ARM, Atom, PowerPC

Well folks I think you overlook the concept of the MID being basically a smart terminal..with the apps running on the server it connects to. That says MS doesn''t need to port anything, the OS on the MID is very small, the memory/storage are small or non-existent, and that all means that ARM wins. I think we have to move to the old time-sharing model we had on mainframes and get the apps off the cellphone/MID/little device. Only a few minor things need to be running on the device itself. Bandwidth is a lot cheaper than hardware and software...

Does it even make sense to talk about computing not connected to network in years that are too soon to follow? Maybe wiser would be supplyunitless device, SUD. IMHO There is no generic ´MID´ that fits all, distinct segments have their own cpu and data bandwidth requiremenets.

Guys, Unit volume and $$ drive everything - developer base, available software for a market segment = future success. Intel x86 is $$ leader, though ARM ecosystem probably actually drives greater $$ in semiconductor sales and definitely 10x unit volume. Intel can decide to win ARM''s markets, but only if they take a 10x haircut on their unit pricing and a 5x haircut on their dynamic and leakage power, while expanding their offering to single chip SoCs. For most markets it''s not about uniprocessor performance anymore. It''s all about scalable muli-core SoCs that offer ultra low power.

I have a few more comments for those of you who are not already annoyed. :-) (1) I think all of the comments so far are good. (2) It is nice to hear from Ross again. (3) The Atom is compatible with the Core 2, so I can develop my software on a Core 2 Quad running at a zillion GHz, using ALL of the development tools available for IA-32. I also have the Intel compiler, which produces amazing code, including mapping loops over small data to SIMD instructions automatically. (4) On the downside, the Intel tools--the above mentioned compiler and VTune--are expensive (about $800/year each I believe). Intel is going to have to learn to make pennies millions of times instead of dollars thousands of times. Cheers! -Todd

I have a few comments. (1) The large amount of power to decode the IA-32 ISA comes from the early to mid 90s when on-chip caches were small and out-of-order execution unit implementations had large instruction windows (causing lots of switching to decide which instruction to issue--not from decoding instructions). One person who gave a talk at my graduate school during this period reported 30% of total pipeline power being used by decode-issue. The Silverthorne Atom is dual-issue, in-order execute, thus it will use much less power for instruction decode-issue than a 4-way Pentium with 100s of in-flight instructions. (2) I have an EeePC NetBook with a single-core (two threads) Atom running at 1.6 GHz. I can run Visual Studio 2005 on it without much annoyance. It comes out of Windows sleep mode in less than 10 seconds, and boots Ubuntu Linux in about 20 seconds. The Atom processor in it costs about $30 in large quantities and draws about 2.5W. For about $40 and 4W, you can go to a dual-core with two threads per core. With the ability to run all of these applications, I do not see how someone will want to be running on ARM instead. (3) Apple has been advertising for people with both an embedded background and experience with IA-32. Hmmm. Cheers! -Todd

Since ''embedded'' may include everything from a nuclear power station to the tiniest portable device in someone''s pocket, it may be a useful technical classification for 1/2 of the computing space (when is a computer not a computer?) but it probably is not very useful in defining the market segmentation for computing and communications. The discussion here of where ARMs, Intel architecture, and Power architecture processors belong or end up in the ''embedded'' market only makes sense if we realise that this market consists of quite distinct subdivisions that often have very little to do with each other. Mil/aero and networking/comms equipment often share very few characteristics with smart phones, ordinary cellphones and other consumer oriented equipment. With MIDs we are seeing an attempt to make two categories grow together - the smart phone and the laptop. Although we might all have opinions on whether MID devices will take over (and whether growing up from smart phones or growing down from laptops etc. will be the dominant trend in the long term), we should recognise that not all embedded is the same thing.

I often find it odd how embedded is measured. If one measures only by chip volume the clearly the Embedded cpu marketplace is exclusively owned by ARM based CPUs. However if one works in US based space applications embedded is exclusively the domain of PPC750 from BAE. It would be interesting to see a breakdown not only by unit sales, but also by $ sales of SOCs and $sales of final products.

Paul, To some degree, I agree with you, just because of the numbers involved. According to industry numbers, ARM "sells" 10x the number of 32/64 bit processors that Intel (plus AMD) does, and the disparity only seems to be growing with ARM''s push into the deep embedded market. Recent comparisons of ARM vs. Atom show 5x reduced power consumption for equivalent performance, though Intel can crank out faster chips that have even higher power envelopes. There was some story the other day that Intel had gotten Atom processor pricing down to 27$, but requires 3-4 support chips, even though a complete ARM-based SoCs like Snapdragon can be had for the nearly the same price. Once Intel solves their power, pricing, SoC completeness and breadth of offering problems, they might be dangerous.

Paul, I really like your blog entries and your insights into the business of EDA. They all seem spot on to me. However, your foray into processors and processor IP seems to contain several elements I take issue with. 1. The claim that the instruction decoder on the Atom is huge is nothing more than a straw man. While this sort of claim might have been somewhat accurate 20 years ago, instruction decoders are tiny compared to the rest of a processor with today?s designs. Atom?s decoder is almost nothing compared to other parts of the processor, which mostly consists of caches by this point (like all modern processors used as CPUs). I have no special knowledge because I don?t work for Intel, but there?s a die photo of Intel?s Atom processor in the April 7, 2008 article on the Atom by Tom Halfhill (?Intel?s Tiny Atom?) in The Microprocessor Report. The die photo outlines the processor?s main blocks, including the FEC (front-end cluster), which includes the instruction decoder, branch predictor, I cache, prefetch buffers, and instruction TLB. The entire FEC consumes about 10% of the die (that?s a quick eyeball estimate) and the instruction decoder is a small portion of that 10%. The very visible, regular features of the I cache, prefetch buffers, and TLB constitute most of the area devoted to the FEC. Put the instruction decoder down as consuming a mere 1-2% of the Atom?s real estate. 2. ARM may be the big guy on the block with a big 3rd-party applications book in the embedded space, but in the space of PCs and PC-like products including MIDs, Intel is in the upper-class neighborhood and ARM's stronghold is comparatively in the ghetto. We?re talking entirely different weight classes here. 3. PowerPC is nearly irrelevant in the embedded space from what I can tell. It has a big, power-hungry architecture not commonly in demand for new designs. Other RISC cores have passed it by and are far more common in the SOC arena. That's why IBM's pretty much giving PowerPC away through Power.org. Heck, even Apple has abandoned PowerPC. 4. No one needs to emulate an Atom in an FPGA when they can already buy one that runs full speed for $25 and interface that chip to the rest of their breadboard design. What good is an FPGA-emulated processor that runs at 100 MHz and takes up more than half of your $4000 FPGA when you can get one that runs at 10x that speed for 1% of the cost? In fact, FPGA synthesis tools routinely do a relatively poor job of fitting any 32-bit processor core into FPGA fabrics because of the inevitable wire congestion around the centrally located register files. That?s why Xilinx and Altera offer their respective soft processor cores as manually fitted blocks: to get more performance from those processors.