Show report – ARM hands DSP extensions to ARM9 processor core

Show report – ARM hands DSP extensions to ARM9 processor coreRichard Ball
ARM has chosen to add DSPextensions to its ARM9 processor core. It is aiming the enhanced processor at hard disk drive chips, audio and speech coders, networking and automotive products.
“Hard disk is a good market because they are on the verge of moving to a single chip solution,” said Guy Larri, ARM’s CPU product manager.
The ARM9E is a full ARM9TDMI with DSPextensions. Thus it shares the Thumb instruction set with its smaller sibling, the ARM7TDMI.
In order to develop the extensions, ARM’s engineers looked at the bottlenecks in the existing architecture.  
  Select your source… Several DSP instructions have been added along with the 32 x 16 multiplier. Source operands for the multiply come from the higher or lower 16-bits of Rm and Rs. Source and destination for the accumulate can be 32 or 64-bits.
“We needed to significantly improve the performance of our multiplier, especially for 16-bit data,” Larri said. “A lot of DSP code represents data as a fraction and lots of algorithms use saturating arithmetic.” These are catered for in the new architecture. The multiplier can access either half of the ARM’s 32-bit data registers through a new set of instructions (see diagram). However, unlike the ARC and Lexra cores, ARM cannot carry out two 16-bit multiplies at once.
But, Larri pointed out that the single cycle 32×16 multiply array is efficient in terms of silicon. “This gives us enough performance for the target applications which are all cost sensitive,” he said.
Larri claims the addition of the DSP extensions will not affect the speed in a given process. The company says ARM9 is capable of 160MHz in a 0.25?m process.
Simulations have shown that implementing a full duplex G.723.1 speech codec uses 25 per cent of a 160MHz ARM9E’s capability. This is half that of the ARM9. Likewise a V.34 soft modem uses 28 per cent of the performance.
The core runs a 1,024 point, radix-4 fast Fourier transform (FFT) around twice as fast as an ARM7 could do it.
Both hard and soft macro of the ARM9E will be available. The physical version occupies 2.7mm2 in a 0.25?m process, said Larri. This is 30 per cent more than the ARM9TDMI.
Power figures are not available, though Larri said the increase when using the DSPinstructions will be ‘modest’.


Leave a Reply

Your email address will not be published. Required fields are marked *

*