Apple has revealed that IBM's PowerPC 970 chips power its next-generation G5 Macs. But will the new processor(s) provide the speed boost necessary to best Intel and AMD?
Introduction
Even before their June 23rd announcement by Apple (following a widely reported leak the previous week!), it was well known that the next generation of Power Macintosh computers would be based on the PowerPC 970 chip. This IBM-developed chip, which Apple calls the G5, is much more powerful than the chips referred to as G3 and G4. In this article, we'll look at the new chip, its features and performance.
What IBM Said
In early March 2003, a press release from IBM suggested that the Power PC 970 would be clocked up to 2.5 GHz. Mac rumor sites immediately announced that the 2.5 GHz chip was already running in the lab and that Macs based on this part would soon be blowing away Intel's fastest chips. Although the real story turned out a little differently, Apple can still claim "the world's fastest personal computer."
On Mar. 11, 2003, IBM announced that initial shipments of the PPC970, the first in a new family of high-performance 64-bit PowerPC microprocessors, would be delivered at clock speeds up to 1.8 GHz, with volume production slated to begin in July 2003. This seemed consistent with earlier statements and technical information posted on IBM's Microprocessor Forum website in Nov. 2002.
Fortunately, the company was somehow able to push the clock rate up to 2.0 GHz for the dual-CPU model touted by Apple as "the world's fastest personal computer." IBM now says that 2.3 to 2.5 GHz chips will be the generation after the next. In other words, they're not coming anytime soon.
What Apple Says
Apple, of course, says its G5-based Macs are the world's fastest personal computers, and Apple VP Phil Schiller demonstrated the G5's prowess in 3-D rendering, Photoshop imaging and mathematical number-crunching to prove this point, at the product's introduction on June 23rd.
At his announcement of Apple's G5 Macs, Steve Jobs said the new machines were slated for delivery in August. One day later, the online Apple Store listed the earliest available delivery date as Sept. (There's usually at least a few months' gap between the onset of production and the release of commercial products.)
Apple has a developer note highlighting the differences between the G4 and the G5. The "DOs and DON'Ts" section, which highlights some incompatibilities between the two chips, is particularly interesting.
It is also known that the PPC 970's initial production run is based on a .13 micrometer fabrication process and Copper/SOI CMOS technology. Even the 2.5 GHz version of the chip, when it appears, is expected to be based on this .13m fab. The company has also confirmed that the 64-bit CPU has an on-chip 512 KB L2 cache, AltiVec Vector/SIMD unit, and 6.4GB/s I/O system bus throughput. According to Apple, the FSB runs at speeds of up to 1 GHz. More precisely, it runs at one half the CPU speed. In other words, the 1.6 GHz model's FSB is 800 MHz; the 1.8 GHz model's FSB is 900 and the 2GHz Mac's FSB runs at 1GHz.
At this writing, it's unclear how many PPC970 processors IBM can deliver to Apple -- after all, the chip, at the time of the Apple introduction, wasn't even in "volume production" -- that isn't slated to happen until July 2003. Thus, it's unlikely that "mass market" Macs based on the part will be made available anytime soon, and even less likely that Apple will announce faster models before Q1/04. The real risk is that the PPC970, if the entire Mac line does not migrate to it quickly, might dampen sales of all G4-based models. It's also interesting to note that Steve Jobs felt compelled to discuss an unavailable 3 GHz version of the CPU. Conventional wisdom holds that companies should never pre-announce products they can't deliver, as this can impact current sales. (Steve Jobs is old enough to have learned that lesson from Adam Osborne.) However, in this case, Jobs' comment was probably justified: Mac users, jittery after years of near-stagnation from Motorola, need to feel that Apple and its G5 chip partner IBM have a roadmap for future chip development. With a single statement, Jobs unequivocally confirmed this.
What the Experts Say
A May 5, 2003 commentary by Hannibal of Ars Technica on Slashdot suggests that the PowerPC 970 may not be as well-rounded as its most ardent proponents hope. He characterizes the 970’s group dispatching scheme as "bad" and says the way it dictates how one ALU is fed from dispatch slots 0 and 3, while the other is fed from dispatch slots 1 and 2 is a performance killer.
He continues, "Now, this is potentially bad enough already. But when you factor in the fact that the ALUs are not symmetrical, and that certain types of ops can only go to one ALU and hence MUST go into one of only two dispatch slots, then you get a recipe for further choking of dispatch bandwidth."
The end result, he says, "is that the 970’s ALU hardware is weaker than that of the G4e, the P4, and the Athlon. So its clock-for-clock integer performance will be worse, at least this is what I’m predicting. We’ll see if I’m right."
The Media Chime in....
Interestingly, however, it appears that Apple may be fudging some of the numbers a bit to make its point. There is a significant discrepancy between the performance Apple claims its competitors offer, and the claims of Intel and Dell. Indeed, there is evidence to suggest that the 2 GHz dual-CPU Power Mac G5 is in fact slower than a 3.06 GHz single-processor Dell PC. And that's not even Dell's fastest Pentium 4: it also offers a 3.2 GHz model.
Even more damning are the Oct. 2003 tests conducted by PC World, which found that Apple Power Macs did well on Photoshop, but the 64-bit AMD-based systems won handily on most tests -- and ran at least twice as fast in several instances.
Elsewhere, critics charge that Apple's SPECmark test results don't gibe with the official SPEC results obtained by other companies.
Some are accusing Apple of deliberately using the non-optimized GCC compiler on the Intel platform to make its competition look bad in this comparison. The Register casts a skeptical eye at Apple's numbers:
On one hand we have figures that suggest the 2GHz G5 outperforms the 3GHz Xeon in certain benchtests, and on the other we have numbers that show the exact opposite. What gives?
Firstly, Dell's own figures were calculated using different compilers and host operating system: Windows XP Pro, Intel's own C++ and Fortran compilers, and the MicroQuill SmartHeap Library 6.01. Secondly, the compiler used by VeriTest, GCC, is said to generate code that less well optimised for x86. Thirdly, VeriTest seems to have adjusted the test hardware to favour the G5. Again, all the details are there in the documentation.
VeriTest admits it used an Apple-supplied tool to adjust the G5 processor's registers "to enable Memory Read Bypass" and "to enable the maximum of eight hardware prefetch streams and disable software-based pre-fetching". The company also installed a "high performance, single-threaded malloc library... geared for speed rather than memory efficiency". That, says VeriTest, "makes it unsuitable for many uses".
We'd guess these are hardly standard system configurations.
VeriTest also says it tweaked the Dell boxes. For example, when it came to the SPECint and SPECfp rate tests, it disabled HyperThreading, though enabled it for the base SPECint and SPECfp tests. While the compilers were set to optimise code for the Pentium 4, SSE 2 instructions were not used to speed floating point maths operations, only SSE 1 instructions were enabled. VeriTest provides no clear rationale for these choices.
See the following links for further details and Apple's response.
There are also people accusing Apple of ignoring the Athlon MP in its comparisons, knowing full well that the AMD chip would have turned the tables in virtually all of the integer math tests. And it has already been shown that the AMD Opteron, using Intel's compiler, manages to beat the 970 in both SPECint and SPECfp tests. One thing's for sure, though: the new Macs are more than twice as fast as the company's previous high-end G4s in many tasks. And, for many professional Mac users, this is all that really matters.
There are many ways to compare G4 vs. G5 performance. One is the Skidmarks GT benchmarking utility that's part of Apple's CHUD performance tools. Think Secret ran a test providing a rough comparison of G5 and G4 performance. The Skidmarks scale, reports TS, has "100" equal a Power Mac G4 at 1GHz. The Dual 2GHz received scores of 172 for integer performance, 270 for floating point performance, and 208 for vector performance.
In another widely publicized benchmark test, NASA compared a 2 GHz G5 against a 2.66 GHz Pentium 4 in computational fluid dynamics simulation tests. In these tests, which rely heavily upon floating point calculations, the P4 beat any single-CPU G5.
And, in our own tests, a Pentium 4 running at a mere 1.4 GHz beat Apple's 1.6GHz and 1.8 GHz single-CPU G5 configurations in multitrack audio playback and effects processing.
However, as noted in our G5 report elsewhere in this section, these findings don't change the fact that the G5 is by far the fastest computer Apple has ever shipped. And for most Mac users, that's all that matters.
Power Draw
The 1.8 GHz PPC 970, according to IBM, draws approximately 42-48W of power. The 2.5 GHz part based on a .13 micron fabrication process, says IBM, will draw 64W. For comparison, a 3.06 GHz Pentium 4 -- a notoriously power-hungry chip -- draws 81.8W. Things start looking even better for the PPC970 at lower clock speeds. A 1.2 GHz PPC970 is expected to draw only 19W. This bodes well for the prospect of PPC970-based PowerBooks -- probably clocked at lower rates than the desktop models. (PCs, of course, can use lower-power mobile chips such as the Pentium M. IBM already boasts a battery life of up to seven hours for its 5-pound ThinkPad T40 1.7GHz notebook based on the new chip.)
Intel and AMD
Intel and AMD are, of course, hard at work on their own 64-bit families of processors. Even 32-bit chips from Intel and AMD are already faster in many tests than IBM estimates the PPC970 will be. According to information supplied by IBM, Intel Pentium 4 and similar chips (including AMD's Barton) running at 3 GHz or faster are expected to outperform the fastest PPC970 available at introduction.
SPECfp and SPECint are based on real user applications. These benchmarks measure the performance of the processor, memory and compiler on the tested system. They are the most widely used benchmark suites for microprocessor performance.
Based on information supplied by IBM, a PPC 970 running at 1.8 GHz is estimated at:
SPECfp - 1051
SPECint - 937
For comparison, 3.06GHz Pentium4 scores roughly:
SPECfp - 1077
SPECint - 1099
AMD, meanwhile, says its 1.8GHz Opteron 144, intended for uni-processor systems, yields peak SPECint2000 and SPECintFP scores of 1170 and 1219, respectively. Adding a further twist, The Register reported in April 2003 that a senior AMD official hinted at the launch of the Opteron that Apple and AMD are working together on some as-yet unnamed project.
Intel's forthcoming 3.2GHz Prescott P4, announced the same week as the G5 Macs, should boost these scores an additional 10-15 percent, with a 3.6 GHz part, available in the same time frame, providing further gains. AMD's 64-bit Athlon 64 will follow in Sept.
To put it another way, if the PowerPC 74xx/75xx running at 1GHz roughly equals the performance of a 1.6 GHz Pentium 4, a PowerPC 970 at 1GHz equals a 2 GHz Pentium 4.
(More comparative estimates, showing the 1.8GHz PPC970 to roughly compare to an AMD Athlon 2800+ @ 2.25GHz are here.)
More impressive numbers for the PPC970 show up when testing Photoshop, Final Cut Pro and Maya, as the French site MacBidouille reports:
Photoshop: PPC 970 mono 1.4 is 87% faster than a Dual 1.42 GHz
Final Cut Pro: PPC 970 mono 1.4 is 112% faster than a Dual 1.42 GHz
Alias|Wavefront Maya Render: PPC 970 mono 1.4 is 254% faster than a Dual 1.42 GHz
Of course, PC chipmakers aren't standing still. Both Intel and AMD already have processors faster than the fastest PPC970, and both offer SMP-enabled configurations yielding further gains. Intel boasts that, by the end of 2004 -- widely considered the earliest possible date for second-generation 64-bit chips from IBM -- it will have CPUs based on its "Tejas" core running at 5.08 GHz with a 1066 MHz FSB. Tejas will reportedly scale up to 9.20 GHz by the time the 60 nanometre (.06 micron) Nehalem (10+ GHz, 1200 MHz bus) appears in 2005.
Around the same time, IBM's Power5 should become available. Documents we've seen suggest a processor clock range between 1.8 GHz and 3.0 GHz. It promises better error correction and 4x faster performance (!) than the Power4. IBM says it will be available in "both high- and low-end systems." Before you get too excited, however, consider that the Power4 was released in late 2001 and it has taken almost two years for its PPC970 derivative to hit the market. Expect a similar lag between the initial release of the Power5 in late 2004 and its low-end derivative.
Multiprocessing
Prior to the infamous "G5 news leak," few seriously believed that initial shipments of PPC970-equipped Macs would support multiple processors. While Apple has, in recent years, favored symmetrical multiprocessing (SMP)-enabled Macs, it didn't seem coincidental that an IBM press release announcing the "SMP-enabled PPC970" and a second press release announcing a line of PPC970-based blade servers with full symmetrical multi-processing were both rapidly withdrawn after release. In retrospect, it was probably at Apple's request that this information was pulled. For the record, we stated, back in March 2003, that "we believe Apple will probably release versions of the Xserve and desktop (G5?) PowerMacs in dual PPC970 configurations."
As for the prospect of an SMT (simultaneous multithreading)-enabled Mac with a multicore architecture and capabilities similar to those of Intel's HT ("HyperThreaded") processors, it is possible (the PPC970 is, in fact, a single-core variant of the multicore-equipped Power4 family of processors), but not likely. Most industry watchers expect a derivative (probably called the PPC980) of the SMT-enabled Power5 processor to be the path to SMT-enabled Macs in 2005.
Compatibility
The PPC970 is engineered to be backward compatible with existing 32-bit apps designed for current Power PC chips (known as the G4 in Apple parlance). It's unlikely, for cost, power-consumption and availability reasons, that Apple will switch the entire Mac line to PPC970 before 2004. We expect to see the G4 (but probably not the G3, which doesn't support the Altivec instruction set of the G4 and PPC970 series processors), continue to be sold alongside the company's Pro line of 970-based powerhouses by this time next year.
And for anyone who wonders why Apple dubbed the new chip the G5, despite the G3 and G4 families' earlier associations with Motorola, the answer is simple: marketing. It connotes continuity, compatibility and the hierarchy of performance in terms that any Mac buyer can understand.
Conclusion
SPECint (integer) and SPECfp (floating point) tests don't bear much resemblance to real-world tasks with Altivec-enabled applications, where we expect the PPC970 to shine. Although the PPC970 is fully backward compatible with 32-bit code, it seems likely that Apple will recompile OS X and extend certain aspects of the OS (most obviously, in those areas in which larger address spaces are meaningful) to take advantage of the chip's 64-bit addressing capabilities.
"Should Apple move from 32-bit PPC to 64-bit PPC," says Jon Stokes of Ars Technica, "Mac users should not expect the same kinds of performance gains that x86 software sees from the jump to x86-64. 64-bit PPC gives you larger integers and more memory, and that's about it. There are no extra registers, no cleaned up addressing scheme, etc." Elsewhere, igeek.com offers a less PC-centric view of the technical issues with its article entitled "x86-64 : Corrections to Ars-Technica."
As is currently the case, Apple is limited by what chips its CPU partners deliver. We expect that availability of the 2.0GHz version of the G5 will be constrained by IBM's ability to deliver the chips. (None of the products are expected until August.) Until Q3 of 2003, we think the 1.8 GHz model will probably be the "high end" chip in Apple's first batch of PPC970-based Macs. Later, as supply of the faster chips improves, the 2.0 GHz part will offered in a two-way SMP configuration, possibly joined by more affordable, single-CPU 1.4GHz and 1.6GHz models for "mass market" Macs and/or portables. Apple promises that, within 12 months, the machines will hit 3 GHz.
Although the PPC970 should help Apple regain lost ground in the professional markets and will undoubtedly lead to a whole new round of Pentium-beating "bake-offs" at Expo keynotes by Mssrs. Jobs and/or Schiller, the "megahertz myth" is likely to continue to haunt Apple, when Intel, AMD and other PowerPC competitors in 2004 boast high-speed Serial ATA drives, 1000+ MHz front side buses and AGP 8x (or maybe even PCI Express!) or faster architectures.
For Further Reading:
(external links are beyond our control)
Post new comment