August 23, 2006 1:41 PM PDT
Sun's Niagara 2 doubles down with twice the threads
The current UltraSparc T1 "Niagara"-based servers can run 32 threads--eight processing cores that each can run four threads. Niagara 2 still has eight cores, but each can run eight threads, said Greg Grohoski, one of the chip's architects, speaking here at the Hot Chips conference.
"We received first silicon around the end of May, and we booted Solaris in five days. We should have systems to market in the second half of next year," Grohoski said. Though he didn't release any performance statistics, he said Niagara 2 goals include more than doubling the processing throughput and more than doubling the throughput per watt consumed.
The first Niagara was an ambitious design that Sun used to try to restore customers' shaken faith in the company's chip engineering skills. And Sun has had some success, selling $100 million worth of Niagara servers in the second quarter of 2006, only a few months after the radically different design was launched in late 2005.
Niagara is a bet that for lower-end servers, what matters most is how well the server handles many tasks running in parallel rather than how fast it completes an individual task. In contrast, chips from IBM, Intel and Advanced Micro Devices emphasize single-thread speed.
Niagara 2 takes Sun even farther down the multithreading path. Each core has two sections called thread groups that can handle four threads apiece, Grohoski said. Sun's Solaris operating system makes sure the threads are evenly distributed.
Doubling the thread capacity added less than 20 percent to the surface area of the chip, he said; chip size is a key factor in manufacturing expense and therefore in product profitability.
Because Niagara 2 is built on a more advanced manufacturing process that permits 65-nanometer circuitry elements, compared with 90 nanometers for the first Niagara, Sun can pack more features into the chip. In addition to the doubled thread ability, Niagara 2 includes 4MB of cache memory compared to 3MB for Niagara, multiple encryption engines instead of just one, a single PCI Express port, four controllers for FB-DIMM (fully buffered dual inline memory modules) memory, and dual 10-gigabit-per-second Ethernet ports.
Grohoski left unresolved one major question about Niagara 2, however: Will it work in configurations with more than one chip? That would be a significant change from the first Niagara, which only works in uniprocessor machines.
David Yen, who ran Sun's Sparc and Sparc server work until taking over the storage business in May, said in a 2005 interview that Sun plans Niagara 2 servers with more than one chip. With the extra transistors available, he said, "Even more functionality can be incorporated. We started making Niagara to be more scalable. That means now you can put multiple Niagara chips to build a bigger system."
But in his presentation Tuesday, Grohoski said, "This particular design has no multiprocessor capability."
That leaves the door open for multiprocessor-capable Niagara models, a possibility Sun endorsed in a statement Wednesday.
"The current Niagara 2 does not support multichip coherency," Sun said, meaning that the chip lacks mechanisms to make sure it's working well with another--for example, by synchronizing data stored in cache memories. "But we think it's a good idea to modify the coherency mechanism and enable it," Sun said.
It appears that uniprocessor servers hold Sun's initial Niagara 2 focus. "Based on Sun's Hot Chips presentation, it appears the company has concluded that 64 hardware threads will be more than enough for most low-end servers in 2007," Insight 64 analyst Nathan Brookwood said.
Indeed, 64 threads in a system arriving in the second half of 2007 would still be many more than the 16 possible with servers using four AMD or Intel quad-core chips, Brookwood said.
Faster single-thread performance
Although the focus is on aggregate thread performance, Sun is taking steps to improve single-thread performance as well, Grohoski said. For example, each core has a mathematical engine, called a floating point unit, in Niagara 2; its predecessor had only one such unit for all eight cores to share. Each core also has a "stream processing unit" that accelerates a wide variety of encryption and decryption tasks.
"We enable what I call free encryption," Grohoski said, meaning that customers no longer have to pay a performance penalty to encrypt network traffic.
Niagara 2 also includes features to improve server reliability, he said. Hardware can automatically shut off one thread, one thread group, or an entire core if too many errors are detected. "A whole physical core can be offlined, and you still have 56 threads to do work," he said.
A key part of Niagara is doing work while consuming low amounts of electrical power, a major concern for data center operators today coping with ever larger amounts of waste heat. Niagara 2 extended this energy-efficient focus, Grohoski said.
Software can control Niagara 2's power consumption by slowing it down or by shutting parts of the chip down. "Software can turn off threads or cores to manage to a thermal limit," Grohoski said.
4 commentsJoin the conversation! Add your comment