Hacker News Reader

Comment by samfisher83 | Hacker News Reader

Lets look at QCT which I think is the biggest if not the biggest ARM company out there.

http://marketrealist.com/2016/04/qualcomms-chipset-margin-hi...

QCT’s operating margin fell from 16.9% in fiscal 1Q16 to 5% in fiscal 2Q16. The margin was on the higher end of the low- to mid-single-digit guidance as the ASP (average selling price) of 3G and 4G handsets equipped with Qualcomm chipsets rose by 6% YoY to $205–$211. The price rose due to a favorable product mix and higher content per device.

The Margins on cell phone chips are terrible. QCT made 2.5 bil on 17bil in revenue.

http://investor.qualcomm.com/secfiling.cfm?filingID=1234452-...

Would it really make sense to invest in Cellphone business when every dollar you put in gets you less ROI compared to what you have now. From a finance perspective it would make more sense to return it to the shareholders and let them invest in QCOM if they want to.

View on HN

Actually 16.9% isn't much below INTC's 18.2% operating margin for 1Q16. The 5% number is more just an outlier due to revenue dropping much faster than expenses - the article mentions losing Samsung and possibly Apple as customers in a pretty short period of time. That certainly does suggest trouble ahead for QCT but I would argue that it doesn't necessarily say much about the industry as a whole.

The reason there overall margin is 18% is due to mobile. Their mobile margin is -30%. The rest of their business is closer to 30%.

https://www.sec.gov/Archives/edgar/data/50863/00011931251306...

As it stands today you can't really say it was a mistake to not get into mobile. They don't have the competitive advantages they had with x86. They had a monopoly there, but in the ARM space they would be just one of many.

The existing ARM fabrication market is very balkanized. Presumably if Intel had dived in with both feet, they would have been able to dominate it using their incredible aptitude for iterating process technology.

There is certainly that third path though, where Intel embraced fabrication, then regretted it.

I'm not sure if Intel's process lead is what it once was. For the last year or so Samsung have been shipping phones that use their own 14nm FinFET process, very similar on paper to Intel's best. It's entirely possible that Samsung will actually beat Intel to 10nm.

A point that's been made is that Samsung's 14nm process isn't really equivalent to Intel's- many gate feature on Samsung's 14nm process are much larger than 14nm.

Still, just the fact that anyone is in a race with Intel at all would have been unthinkable 10 years ago.

On the one hand it's true, but on the other hand, Intel opened it's fabs to others but nobody accepted, so intel's process might be more expensive.

   > The Margins on cell phone chips are terrible. 
   > QCT made 2.5 bil on 17bil in revenue.

And in many ways ARM is to blame for that, if you have a lot of chip suppliers you get downward price pressure. But the more interesting question is this, what is the marginal cost of that extra margin? Intel is going to run its fabs, its going to buy silicon ingots, its going to test and package chips anyway, so the marginal R&D cost is the additional architecture (not the chip layout) and a channel for moving those chips (which is easy since they is only a half dozen or so people who actually make the phones.

Gasse's point, as I understood it, was that Intel couldn't see past it's self imposed margin requirements on chips to see that this business could be additive to its revenue, margin, and keep it in the game when the world switched their primary IT gadget from the laptop to the phone.

"Gasse's point, as I understood it, was that Intel couldn't see past it's self imposed margin requirements"

And I think Gasse is dangerously ignorant to the basic fact that new Apple chips, even with less advanced fab tech, are quite competitive with chips Intel sells for $281. Core M is somewhere in the same field performance wise, and M's are frequently priced the same as the vastly larger ultrabook cores that themselves only offer incremental gains. And the A9X's vast GPU likely applied much of the pressure on Intel beefing up it's Iris, just to stay competitive (with a hat tip to Apple's colossal drivers/software advantage that no one else has). http://ark.intel.com/products/88198/Intel-Core-m3-6Y30-Proce... http://ark.intel.com/products/90617/Intel-Core-i5-6442EQ-Pro...

I see it a bit differently. There isn't any ARM vendor or user which is building an ARM CPU that has the memory and I/O connectivity of even the "low end" x86 chips. Every ARM vendor (with the exception of AMD's server core and HPs use) are focussed on SoC's like the A9x which are high integration chips with limited memory capacity and no I/O expansion beyond what is built in. Great for tablets and phones, and not great for servers/desktops/laptops. Consider that there isn't an ARM chip available you could use a GPU with that was connected via 16 lanes of PCIe.

I think the point was that phones/tablets are a differentiable enough space that you could go there with a lower margin SoC, especially if you went tit for tat and made it an ARM architecture.

According to Google, the A9X has a memory bandwidth of 51.2GB/sec, whereas the Core i5 mentioned above, according to Intel, has a bandwidth of 34.1 GB/s. So what gives?

No generally available access port to that memory. In the x86 world (for some chips) there is a proprietary bus called the "frontside bus" which connects the CPU to something called the north bridge. Other chips export the memory controller directly with some number of "physical" bits which will constrain the total number of DIMMS that can be connected to the system.

ARM SoC's typically have a memory port that supports exactly one memory chip. So the most memory you will see on them is typically 2GB (32 bits x 512K) although from the Anandtech article on the A9x (http://www.anandtech.com/show/9824/more-on-apples-a9x-soc) it seems to support 4GB. Looking at the iFixit teardown (https://www.ifixit.com/Teardown/iPad+Pro+12.9-Inch+Teardown/...) on the iPad Pro it has two LPDDR4 chips (SK Hynix H9HCNNNBTUMLNR-NLH).

So that makes it one of the first I've seen that actually can take two chips. The issues generally are that they burn up a lot of pins and at the speeds the run signal integrity is really hard (for example all of the data and address PCB traces have to be the same length so the bits all arrive at the same time on the CPU side and the memory side!)

> I think Gasse is dangerously ignorant to the basic fact

Unlikely.

I have to agree. not only was it the right answer to not get into the ARM business, it is also the right answer to get out of mobile phone and tablet chips if they are deeply cutting in to your numbers.

The ARM ecosystem and economy run on different margins, and are put together in an entirely different way.

Ax makes sense because Apple is vertically integrated and they get more value out of outspending their rivals on ARM R&D to be in a unique position than they do from whatever effect vertical integration in mobile CPUs has on their margins.

That means Intel is squeezed between a lo-margin ecosystem and strategic vertical integration. Doubly ugly. There is no good answer short of somehow coming up with a significantly better proprietary technology. That somehow appears to be elusive.

I agree. Qualcomm and the rest can own the frontend and get half a nickel in profits per handset shipped, while Intel owns the backend and sells server CPUs by the millions with gross margins of hundreds of dollars per unit and next to nothing in the way of competition.

Having worked for 1.5 of the companies you mentioned I think there is something to be said for and against each side of that. Intel's margins may be higher, but the volume becomes lower and in semiconductor scale is everything. The cost of plants does not scale well and the cost of developing technology nodes is VAST. That is why you see the majority (all?...and I include Samsung in this) of the ARM design shops going the customer fab route.

The issue for Intel is that they seem to have found an unintentional local optimization point where margins and volume are effectively in sync in a way that prevents them from effectively growing into new sectors. I used to joke that every ARM chip you buy includes the purchase of an Intel Xeon in a data center somewhere, but that is now something less than 10:1 and probably more like 100:1.

Now that IBM has divested from their fabrication capacity almost entirely, and they sold it to a company owned by Abu Dhabi, it will be interesting to see what happens with Intel's government revenue share. I could see them working together very intentionally with Amazon and the NSA on projects. I think the purchase and quasi-customer fabbing of Xilinix may be an attempt to go that direction.

Intel has purchased Altera, not Xilinx.

ah cra, my bad. thank you for the correction

Reply on news.ycombinator.com