Hacker News Reader

Comment by SamuelAdams | Hacker News Reader

It doesn't have to be that bad. Every organization has constraints - particularly people's time. If Netflix, Amazon, and Hulu come up with some new video codec and plan on releasing it in 12 months into their services, and I know 80% of my consumers use video streaming services, then I know that I should allocate most of my development resources / people towards enabling the new codec on GPU's my company sells.

Knowing how people use your products helps allocate new development features. This problem isn't unique to intel.

View on HN

If you have 80% of the desktop market, you can just aggregate web traffic statistics - your customers will be a representative sample

Every organization has constraints and in Intel's case they are constrained to only $70 billion in revenue.

That sounds like a finite number to me.

In 2018, Intel spent $15B on share repurchases and dividend issuances... if more cash could improve their offerings, they aren’t acting like it.

Over 20%. Is there a P/E ratio-analogue, but for this kind of dead weight financialization?

Is this codec scenario plausible, that something so significant would be broadly adopted and Intel would need user usage data to decide whether to support it?

Usage data could be very useful for decisions on what to dedicate die area to, and how to price products.

They might find that a market segment doesn't do a lot of playing 3D games, but does spend a lot of time watching Youtube on battery, and so make the decision to drop integrated GPU shader cores and dedicate more die area to video decoding, to make energy efficiency better when watching video.

But now we are not speculating about potential new additions, we want to know utilization of existing circuits to allocate resources & improve in the future. I'd hope this could be measured without collecting a ton of personal data.

These goals blur together. R&D effort, and usage of die area are all justified by market potential. Knowing how large the market for a technology or tradeoff actually is has value.

> I'd hope this could be measured without collecting a ton of personal data.

Intel's claim is that they don't: "we would like your permission to collect [...] The categories of websites you visit, but not the URL itself" also the privacy safeguarding statement at the top of the page.

The whole purpose of CIP is to gather data on actual users' computing needs. For many computer users, the web is most of their computing experience -- it seems appropriate therefore to include usage of the web in the program.

Much like any customer experience study, privacy is an issue. Intel do at least claim to be mitigating the impact to privacy.

I'm of the opinion they should add FPGAs to everything. Every chip now has an insane amount of die area dedicated to the northbridge or the L3 cache. Anything else turns the junk we buy into garbage the moment everyone switches to a new standard.

Aren’t FPGAs slower than ASICs though?

That depends a bit on whether you count the fabrication time for a new ASIC towards the application startup time when switching applications or not.

Certainly propagation delay is higher. But for many applications it is cheaper to implement them in software than ASICs.

And besides, FPGAs are used for web scale applications, many of them applicable to small computers.

Edit: so they need to know every URL I access because they what to improve GPU or their drivers?

That's not what Intel is collecting. From the article:

> what data is collected and how is it used?

> The categories of websites you visit, but not the URL itself

> If I participate in the program, is there any personal information in the data collected?

> Will not include the URL (web address) for specific sites visited

Unless they're shipping a fairly complete domain -> category map to your machine and doing all categorization offline, unlikely, then they absolutely have the ability to collect it, and will "accidentally" do so at some point. We've all seen this movie now, multiple times.

edit: also the knee-jerk reaction most people are going to have to this, is concern about certain categories of websites.

Third party services (e.g. https://www.webshrinker.com/apis/) exist to perform this kind of classification. They'll be doing the classification API calls from the client (CIP-enrolled) machine, then sending the results (and not the domains or URLs that generated them) back to their own endpoint.

Then webshrinker would know everything you visit right ?

> The information collected includes categorized web browsing history that shows how long and how often you visited specific categories of sites (i.e. social media, personal finance, or news). All site visits are classified into one of 30 categories. We do not collect URLs, web pages titles, or user-specific content without explicit permission from you.

From a technical perspective, how does this work? I can only think of two ways:

(1) Intel downloads a big "map" file containing a huge list of pre-defined domains and categories. This means either the categories are pretty slim or you have a GB's big file

(2) Each URL is sent to Intel (aka "collected"). Maybe they don't "store" it ("We promise") but as soon as it leaves your system you have no control on this

They could use a technique like Google Safe Browsing where you send a partial hash of the url, and they send back the subset of the map containing only urls which could match that hash

They can strip the request down to the domain name, hash that with a OTP to do a fuzzy lookup on their end which would be difficult (not impossible) to reverse per reference, but would require APT to do en masse.

There is probably some flaw with this method... curious if others think there's a "safe enough" way to do something like this...

Yes they can, but they can not

Yes

And intel want it for free

Reply on news.ycombinator.com