After buying Sun back in 2010, Oracle continued Sun’s work on the SPARC T processors and now it is getting ready to present the new SPARC T5 processor at this year’s Hot Chips conference.
Japanese company Fujitsu is also preparing to talk about its own SPARC architecture implementation during the same event, as we reported here
Back when Sun was trying hard to recover from the 2000 dot.com bubble burst, the company gave up on high thread performance microprocessor
development and chose to concentrate on high throughput.
The first-generation SPARC T1 processor has 8 cores
able to handle four threads each. It only has a single FPU unit and it was not suited for HPC use, but rather specialized on HTC use.
Sun adopted Fujitsu’s high-performance SPARC64
processors for its HPC servers and the combination was quite a successful one.
Later on, Sun wanted to give the T series more thread performance and thus brought 8 FPUs and 8 threads per core in the T2. The T2 Plus managed to achieve over 65 times the performance of a 2001 SPARC III Sun processor.
Sun’s T3 was a 16-core design with 8 threads per core and the current SPARC T4 dropped the number of cores from 16 to just 8, but it greatly improved the per-thread performance.
The initial T1 was years ahead of its time.
We’re still waiting for applications that would know how to best use a 8-core / 8-thread processor now, in 2012. Back in 2005, a 32-thread CPU was quite difficult to be put to good use.
The T4 has 300% to 500% single thread performance of the previous generations and the new T5 is promising to improve on that.
Oracle’s T5 has already taped out at TSMC
late last year and the company has been working on the chip and servers ever since.
Besides offering considerable single-thread performance improvements, the T5 will be the second Out-of-Order processor in the T series after the T4 and will also integrate special, per-core encrypting units along with database accelerating features.
The T5 is built on 28nm
technology and will feature memory versioning, in memory columnar database acceleration, hardware decompression and other database specific compute units.