Astonishing performance numbers from Freescale

a single core Intel "Yonah" CPU, if measured against either Microsoft's Xenon or the MPC7848would look like a high school kid trying to play in the NFL

Freescale's G4 successor, the e600 based series, is now making its way into the world with the 1.7Ghz MPC7448 available since June of last year and the MPC8641D apparently coming out this spring.

It's the MPC8641D that's most important in the long run because its internal architecture combines I/O, bus, and memory management components to produce a two-way SMP capable CPU board on a chip in much the same way that Sun's Niagara line combines eight cores today. A mid 2005 presentation Inside the New MPC8641D Dual-Core PowerPC® Processor gives an interesting overview and is well worth a careful look.

Freescale, however, has not responded to my request for information on availability and pricing for the MPC8641D - so I don't know what its status is, or when it will be widely available.

From my perspective, however the MPC7448 has two great virtues: first, it's a pilot for the MPC8641D and its 64bit e700 based successors, and secondly it's been out long enough for benchmark results to be available - and they're astonishing.

Since everybody's familiar with Intel's CPUs, it would be nice to be able to compare it directly to something like the Pentium M or Core Duo, but Intel just isn't a player in the embedded processing market and those results aren't there. The closest we get are two AMD results, one on an ancient 550Mhz K6 and one on the 1Ghz Athlon Mobile ("Geode") variant whose design rather heavily influenced both the Pentium M and its "Yonah" successors.

EEMBC, the " Embedded Microprocessor Benchmark Consortium, was formed in 1997 to develop meaningful performance benchmarks for the hardware and software used in embedded systems." Today it publishes benchmark results for about 60 microprocessors including the two AMD machines mentioned above and Freescale's MPC7448 G4 successor.

Here's the benchmark consortium's description of their "DENmark" digital entertainment benchmark:

DENBench Version 1.0 is a suite of [69] benchmarks that allows users to approximate the performance of processor subsystems in multimedia tasks such as image, video, and audio file compression and decompression. Other benchmarks in the suite focus on encryption and decryption algorithms commonly used in digital rights management (DRM) and eCommerce applications. Addressing such end products as PDAs, mobile phones, MP3 players, digital cameras, camcorders, DVD players/recorders, TV set-top boxes, and in-car entertainment systems, DENBench Version 1.0 expands on EEMBC's ConsumerBench Version 1.1 with new benchmark kernels and a larger number of datasets. Several mechanisms, including Peak Signal-to-Noise Ratio (PSNR), check output quality. The DENBench components include the following algorithms and mini-suites:

Note that the 69 tests in this benchmark amount to a hand and glove fit to the Digital Living Network Alliance (DLNA) functional specification for home entertainment devices supposedly to be supported by Intel's "Viiv" Integrated Media Server (IMS) software and the "Yonah" style chips supporting it. Intel, however, is not represented in the results -leaving us to guess where the Core series would fit by extrapolating from AMD's results.

The actual numbers shown are normalised scores - for details click on Calculating the DENmark and other DENbench Consolidated Scores on the on the benchmark page.

The individual results, however, are very specific and correspondingly somewhat safer to interpret and generalise from than is true for most business benchmarks. On the fast fourier transform test, for example, the 1.7Ghz MPC outperformed the 550Mhz AMD K6 by a factor of about 38. Do the arithmetic on a cycle equivelance basis, and it's clear that the Altivec on the G4+ outperformed AMD's SSE/MMX set by a factor of 12 - exactly the ratio claimed for Altivec in Motorola's original documentation.

With that in mind, consider the relative performance suggested by these composite scores:

 1Ghz AMD1.7Ghz MPC7448PPC result/(x86 result x 1.7)
MPEG Decodemark785.1 3524.8 2.64
MPEG Encodemark 587.4 4978.9 4.98
Cryptomark 509.3 3819.6 4.41
Imagemark 918.9 5158.6 3.30
MPEG2 Encode (Floating Point) 30.2 275.3 5.36
DENmark 131.7 762.0 3.40

Note that the overall composite scores suggest that each PPC cycle is roughly equivelent to 3.4 x86 cycles.

As I discussed last week (see The Megahertz Myth and the Ultrasparc T1) Intel's best x86 machines, the big cache Xeon MPs, run about 1.75:1 in terms of throughput equivelance to PPC per cycle, while results produced on dual Pentium 840Ds suggest a cycle equivelance ratio of about 3.5:1 - meaning that the estimated 3.4:1 ratio shown here on the DENmark benchmark is slightly better (presumably because of the SSE/MMX instruction set) than one might otherwise expect.

To the extent that these numbers predict anything, they suggest that a single core Intel "Yonah" CPU, if measured against either Microsoft's Xenon or the MPC7848 would look like a high school kid trying to play in the NFL - and therefore that a key reason Apple could not bring a "MacViiv" Mini replacement to market at January's MacWorld was simply that the CPU, even had it been either affordable or available in the quantities promised, would have made Apple a laughing stock in the home entertainment performance stakes.