Intel's (INTC) most recent investor meeting was going well until 2014 financials were forecast, and much of the bullish sentiment that had been building throughout the event turned bearish.
Earlier when Bay Trail was released, I was skeptical of the pricing and performance. I will say that the performance is impressive, as well as the pricing - at least in the tablet models.
In this article, I will take a look at Intel's product stack aimed at the low end, and explain why I feel Intel making a more targeted attack on the low end of the PC market comes with a catch-22, and may end up compounding some of the issues Intel is facing.
Bay Trail Performs Well and Is Priced Aggressively
You may wonder why I begin a paragraph with a positive heading to explain a negative point. In August, before Bay Trail launched, I made the following statement:
"In my best case scenario, it seems to me that Intel will be stretching it to generate an equal revenue per wafer when comparing Bay Trail to Haswell, so I am having a hard time coming to the conclusion that Intel will be able to maintain ~60% GM for their mobile products."
Note that in the assumptions used in the above scenario, I had actually underestimated Bay Trail-T's die size by 25%.
My skepticism was essentially that the performance figures would be inflated at a given TDP, and that I did not feel Intel would be able to price the part, or at least move them in sufficient quantities, to have much of a positive effect on financials.
I will be the first to say that the CPU in Bay Trail is very good, and the chip is very frugal with power consumption. The tablet version is also priced extremely aggressive in order to win sockets. CEO Mr. Brian Krzanich stated he expects to see 40M Intel tablets by the end of 2014.
However, despite this healthy increase in tablet shipments, 2014 forecasts ended up disappointing investors. A 40M rise in chip shipments would represent a ~15% increase over current levels.
Low Margin Business Is The Antithesis of Intel's Business Model
Watch this video. I would not normally be so pointed, but it explains very clearly how technology adoption changes over time to drive the cost down. And it was sponsored by one of Intel's competitors, and is presented by a big wig in Amazon's (AMZN) server division.
Now read this article. I will highlight the following comment from the article regarding Intel's microserver chip:
But Intel competitors are saying that's not true. Both Feldman and Karl Freund, VP of marketing for Calxeda, another SoC vendor using the ARM architecture, say they have not been able to actually buy an Avoton SoC. "You can't buy it," Feldman says. "They will be available beginning of the year." Calxeda has tried to get its hands on the SoC both in the US and in Taiwan and was unable to, Freund says. "It's not available today."
Still, the two concede that Avoton will be available before any competing ARM product for servers. "Avoton will be four-six months ahead," Feldman says about Intel's lead on Seattle. While he has not actually seen one, he does not believe Avoton will be that good of a part. "Avoton is better than the existing Atom - better than Centerton - but not a big-enough step forward at all," he says. "Remember, I'm the only guy with any experience actually building servers with Atom parts." SeaMicro has been - and continues to do so today as part of AMD - building servers using both AMD and Intel chips.
So if Intel has a part that would be perfect for microserver workloads, why is it so hard to find? Completely my opinion, but I believe it is because of the overall effect the part would have on revenues and the bottom line. These microserver parts are designed to lower the TCO for enterprise users. And if businesses are spending less overall, that means the companies supplying the products are likely taking a revenue hit. For every Avoton Intel delivers, it has potentially lost the ability to deliver a higher priced Xeon. But if you are AMD, Calxeda, or any other company selling microservers, your market share is pretty meaningless, so for you, a win is a win.
How Does This Tie Into The PC Market?
Again, Bay Trail-T is an impressive part. CPU performance is great, and although GPU performance lags some of the competition, few of these devices have GPUs that are really good enough to drive a drastically different experience from one device to the next.
But if Bay Trail is too good, then the concerns of cannibalizing Core sales become more relevant. And in my opinion, this concern shows through in Intel's pricing segmentation between Bay Trail-T and Bay Trail parts destined for traditional laptops and desktops.
PART: Bay Trail-T Chip
PART: Bay Trail For Notebook
I've specifically chosen the two examples above for various reasons, but namely because there are benchmarks available on THG of the dual core Celeron (note, it is the desktop version) to allow for a performance comparison so you can verify the claims I am going to make.
The few things I will draw attention to are pricing, design power, core count, and frequency.
Intel is charging 4 times more for half the core count, higher power consumption, and a lower turbo frequency for the company's notebook Celeron chip. Note, I doubt OEMs will pay tray pricing for the notebook Celeron, but who is going to buy a notebook that is more expensive when they can purchase a device such as the Asus T100 and get a quad core convertible for likely the same or lower price?
So when people thump on about Intel's stellar performance in Bay Trail, they're referring to the quad core, aggressively turbo-ed, cheapest part with the lowest power draw that is specifically aimed at tablets. According to Intel's ARK website, the company appears to create pricing segmentation between the tablet and notebook/desktop equivalent versions of Bay Trail, while charging a premium for the top tier Bay Trail M/D chips. This premium prevents the Bay Trail chips from looking too attractive against Core chips.
Vertical Integration Vs. Fabless
Looking specifically at these chips as a function of die size, fellow contributor Ashraf Eassa points out that the Bay Trail die size (pictured above) is 100 mm^2.
According to AnandTech, a quad core Jaguar part from Advanced Micro Devices (AMD) is slightly bigger at 107 mm^2, so Intel's process advantage isn't necessarily giving them a price advantage in respect to die size.
Comparing the cost structure between AMD and Intel, AMD has to sell enough Jaguar chips to recoup the R&D costs associated with developing the chip. They do not have foundry overhead; this gets baked in when AMD buys the chips from TSMC (TSM).
Intel on the other hand has to sell enough chips across its various products to amortize the R&D required for the chips, along with the R&D + CAPEX required to maintain leading edge fabrication facilities. And when these facilities aren't running at capacity, it's more costly and excess capacity charges in the form of write-downs happen. This last point was one of the motivating factors in AMD deciding to spin off the company's fabrication facilities to create GlobalFoundries; to shed that financial responsibility.
Now TSMC, for example, has the obligation to the company's shareholders to make money, as well as keep up with manufacturing technology and spend the required CAPEX in order to remain competitive. And TSMC still has to amortize this cost across the products the company produces, but this is the responsibility of TSMC, not the fabless players.
And although the argument that Intel does not have to pay foundry margins, and is therefore in a better position to compete on price, makes sense superficially, it still warrants a closer look.
Intel's massive R&D budget means the company is the first to ship finFETs, but by the same token means Intel must maintain high gross margins in order to fund R&D+CAPEX. Competing for low margin business at the risk of sacrificing the higher margin business may not be the best move.
To put numbers to paper, I present two BOMs. The first is for the iPhone 5S.
According to Intel's ARK website, the company charges $37 for a Bay Trail-T chip. Apple (AAPL), by designing its own SoC and having Samsung (GM:SSNLF) fabricate it, is able to create a chip that competes well against Intel in regards to CPU performance, bests it in GPU performance, and includes a separate, smaller core to save power by allowing the big cores to power down. And Apple does this all in almost half the price as what it would cost to purchase a Bay Trail chip. If Intel were to sell Apple a chip for $20, would the company make money on that chip? Apple's A7 is also ~100 mm^2.
As a second example, look at the PS4 APU.
A stark contrast to the tiny mobile chips, the PS4's APU is a massive 348 mm^2, according to ChipWorks. Although the GMs are low on this part, AMD is estimated to pull in 15% operating margins. Larger chips are harder to manufacture, and therefore have lower yields than smaller chips.
A 300 mm wafer has ~70,000 mm^2 of usable die space. For ease of demonstration, assume both a standard Jaguar chip or the PS4 APU yields 100%. AMD would get about 200 PS4 APUs per wafer, or 640 Jaguar chips.
Given revenues of $100 per PS4 APU, this would mean AMD generates revenues of about $20K per wafer on PS4 APUs (using my 100% yield assumption).
To generate the same revenues per wafer on standard Jaguar parts, AMD would need to sell those 640 Jaguar parts for $31.
To bring this example closer to reality, Sony (SNE) pays for a very specific chip. Each chip Sony gets has the same number of cores, similar power draw, and operates at the same frequency. Unless Sony is buying lesser chips to release some sort of PS4 lite, chips that do not make the cut are likely trash.
Compare this to AMD's Jaguar line of chips. If a chip cannot operate at a higher frequency, it is sold at a lower frequency. If there is a quad core chip with a broken core, they turn off a couple of cores and sell a dual core part - some broken chips are salvageable in this scenario, unlike with the PS4.
Why Does This Matter?
If the BOM above for the PS4 is close to accurate, AMD is making money by selling 350 mm^2 chips for $100. Jaguar chips are much smaller, and come in different flavors, meaning AMD likely gets more good chips per wafer.
Now unlike the PS4, AMD fronts the engineering bill for Jaguar, so the company will have to charge more for the chips to recoup the development cost to profitably build Jaguars. Typical corporate margins for AMD run around the 40% mark (excluding consoles), and we know that console chips carry lower than corporate average GM. So even if AMD needs to make $30k per wafer in order to cover operating costs, AMD can do this by charging around $50 on average per Jaguar chip.
In Intel's higher margin big core business, the company enjoys a performance lead over AMD at a healthy die size advantage, and while consuming much less energy. AMD's big core APUs, at a die size of 246 mm^2, are much larger than Intel's equivalent chips. Therefore Intel is able to charge a premium over the competition, boosting gross margins.
But AMD's more powerful Kabini processors (highest clocked Jaguar parts) start creeping into mobile Core i3 performance territory. Intel's quad core Silvermont parts have higher CPU performance at better power draw than AMD's low power chips, but struggle to keep up in the GPU arena. However, based on the calculations above, I believe it is likely AMD could still do alright selling Jaguar parts around $50 or so, which is at a much lower price point than Intel's Core line of CPUs.
So connecting all these dots, AMD's most competitive offerings are at the low end. Unlike the high end, AMD offers similar performance (feel free to split hairs here, I'm talking about ballpark figures) to Intel, and based on the pricing of the PS4 APU and AMD's ability to make money off that chip, I believe can do so at a price point well below Intel's Core line of chips.
Yes, Intel could come after the low end of the PC space, but in order to do so, the company will have to do so on both price and performance. A 2 core Bay Trail Celeron at 2.13 GHz is going to have a much harder time competing against a Kabini part. A quad core 1.5 GHz Jaguar APU will likely have slightly better CPU performance. And by slightly, I mean kicks it to the curb and leaves it out in the cold.
Respectively, the benchmarks above are of a 1.5 GHz Kabini against a dual core Bay Trail Celeron part, and a 2.0 GHz Kabini against a Haswell mobile Core i3 part.
In the first benchmark, the Aspire E1-522 offers 50% more single threaded performance when compared to the N2805 Bay Trail, and almost 3x the multi-threaded performance. I did not bother with looking at GPU performance.
The second benchmark shows a Haswell Core i3 ULV part besting a 2.0 GHz Kabini in single threaded performance, but tying it in multi-threaded performance (hyperthreading vs. physical cores).
And you may think I'm being unfair by picking a dual core Bay Trail to pit against a quad core Jaguar, but I am showing pricing and performance segmentation between Intel's Core line and Celeron line of processors.
To explain simply, if Intel wants to go after AMD's market share at the low end, the company will have to do so by releasing higher performing Bay Trail parts at lower prices. AMD can scale Kabini squarely to Core i3 levels, and likely sell these chips at a fraction of the cost as what Intel charges for a Core i3 chip.
But by going after the low end of the segment, Intel drastically increases the risk of cannibalizing the more expensive Core CPUs. And without the more expensive CPUs to offset the CAPEX+R&D spending, Intel's business model starts to look a little shakier.
Intel's PC and data center sales are where the company's bread gets buttered. The video regarding microservers I linked to above laid out the case as to why microservers make sense and lower costs, and if Mr. Feldman's comments are correct, the lack of availability of Avoton is likely a smart move on Intel's part. Intel needs to hold on to high margin business as long as possible.
Similarly, releasing parts to compete with AMD in the low end could obsolete part of Intel's high end offerings, and could easily prove to be an exercise in futility.
Additional disclosure: I actively trade my AMD position, and am long both shares and options. I may add/liquidate shares/contracts at anytime.