Qualcomm: Krait, Speed Demon

When a presentation slide claimed that the dual-core MSM8960 would deliver 5x the CPU performance of the 1GHz Snapdragon, everyone assumed that mostly meant significantly higher performance per clock than Scorpion. But no, it actually runs at 2.5GHz! I also learned that it would definitively have slightly higher performance per clock than Scorpion so it's actually more than 5x. It still won't be anywhere near as fast per clock as the Cortex-A15, but unlike the ST-Ericsson A9600 it will be available in a quad-core variant (the dual-core MSM8960 is sampling in Q2 2011, the single-core and quad-core MSM8930 and APQ8064 are sampling in early 2012). 

The APQ8064's aggregate CPU performance should be class-leading in that timeframe (even more so than NVIDIA's 28nm Wayne, which is presumably still a quad-core Cortex-A9 at 2GHz or less) but unfortunately the details are sparse both in terms of CPU architecture (Out-of-Order Execution or only small incremental changes?) and all the other components (however we've heard some very impressive things about the GPU and memory bandwidth off the record, the 15x Adreno 200 GPU performance estimate may turn out to be quite conservative). It can also both decode and encode 1080p 3D video, although it's possible it can't do 1080p60 3D decode like the A9600 or 1440p 2D decode like Kal-El (not that either of those modes are anywhere near as important). Interestingly, it even includes PCI Express interfaces (and multiple USB ports, hopefully at least one USB3). Overall, it's a strong ultra-high-end platform.

TI, Samsung and Broadcom

Unfortunately, TI wasn't showing anything about OMAP5 publicly, and I didn't get a private meeting with them. However OMAP4 is easily one of the fastest solutions of this generation (e.g. Imaginations' SGX544MP is still very solid and could certainly be faster than Rogue in the A9600 with sufficiently more cores, although it wouldn't support as many APIs and TI probably won't go that far), whereas OMAP5 looks to be very slightly lower-end than both the A9600 and APQ8064. But in the grand scheme of things that's not really important, and it could still be incredibly successful if they execute well and are sufficiently aggressive on pricing. Believe it or not, there is a lot more to the handheld market than ultra-high-end smartphones, and even there OMAP5 is extremely competitive. It should be little surprise, however, if their competitors manage to grab most of the headlines in the coming months by virtue of their shorter time-to-market (NVIDIA) or their slightly greater performance/innovation (Qualcomm/ST-Ericsson). Maybe a few more OMAP4 design win announcements would help...

The other application processors vendors were even more low-key at MWC (e.g. Hospitality Suites in Hall 4 for customer meetings), but that doesn't mean they're out of the game. Samsung's Exynos (formerly known as Orion) was obviously very prominent through the Samsung Galaxy S2, and that processor business can probably be profitable through their in-house phone business alone, even if they won't necessarily be the main supplier for Samsung Electronics on any given generation - the two parts of the company are fundamentally very separate and they have to work hard for the business. Samsung also multi-sources just about everything. For example, at least one S2 variant will actually use NVIDIA's Tegra 2. They are also one of the three lead licensees for the Cortex-A15 (along with TI and ST-E).

It is very important to understand that Samsung's application processor division is in no way responsible for Apple's latest A4 SoC (although they were probably mostly responsible for the SoC in the original iPhone) as Apple has its own design division and uses Samsung only as a foundry, exactly like Qualcomm or NVIDIA use TSMC as their foundry. Any analyst report that implies the contrary should immediately be forwarded to the recycling center. Samsung's processor division must make its own way in the world, and unfortunately they are not in a good position to convince other OEMs to use their solutions. The proposed joint-venture with Infineon's wireless business made a lot of sense for them as a way to diversify their customer base - unfortunately it didn't for Infineon (who wanted cash) so that didn't happen. While Samsung's technology is solid, it remains to be seen whether they will find a way to attract other customers, and whether it makes sense for the overall company to maintain their current level of investment if they do not.

Unlike TI and Samsung, Broadcom has historically focused on low-cost basebands rather than high-performance application processors. They've got high volume 2G wins at Nokia and 2G/3G wins at Samsung, and now they're trying to both diversify their customer base and enter the smartphone market. They also make high-end discrete multimedia coprocessors (used in e.g. the Nokia N8 for 720p & the original iPod with Video for VGA), but that business will not survive going forward (it's absurd they even bothered with the 40nm 1080p BCM2763) so they are now using the same multimedia IP (based on an exotic vector architecture developed in the UK by Alphamosaic which also handles 3D graphics with a bunch of fixed-function HW around it) in their new high-end application processors (BCM11311 for tablets with 2x1GHz Cortex-A9) and basebands (BCM28150 with 21Mbps HSPA+, 2x1.1GHz Cortex-A9, 2x32-bit LPDDR2).

The BCM11311 looks decent enough but is not very differentiated (it's hard to imagine why a tablet manufacturer would use it rather than a more mature solution from any of many competitors), and while the BCM28150 for smartphones seems nice on paper, the decision to use a dual-channel (2x32-bit) LPDDR2 memory controller seems dubious (there's probably a small performance advantage, but its main competitor is likely to be the Qualcomm MSM8x60 which uses only single-channel LPDDR2, so that gives it a cost disadvantage to start with). The 3D GPU architecture is also a potential problem - even if they get away with their in-house solution (they did on the Nokia N8 after all), it doesn't make much sense to invest so many resources into a dead-end as they will have to license a 3rd Party solution eventually to match the ever-increasing 3D performance of competitors. And as always, the devil is in the details (as are most potential advantages).

Broadcom's real strength here is still low-cost platforms and analogue integration, not high-end diversified solutions, and I think it's quite likely that they will become extremely successful in low-end 3G smartphones but never nearly as much in tablets or high-end smartphones (excluding WiFi/Bluetooth/GPS connectivity of course, where they already are the largest supplier today). I'll examine the potential of that market in my smartphone market predictions on the next page, but for now let's focus on a few more companies...

Renesas, Marvell and Freescale

Renesas Mobile is the result of the merger of Renesas, Hitachi, NEC, and Nokia's baseband group. They have good market share in Asia (especially Japan but not only there). There's nothing fundamental that prevents them from winning more business worldwide and they're willing to invest quite a bit of money to make it happen. Their new ultra-high-end platform looks very good on paper, both for the baseband (LTE/3G/2G from the Nokia team with a 65nm RF chip) and for the application processor (APE5R on 45nm with dual-core 1.2GHz A9s, SGX543MP2, 1080p60 2D/1080p30 3D encode & decode). It remains to be seen how competitive they are on cost (e.g. fully fixed-function baseband architecture with little reuse between standards) and whether they can move fast enough to deliver 28nm solutions and lower-end derivatives in a similar timeframe to some of their competitors. In theory they have a real fighting chance, but in practice it remains to be seen whether they offer enough differentiation to penetrate new customers who already have a diversified supplier base.

Marvell and Freescale's cases are more problematic. They have had little success in the last generation (Freescale i.MX5x and Armada 610 are practically inexistent in the market) although they still have some valuable strongholds (RIMM and China Mobile's OPhones for Marvell - Automotive, eReaders, and various other markets for Freescale). Both of their next-generation 40nm platforms are also big improvements: for example, Marvell's Armada 628 is the first solution with heterogeneous multiprocessing (2x1.5GHz PJ4s plus a lower-leakage 624MHz PJ4) and it supports 1080p 3D encode/decode (and multi-stream) along with a Vivante GC2000 GPU (which is probably quite powerful given the GC8xx's public performance results). But it's not obvious that there is enough room in the market for everyone to be profitable long-term. Given their obvious Tier 2 status, Marvell and especially Freescale are the most at risk of being marginalised.

Then again, a few years ago every analyst seemed to think all application processors would have to be integrated with the baseband by now. That's looking very silly in retrospect given what was announced at MWC11.