectable 768 multithreaded processors per chassis. As Sun says, it comes down to the TPC-C numbers. I'm no Sun fanboi, far from it, but I could be convinced by the right performance/heat ratio.
Have you got your LWN subscription yet?
But why? 10 years ago I thought sharing an 8 CPU Sun with a big devel team was a privilege. Now any decent Dell workstation has that. What does SPARC have over Intel? (No vague claims of superior "throughput", please!)
It has throughput. Back in 2006, when the first T2000 was released, a dual Xeon could handle 980 req/s from Apache and the T2000 could handle 15,000 req/s:
http://www.stdlib.net/~colmmacc/2006/03/23/niagara-vs-ftpheanetie-showdown/
http://www.stdlib.net/~colmmacc/2006/03/27/niagara-benchmarks-update/
At the same time the Xeon used a peak of 2.2 Amps, while the T2000 peaked at 1.2 A. Things have only gotten faster.
Throw-in on-board crypto, and you can do AES-128 at 38.9 Gb/s with a single socket (eight core) T5220:
http://blogs.sun.com/bmseer/entry/ultra_fast_cryptography_on_the
A T5440 can do 22,932 MB/s (183,456 Mb/s = 179 Gb/s):
http://blogs.sun.com/yenduri/entry/t5440_crypto_performance_numbers
If you're a site that cares about SSL/TLS, how many x86 machines would need to buy, maintain, and cool to handle that load? How many F5 load balancers/SSL accelarators would you purchase? According to F5's own data sheet, the 8900 (with dual 850W P/S) can handle 9.6 Gb/s--and you still have to buy web servers on top of that (more power)
So the T5120 can do roughly four times the raw encryption rate, uses dual 720 W P/S, and also do work as web servers. You're also using less rack space.
Let's also compare to AMD-based systems (which Sun also sells):
http://blogs.sun.com/bmseer/entry/web2_0_consolidation_sun_sparc
Now the Niagara (UltraSPARC-Tx) CPU isn't good for every work load out there, but if it's highly parallel then it's something that you should be looking.
At this point, it looks pretty much the same as a (DEC) Alpha or Itanium roadmap.
Except for one thing: the SPARC circuitry is entirely open source. This has interesting implications, such as the fact that enthusiasts can build these things as FPGAs or even ASICs as fab costs come down to within the reach of clubs and schools. And emerging economies can fab these things by the bazillions without paying royalties. Not to mention big rich countries too.
Have you got your LWN subscription yet?
IBM seems to do ok keeping POWER up with and ahead of Intel/AMD. Sun just rested on their laurels and kept selling essentially the same very expensive product as the rest of the computer world was rapidly advancing around them. The only truly cool thing I saw come out of Sun was the T2 and Oracle killed it for me when they dropped the per core price break when T2 started getting too fast. Perhaps now that they own Sun the pricing will go back to reasonable but I'm certainly not holding my breath.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
> Now the Niagara (UltraSPARC-Tx) CPU isn't good for every work load out there, but if it's highly parallel then it's something that you should be looking.
:). Or Intel could decide to use some transistors to do it - they have lots of transistors to play with on their chips, it's just a matter of priorities.
If Oracle still charges per core, the Niagara approach of many core CPUs could be more expensive.
Looking at the roadmap they seem to be going fewer cores, or at least sticking with 8.
As for power consumption, I wouldn't bet on the Intel x86 always consuming more power than a SPARC for the same performance. They are a scary competitor. They keep introducing consumer grade x86 cpus that are more powerful and yet consume less power.
Can Sun/SPARC keep ahead of them? They might only be ahead in SSL/TLS. And if that becomes a big enough demand, some taiwanese/chinese company start producing cheap pcie cards to do that
IBM seems to do ok keeping POWER up with and ahead of Intel/AMD
Depends on your metric. Raw performance? Sure. Performance per Watt? Maybe. Performance per dollar? Not for most workloads. And don't forget that IBM's latest POWER chips use the same execution engines as the SystemZ CPUs, just with different instruction decoders and a few specialised parts unique to each design; the majority of both chips is the same. When you have customers who think $1m is cheap for a machine, this helps subsidise your workstation processors.
I am TheRaven on Soylent News