In Defense of Discrete Processors

One big surprise is that both the ST-Ericsson Nova A9600 and the Qualcomm APQ8064 are discrete application processors, despite both companies having very strong baseband technology they could easily have integrated in these chips (unlike NVIDIA and TI). ST-Ericsson is also pushing discrete solutions in the high-end smartphone market, not just the ultra-high-end. Why is that?

Qualcomm's explanation is a single word: tablets. OEMs often want to sell WiFi+3G (WCDMA and CDMA), WiFi+3G+LTE (or even WiMax), and WiFi-only versions of the same tablet. For faster time-to-market and lower development costs, they are also likely to use 3G/4G modules rather than design the radio subsystem themselves (which they would be forced to do with an integrated solution). Based on my conversations with both Infineon and Icera, 200K units (for a given configuration) is about the point where it makes financial sense to do it yourself although NVIDIA mentioned the royalties might be slightly lower with modules so it could sometimes be higher than that. And more importantly, 3G/4G modules are prequalified at operators (you can even pick and choose different ones for different operators). This allows products to skip most of the certification process and be available months earlier than a more custom approach.

On the other hand, ST-Ericsson sees the benefits of discrete application processors as being much more widespread, and they plan on remaining discrete in the high-end, not just the ultra-high-end (e.g. the Nova A9540). They believe it makes sense in terms of time to market, performance, and flexibility. Then they'll release derivatives afterwards with an integrated baseband and slightly lower performance for the mainstream market. Obviously, other discrete baseband vendors have very little chance of winning sockets next to these application processors in smartphones (all the reference platforms have the processor pre-integrated with basebands from the same company so it would take a lot more effort by the OEM) but they might have a chance to do so in tablets (where modules are very easy to integrate and a bit of a commodity except for operator requirements).

At a more fundamental level, the cost advantage of integrating the baseband is small and has actually shrunk over time (e.g. it is now possible to reuse system NAND flash and stack the DRAM and/or power management). You're probably leaving $2 to $3 on the table with a two-chip solution (worst-case including DRAM and packaging costs), and a more efficient baseband architecture might actually make a bigger difference than that. The most obvious example is Icera, which I'll cover in a future article. Even otherwise, that's a very small price to pay for a time-to-market and flexibility advantage on products that cost ~100x as much to make overall. As for power consumption, anyone who claims there is a non-negligible difference between the two approaches is a marketing drone - the extra bus is barely a blip on the radar. So clearly the amount of differentiation you can bring on the application processing side is much more important.
Everyone seems to agree that discrete is best in the ultra-high-end, and unless your main differentiator continues to be baseband technology (i.e. Qualcomm), it probably also makes more sense in the high-end both today and going forward. However, integrated solutions continue to dominate the lower tiers, and single-chip solutions (integrated RF & analogue) already own the ultra-low-cost market. So if the opportunity for discrete solutions is mostly in the high-end, how big is that market anyway?

Smartphone Market Predictions

Here's a simple yet bold prediction I'd like to make: more than 95% of 3G phones will be smartphones within ~3 years (early 2014). There's a very simple technical justification to it: 3G basebands are expensive, low-end application processors are not (in terms of silicon die size, 3G basebands are still much bigger than most people realise). And there's also a very simple market justification: most people want either smartphones or voice-centric devices, not a low-quality mix of the two. Feature phones today do not sell because they are desirable; they sell because they're a good value.

The question is not whether the Average Selling Price (ASP) of the *existing* smartphone market will go down. It might slightly, but there's still an unbelievable amount of innovation on the horizon, and smartphones are becoming a more important part of many people's lives. The application processor especially is still a relatively cheap component (e.g. $15 for Tegra 2) in the overall bill of materials, and it will remain a strong differentiator, so its ASP in high-end smartphones might even go up slightly rather than down in the next few years. The *existing* smartphone market is based on large operator subsidies tied to expensive voice/data contracts, and the market is dominated by high-end devices (arguably superphones). That's fine, and this model will probably continue to do very well for a long time.

The point is that another smartphone market is developing at the same time. You could say these are not cheaper smartphones; they are smarter cheap phones. One great example of that in the market today is the ZTE Blade (aka Orange San Francisco in the UK) which sells for about £99 ($160 including tax) and is based on the 65nm Qualcomm MSM7227 (7.2Mbps HSDPA/5.6Mbps HSUPA, dedicated 600MHz ARM11+L2, OpenGL ES 2.0 Adreno 200, WVGA Video) and a 800x480 capacitive touchscreen. The low-end Android market is already going down to $99 and below with cheaper components (e.g. 480x320 resistive touchscreens) and cheaper chips (e.g. 65nm ST-Ericsson U6715 with a shared 468MHz ARM9 or 65nm Broadcom BCM2157 with a dedicated 500MHz ARM11). These solutions have to make many dubious compromises, but some of them are already more attractive than feature phones to many people.

The next generation is when it gets really interesting. The 40nm ST-Ericsson U4500 and Broadcom BCM21654 both have very similar specs (U4500: 7.2Mbps Release 7 HSPA+, dedicated 800MHz 1xCortex-A9, 1xMali-400, WVGA video encode/decode, external RF & Audio/Power chips), and ST-Ericsson claims 2012 smartphones will hit unsubsidised price points below $100 with their solution. There's really no good reason to still buy a 3G feature phone at that point in time, and the price point for that level of performance is only going to drop further as the level of integration increases. For example, Broadcom has indicated that they're working on a 40nm 3G RF block to integrate with their 40nm basebands (and its die size is supposedly half what it was on 65nm - good engineers can apparently make anything scale, even analogue/RF) - that kind of integration is clearly one of Broadcom's historical strengths, and it will certainly be valuable as 3G (smart)phones enter that kind of price point.

The existing smartphone market is unlikely to be cannibalised by these solutions - the target market is completely different and only 3G features phones will be obliterated in the process. The existing 2G market will remain continue to exist for a long time (based on ultra-low-cost single-chip solutions with varying levels of multimedia performance) although it might shrink slightly (in both unit and dollar terms) because of the value in paying just a little bit more for a perfectly decent 3G smartphone (with WiFi so it could make sense even without 3G access for some people). Actually there is one scenario where high-end smartphone sales may be affected: people may prefer to have a cheap smartphone along with a high-end tablet (or a similar device). I don't think it's likely, but if it does happen, then OEMs and semiconductor suppliers are unlikely to complain since there's probably even more money to be made on a tablet.


It was a very impressive year for smartphone application processor announcements, and as evidenced by the analysis above, there are still a lot of different companies investing in the market (and we're not even considering Intel or some of the niche players here). It will be interesting to see whether the market can sustain this many players. For example, will Freescale even bother with a Cortex-A15 license? Could CSR (historically a Bluetooth/connectivity company) try to enter the smartphone/tablets market, and would they have enough technology for it already (via their SiRF and Zoran acquisitions) or might they buy one of the smaller players? More interestingly, what if Broadcom bought Samsung's application processor division and started using Samsung as a foundry for other products as part of the deal?

All of these questions will answer themselves in time. And the next article will open up even more food for thought, so stay tuned! (and yes, there will be some very interesting Handheld 3D/GPU-centric content... eventually, patience! Just be glad 'eventually' isn't part of Valve Time!)