In this article, I roll up my sleeves and do a competitive analysis of Intel's (INTC) upcoming "Merrifield" platform.
First off, let me define some terms.
- Merrifield - This is the name of Intel's next generation smartphone PLATFORM.
- Tangier - This is the name of the SYSTEM-ON-CHIP that powers the Merrifield platform.
Now, Intel hasn't announced the official specifications for the Tangier SOC, but the company gave the following slide at its investor meeting that allows me to do a meaningful competitive analysis:
The slide tells us that Merrifield will offer 1.7x the performance of Clover Trail+ in CPU performance and 2x the graphics performance. At first glance, this seems impressive - until one realizes just how slow Clover Trail+ is compared to Qualcomm's high end parts. To illustrate, here is a comparison of the top-bin Clover Trail+ with the Qualcomm Snapdragon 800 found in the Google Nexus 5 in a CPU test:
It is clear from these results that on a per-core basis, the Krait 400 cores found in the Snapdragon 800 are - on an individual basis - about 60% faster than the Saltwell (this is the 5-year old, first generation Atom on 32nm). Of course, Qualcomm packs four of these cores onto the same die, so it extends its lead over the antiquated dual core/four threaded Atom in multithreaded benchmarks where it offers 113% greater performance.
So, let's take Intel's claims at face value (since the "Silvermont" cores are a known quantity and do jive with the claims made) and assume the 1.7x performance improvement. This would lead to the following hypothetical results:
|Single Thread Score||Multithread Score|
Doing a quick "sanity check", this tends to roughly jive with the Geekbench results for Intel's "Bay Trail" system-on-chip for tablets based on the same processor core. I do expect that the single threaded results estimated here are still a bit TOO LOW based on the Bay Trail results (especially if Intel boosts the top clock speed of the Merrifield compared to Bay Trail), but we are indeed in the right ballpark.
So, as far as peak CPU performance goes, Merrifield will lag the top end Snapdragon 800 in multithreaded applications. However, while this may look rather embarrassing for Intel's chip (particularly as it has a significant process advantage), it is important that I now make more widely known a very seldom talked about "problem" with Qualcomm's Snapdragon 800: throttling.
Subtlety Required To Interpret The Benchmarks
Qualcomm and the other ARM (ARMH) SoC vendors appear to be overstating performance/watt that they are able to achieve at the TSMC (TSM) 28nm HPM node. While these firms often advertise high clock speeds (2.26GHz in the Snapdragon 800's case) and lots of cores, and while in very short benchmarks like Geekbench for mobile devices that level of performance is sustainable (which makes benchmarks look good), these performance levels are unsustainable in real world, fairly long workloads.
This is issue is illustrated in Anandtech's review of the Google Nexus 5 (powered by a Snapdragon 800) in which Anand Shimpi showed the clock speeds of the Snapdragon 800's CPU during a long workload under two scenarios: the first, in a normal, room temperature environment, and the second by running the device on a cold compress. The results are rather eye-opening:
Under the cold compress, the CPU (in red) and GPU (in blue) are running at full tilt. But watch what happens when the device is run at room temperature:
All of a sudden this quad core "2.26GHz" processor is reduced to a sub-1.2GHz processor in many cases. This is the problem with the multi-core race: yes, Qualcomm can brag that its chip gets the highest CPU scores in <2 minute micro-benchmarks, but a dual core Intel "Merrifield" built on a much more power efficient 22nm FinFET process is likely to sustain much higher peak performance. Indeed, investors/OEMs will also likely see that the quad core variant of Merrifield known as "Moorefield" will - again, thanks largely to a meaningful process advantage - probably have similar peak performance as a Snapdragon 800 but will be able to actually sustain more of its peak performance over a long period of time
This is simply physics: the better your transistors, the less power you consume for a given function.
What About Graphics?
As far as graphics goes, Intel is claiming 2x over the Imagination Technologies (OTCPK:IGNMF) SGX 544MP2 found in Clover Trail+. I loathe to use a single benchmark as the be-all, end-all, but 3DMark Ice Storm is currently the best proxy for GPU performance on a mobile device today and it makes life easy for investors as Futuremark (the creators of the benchmark) keep a fairly comprehensive database of scores for various devices.
In this particular comparison, I will use the Dell Venue 8 (not Venue 8 Pro) which uses the top-bin of Clover Trail+ (the Atom Z2580) to try to get a performance estimate for Merrifield's GPU. In the comparison I will also include a Snapdragon 800 (the one used in the Samsung Galaxy Round). Here are the results (the GPU score, mind you, not the overall 3D Mark score) for the S800 and Clover Trail+ along with an extrapolated Merrifield score:
|3D Mark "Unlimited" Graphics Score|
|Dell Venue 8 (Atom Z2580 w/ SGX 544MP2 @ 533Mhz)||7898|
|Samsung Galaxy Round (Snapdragon 800 w/ Adreno 330 GPU)||19430|
As far as peak GPU performance is concerned, it seems that the Snapdragon 800 with Adreno 330 will be about 23% faster than the GPU inside of the Merrifield platform.
What GPU Will Merrifield Use?
A question that is of interest to me (and likely to others) is what GPU will the Merrifield platform sport? In my conversations with Intel's architects at the Intel Developer Forum, I learned that the platform would be sticking to a GPU block from Imagination Technologies (Series 6), but as Anandtech shows in this handy chart, the Series 6 GPU comes in many different flavors:
It may seem "easy" to look at this projected performance and try to guess which configuration Intel went and used (it can be ball-parked by using the iPhone 5s performance level as reference point), but Intel's process advantage allows it to clock a given IP block higher than what can be done on, say, a TSMC 28nm HPM or Samsung 28nm HKMG process. Even so, I think we can figure it out.
So, let's go back to an article I wrote that had some iPhone 5s performance numbers that I measured with an actual device:
This means that a "Rogue" G6430 @ 450MHz scores about 19% higher than our hypothetical Merrifield. This means that Intel is probably not using the G6430 block. Further, since Intel's process lead (in transistor performance) allows it to clock its GPU block meaningfully higher than what can be done on a lesser process, I suspect that Intel - in the interests of cutting costs - went with a smaller GPU block and jacked up the clock speed to compensate.
This leads me to my guess that Intel isn't using the G6430 or even the G6400 (area-optimized version of the G6430), but instead the G6230.
The Integration Dilemma
Frankly, Merrifield looks like it'll offer rather superb performance for sub-5" phones (where a Snapdragon 800 would have a tough time playing), but the big problem here is that this platform lacks the integration necessary to really be cost/board area effective in the mid-range of the market (I'm not sure whom other than Apple can get away with a 4" flagship device these days). Indeed, take a look at the following depiction of the Snapdragon 800:
Notice how the Snapdragon 800 - at least in its most highly integrated SKU (there are SKUs without some of these other IP blocks) - comes with integrated Wi-Fi, LTE-Advanced, and bluetooth? The Tangier SOC will not have an integrated cellular baseband nor will it come with integrated connectivity (Wi-Fi, Bluetooth, GPS, etc.). This is perfectly acceptable for the high end of the market where device vendors often use discrete components when they are better than what comes integrated onto the chip (performance/quality matters more so than cost here), but for the mid-range where Merrifield is likely to play, cost is much more important.
If Merrifield had an integrated LTE modem and connectivity, it would be a winner from a technical perspective. Not quite the speed demon that Snapdragon 800 is, but it would be a lower power chip suitable for more svelte, smaller-screen phones. Unfortunately, it has neither, but Qualcomm - with its complete array of smartphone system-on-chip products - does offer parts for this market with all of that integration (albeit at lower performance levels than what Merrifield will likely provide).
The Intel Merrifield platform will be launched in Q1 2014, probably around Mobile World Congress. It is likely, however, that actual devices may come a bit later than that (MWC is held in February) as the platform is likely to be paired with Intel's XMM 7260 LTE-A modem in handsets rather than the rather uncompetitive XMM 7160 LTE modem available today (although this is a prediction that I'm not 100% confident about - we may very well see Merrifield handsets initially paired with XMM 7160).
But the real question is what sorts of design wins Intel will actually garner with Merrifield. Frankly, I wouldn't hold my breath on any big ticket ones - Qualcomm's product offering is still too strong for Merrifield to really rip away sockets there. Things get much better with Moorefield (quad core version of Merrifield with a faster GPU - probably a G6430 or even the G6630 GPU block from Imagination), but that's a second half of 2014 event.
Really, the problem here is that if Intel can't afford a "second best" type of chip. This market is BRUTAL and Qualcomm is rather entrenched precisely because it has technology leadership. Further, in this market, marketing is important. How does Intel really expect to sell a dual core part into a market in which "quad core" - from the lowest end Motorola G to the highest end Galaxy Round - is such an important buzzword? While device vendors do care about performance/watt, marketing - especially in a crowded environment - is even more important.
No flagship phone other than the iPhone would be caught dead with a "dual core" processor, no matter how great those cores are.
While Intel is closing the gap and finally has a more competitive pipeline of products, it's tough to not feel incredible disappointment from both an observer's standpoint as well as a shareholder's standpoint. At the 22nm node, Intel could have done something truly a cut above what its competitors are doing at 28 nanometer if it had started on its designs and had the right SOC development methodology in place earlier on.
Unfortunately, while the 22nm parts are much more competitive relative to the competition than what the 32nm parts had available at the time, it's still not enough. Until Intel has decisively leadership products for the handset space, its market share will continue to be negligible.
Moorefield will be the company's next shot at bat, although I strongly suspect that it won't be until Broxton that Intel can finally win the designs that truly matter and finally drive meaningful revenue growth in the high end of the handset market.
Additional disclosure: I may go long IGNMF.PK at any time.