What Arm's CEO makes of the Intel debacle
theverge.com
theverge.com
CPUs are still relevant. And x86 can still provide a lot of value and iterate forward. AMD has spent 20 years getting to the front. It is the short term vision that is killing Intel.
AI accelerators are going to be useful and valuable until the point they are commoditised just like CPU. Until then someone (Nvidia) will gain a lot of profit. They used this money to buy their way into the datacenter with companies like mellanox. Nvidia will be on top for a while. And then the cycle continues and we will see a new company on top.
Laptops using AMD CPUs are at their most available and they're still lucky to be 20% of any given manufacturer's SKUs.
Intel has an opportunity to reclaim some of those areas with their more power efficient chips, but both they and AMD are facing more competition from non-x86 manufacturers than ever, and that competition is likely only to grow.
Apple's CPUs seemed to be a product of engineering without concern for backwards compatibility, e.g.: analyze current software, make the CPU do those things well. I wonder what Intel could produce if they came at a new line of processors with the same mindset, though that would require engineers to be in charge of the company again.
What is needed is a sober look and re-engineering.
> Arm is also rumored to be eyeing an expansion into building its own chips and not just licensing its designs
> if you’re a vertically integrated company and the power of your strategy is in the fact that you have a product and you have fabs, inherently, you have a potential huge advantage in terms of cost versus the competition
I don't get this strategy at all. Why is vertical integration of the designs and the fabs any advantage? Sure, there are probably small savings here and there from changing the design to suit the process. But like Intel you'd get fat and lazy and end up with weird tooling that only your one company uses. And you can't concentrate on being the best at one thing with huge volumes like TSMC.
Arm is way overvalued as it is, and now the CEO is going to try something stupid.
Unfortunately I think that the temptation is to give in to the "faster horse that eats less" requests and to accept "we just can't do that boz" on the other end. China seems to be moving to a model that couples integration with bouts of intimidatory violence and so far that seems to be working well, but it does have costs.
The point of specialization and outsourcing is to amortize fixed costs among a larger customer base. If fixed costs were zero, then it would never make sense to outsource, since the effective marginal cost is identical in the non specialized company versus the specialized company. Every company would produce everything they need in house.
Fabs have enormous fixed costs. Any form of vertical integration is going to make your products more expensive, unless you specialize in some kind of underserved niche with no competition. TSMC will always have the lowest marginal cost.
What Arm should be doing is partnering with TSMC and other fabs to build optimized cell libraries for Arm designs and investing into software tools for SoC integrators and become a one stop shop for proprietary IP cores and custom ASIPs so people don't even think of talking with Cadence or Synopsys.
Intel got arrogant, remained arrogant and despite getting absolutely pummelled by competitors on share price value, believes it is special.
Intel needs to realise it's not going to catch up TSMC and so should focus on cannibalising all other competitors and moving into spaces where TSMC doesn't operate. It's going to be a lower margin game from now on but Intel can survive. Oh and be the nicest folks in town, not the shit on the people we need like Gelsinger did.
Intel is in effect a Will Ferrell movie character. Character is arrogant, becomes arrogant and stupid in defiance, and potentially finally sees the error of its ways and grows a bit.
And commit to building a simple, powerful, enthusiast/startup friendly platform for running mini-clouds at home/on-premise. And start bringing tech like Optane, Tofino etc to developers so they can get better price/performance than the soon to be ARM dominated cloud.
https://www.intelcapital.com/investing-in-oxide-unlocking-cl...
And listening to the most recent Oxide podcast [1] I can't think the Oxide team would want to stay long at Intel unless the new CEO took radical action.
[1] https://oxide.computer/podcasts/oxide-and-friends/2218242
Which meant no developers could really develop for it so you're in a vicious circle.
It's why I think they need to have some way for people to easily use their tech at home/on-premise.
I understand the change, eventually there won't be any publications left, if people keep refusing to pay for them, and use ad blockers as well.
I know! it's a crazy idea. it'll never work. but still.
Unclear why "service" is in quotes. Not many people get to be in the room where it happens. The people that do get to be in the room, absolutely do provide a real service to the rest of us who don't.
You answer your own question. There's no reason at this point in history the "room size" needs to be artificially limited. Was the Arm CEO paid for their time? If not, why not?
> absolutely do provide a real service to the rest of us who don't.
It's service for Arm not you. If you wanted unbiased analysis of Intel's position you might seek someone other than a direct competitor. Which makes paywalling this all the more baffling.
Especially since - see another comment thread here on HN - their "AI" thinks you're a regular reader when you click on them once per month.
> Our original reporting, reviews, and features will be behind a dynamic metered paywall — many of you will never hit the paywall, but if you read us a lot, we’ll ask you to pay.
So... pretty sure i don't read verge a lot either. And I get the paywall.
To see how the Intel CEO is thinking, I now scrolled through a few interviews with him.
In this interview, Intel's CEO is asked to describe the difference of a CPU and a GPU:
https://www.youtube.com/watch?v=d07wy5AK72E
His answer is that the CPU is capable of doing general computing, while a GPU is made for very specific tasks.
I'm not sure if that is a good way to put it.
The way I would put it is that a GPU is a bit like an array of many CPUs. And to manage this array of processors all doing computation at the same time brings constraints and management overhead. So it is usually harder to program a GPU but results in better performance.
Personally I feel that the 5 years of reintroducing 14nm instead of getting 10nm out the door was what ate up any lead that Intel had over its competitors.
And don't forget that rumors has it that Apple reported more bugs with Intel Skylake than the QA team at Intel had managed to report. Never a good thing when one of your biggest customers finds more bugs in your product than your own engineers have managed to find.
I wouldn't really blame all this on a single person though.
The largest part of that “overhead” is the need for much higher memory bandwidth and caching due to having many more cores to feed.
And the added complexity of programming is managing groups of cores (and multiple groups of cores) so they are all as active as possible without blocking each other algorithmically or due to memory misalignments, page read/write inconsistency, or cross interference.
With many more cores achieved with smaller cores: streamlined Arithmetic Control Units, less pipelining (?) and a simpler/shared Instruction Control Unit.
Somebody correct me if I am wrong.
If your program has one or two threads and spends all its time doing branchy control flow, a CPU will run it adequately and a GPU very poorly. If the program has millions of mostly independent tasks, a GPU will run it adequately and a CPU very poorly. That's the two limiting cases though and quite a lot of software sits somewhere in the middle.
The most concise distinction to draw between the hardware architectures is what they do about memory latency. We want lots of memory, that means it ends up far from the cores, so you have to do something while you wait for accesses to it. Fast CPUs use branch predictors and deep pipelines to keep the cores busy, fast GPUs keep a queue of coroutines ready to go and pick a different one to step along when waiting for memory. That's roughly why CPUs have a few threads - all the predictor and wind back logic is expensive. It's also why GPUs have many - no predictor or wind back logic, but you need to have a lot of coroutines ready to go to keep the latency hidden.
Beyond that, there's nothing either CPU or GPU can do that the other cannot. They're both finite approximations to Turing machines. There are some apparent distinctions from the hosted OS abstraction where the CPU threads get "syscall" and the GPU threads need to DIY their equivalent but the application doesn't care. Threads on either can call fprintf just fine. It makes a bit of a mess in libc but that's alright (and done in LLVM already).
> If the program has millions of mostly independent tasks, a GPU will run it adequately and a CPU very poorly.
Now I would say:
CPU - optimized to execute threads with independent instruction streams and data use.
GPU - Optimized to execute threads with common instruction streams and data layouts.
—
CPUs
As you noted: Optimizing conditional branches is one reason CPU cores are more complex, larger.
CPUs also handle the special tasks of being overall “host”. I.e. I/O, etc.
—-
GPUs
One instruction stream over many cores greatly reduces per core footprints.
Both sides of conditional code are often taken, by different cores. So branch prediction is also dispensed with.
(All cores in a group step through all the instructions of both if-true and else-false conditional clauses, but each core only executes one branch, and is inactive for the other. Independent per-core execution, without independent code branching. Common branching happens too.)
—
Both CPU and GPU can swap between threads or thread groups, to keep cores busy.
Both optimize layers of faster individual to larger shared memory (i.e. registers, cache levels & RAM).
board was focused on extracting value from the company
A simple way to check if a board/CEO wants to extract a lot of future value _now_ is to look into their accounting, and especially how they count physical assets.
If they count physical assets at their current selling price, you can probably get in long-term.
If they count them at their buying prices you might have to worry.
If they do 'future accounting', tying into assets their estimated future added value, you can be sure the current owners are extracting the most they can _now_ and will try to drop the bag into futures investors (or ask for a government bailout)
And how would counting physical assets at their buying price move money from the company into the pocket of the CEO?
In normal/traditional accounting, you put the selling price of the machine (if you have to liquidate the company) on the side of the asset: let's say 8 millions the first year, and 5 million on the debt side (they bought it with 5 million of investment money, leveraged with debt). Company has 8 millions in assets, 5 million in debt: good ratio, healthy company.
Company B use future accounting: it buy the 10 million dollars machine. On the 'asset' side, you put 16 millions: the machine is worth 8, and is expected to earn the company a net profit of 8 during its lifetime. On the 'debt' side, you put 5. 16/5, the ratio is truly excellent. So good even that you can afford to take another 5 millions, and buy share/issue dividends. Now your asset/debt ratio is 16/10=8/5, same as Company A, except either your current shareholders have already reimbursed the initial investment, derisking themselves totally if the company ends up failing.
I think you have a Republican primary candidate who used that with his drug company, the company failed to produce anything of value, but only late investors lost anything.
But then they failed at that as well. Their share price almost never increased after 2000s. Currently they are valued lesser than their 1996's valuation, even if you exclude inflation. Highest they went was $263B valuation in 2021, which is nowhere close to tech giants.
His explantation is actually closer to the truth than your array of many CPUs. Because GPU isn't that. It is / was moving in that direction but it is certain not many CPU or Intel Larrabee.