IBM confirms PowerPC 750GX ‘Gobi’ spec.
It's all in the architecture
IBM has come clean and detailed the attributes of its latest 32-bit G3-class PowerPC processor, the 750GX. The new chip is the successor to the 750FX, aka 'Sahara', that powers Apple's consumer-oriented iBook notebook.
As revealed by The Register back in May, the 750GX is the chip that will take the G3 to 1GHz and beyond. The chip will launch at 733MHz to 1.1GHz, IBM said yesterday. The 750FX already clocks at up to 900MHz.
Beyond clock speed, the 750GX offers double the on-die L2 cache of its predecessor - 1MB - designed to improve performance further by ensuring data the CPU needs is less likely to have to be pulled in from main memory via the chip's 200MHz 60x frontside bus.
Other improvements are more deeply architectural. The 750GX's L2 cache is also four-way set associative - the 750FX's cache is two-way set associative - essentially breaking it down into four 256KB units. Set associativity is a mapping technique used within the cache to improve data access times. The 750GX's four cache units can be locked to ensure that specific blocks of data are retained.
IBM has also added cache line miss buffers between the L1 data cache and the L2 cache, and the L2 and the bus interface unit (BIU), which controls the flow of data and instructions in and out of the chip. Essentially this provides further fall-back positions if the data the processor needs isn't in a particular cache. The BIU can stack up to five bus transactions - the 750FX's BIU could pipeline up to three transactions - just in case the cache doesn't contain what the processor needs. This is all about readying memory transactions just in case they're needed, speeding things up if they are.
Less relevant to Apple's designers but of interest to embedded applications developers will be the 750GX's 'instruction only' cache mode, in which the L2 is used to store only instructions, not data, which is instead pulled directly from memory each time.
Early speculation suggested the 750GX would be produced using a 100nm (0.10 micron) process. This proved unfounded: like the 750FX, the 750GX will be fabbed at 130nm, again using a copper process with silicon-on-insulator technology. So it's the architectural changes that have brought the 200MHz increase in maximum clock frequency rather than the die-shrink we were anticipating.
The extra cache and circuitry means that the 750GX is bigger than the 750FX: 51.9mm squared compared to 36.6mm squared, boosting the chip's heat output. That said, it's still an impressive less than 8W at 1GHz. The core voltage stays at 1.45V.
The 750GX begins sampling in July, IBM said, followed by full-scale production in December, so don't expect iBooks based on the new chip any time soon. In any case, it's probably about time Apple upgraded its consumer notebook line to the G4 to take allow it to advantage of AltiVec. Then again, IBM's anticipated follow-up to the 750GX, codenamed 'Mojave', is believed to contain an AltiVec unit, and Apple may be targeting that processor for future iBooks. ®