MIPS Technologies has unveiled its first parallel processing core. 

"The 1004k is our first multi-processor multi-threading core," MIPS’ v-p marketing Jack Browne told Electronics Weekly.

The processor is similar to the firm’s existing 34k mid-range CPU. "It runs the same binary as the 34k and 24k, there are no new instructions," said Browne. "We have a fairly short pipeline and two threads per core." 

Added for multi-core operation is a coherency management block that ensures requested data is not out of date, and cuts main memory access if data is already with another processor. "It has a cache-sniffing coherency manager to keep the caches fresh," said Browne.

As well as the memory manager, there is also an I/O coherence unit. Between one and four 1004k cores can be specified in a design, running at up to 800MHz.

As such, the possible performance spans that of the the firm’s high-end 1GHz superscalar, out-of-order pipeline 74k core MIPS announced last year.

At 1.95DMips/MHz, a 74k can achieve 2,000DMips at just over 1GHz on 2.5mm sq. of silicon, said Brown. Compared with 2,400DMips from two 800MHz 1004k cores occupying 3.8mm sq.

800MHz is predicted from a core synthesised onto 65nm silicon. "We expect it to move to 45nm, but there are no libraries yet," said Browne.

Well-executed parallel processing tends to save power on a given task, but whether the 1004k can do this is unclear as MIPS is not yet releasing power figures. "We will finish the RTL in April," Browne told EW.

Beta customers have had versions of the core since December and the final RTL will be shipping in June. "We anticipate the first chips in calendar year 2009 and expect in 2010 there will be consumer products based on the 1004k," said Browne.

Applications are foreseen initially in second-generation Blu-ray players which include facilities like picture-in-picture, fast printers, and Internet gateways with advanced security features.

MIPS guide: 74k or multiple 1004ks?

Use 74K core when:

• When staying with an operating system that does not yet have a symmetric multiprocessor version.
• An application is predominantly single-threaded
• Up to 1GHz of performance is sufficient
• The easy or quickest performance migration path is needed

Use 1004K core when:

• New levels of performance are needed
• Multiple tasks are running concurrently
• Scalable performance is required across a suite of products
• The systems has high DMA traffic rates

Why does multi-threading boost performance?

Fast processors can stall while data is exchanged with memory which frequently runs slower than the core.

If thread-switching is incorporated in the core, and the task can sensibly be split into more than one thread, a second thread can be processed while a first is stalled.

"Multi-threading gives 15 per cent improvement because stalls in memory access are not wasted," MIPS’ v-p marketing Jack Browne told EW. "The area taken to support multi-threading is 10 per cent."

The 15 per cent figure comes from processing a single JPEG decode using benchmarks from embedded benchmarking organisation EEMBC.

According to MIPS, in multi-task applications - DVD players and set-top boxes for example - benchmarking suggests multi-threading is likely to give 35-45 per cent more performance.