Nvidia Announces A100 80GB GPU for AI(nvidianews.nvidia.com) |
Nvidia Announces A100 80GB GPU for AI(nvidianews.nvidia.com) |
Their lead example is recommendation systems, but I can't say that I have received many good suggestions recently.
Spotify and Deezer both suggest the chart hits, regardless of how often I dislike that kind of music.
Amazon keeps recommending me tampons (I'm a guy) ever since I've had a coworking office with a female colleague in 2015.
For all the data that they collect and all the AI that they pay for, these companies get very little revenue to show for it.
There are 1 billion users, each of which spend x$ on the platform. The recommendation system does not need to get 100% accurate (it'd be very hard to get to something that 10% nonaccidental/nonfraud click through rate for example). It needs to be slightly accurate. The difference between 0.5% ctr and 0.6% CTR or conversion is probably 20% increase in revenue, and much more in profits (assuming fixed cost).
But it’s not, it’s a tractable similarity based solution to the question of, which of millions of ads to show, in order to increase CTR.
The biggest difference is motive: a good recommendation is made in good faith for the benefit of the recipient while ads are businesses trying to turn a profit. Maybe that’s why this meme never dies.
People aren't saying recommendations need to be good, but just they need to be better than they are right now. To justify the cost of investment in AI, AI hardware, and AI developers an AI recommendation system needs to be more accurate than the next best recommendation system. The real point here is that no matter how clever you think your AI system is the end user still thinks it's worse than a very simple system based on "people who bought X also bought Y".
Good example!
Not really, they make a shit ton of money from their recommendations and perform extensive A/B testing to keep track of how much money it makes them. Anecdotes aren't data and large tech companies don't spend money for no reason (especially when they can A/B test the impact trivially). Remember that at their scales even a 0.01% increase in revenue is worth $10+ million per year so they don't need to be perfect to make a shit ton of money. There's a reason ML engineers get paid $1+ million and it's not corporate stupidity.
I bought one toilet seat on Amazon and now it thinks I'm a toilet seat collector. No I'm not going to buy another one any time soon, no matter how many color and size/shape/design variations it presents to me.
I really think I would watch more videos if the recommendations -challenged- what I enjoyed, rather than couching me, but I have to assume that someone has run that experiment and ad revenue went down.
I have been an Apple Music user, but I am subscribing to Spotify just for its recommendations.
Just within electronic music, look at how many genres and sub-genres have been categorized by someone simply as a hobby project. Now updated for the web 2.0 era.
Do you have a memory problem? My weekly have always been 30-70% the same recordings of songs I have already listened to at least ten times. And it's always been this way.
You just bought a light bulb? How about a dozen others?
But I've realized that accuracy is not really a problem for them, because a false recommendation at most slightly annoys the user and get glossed over. Among 1000 wrong recommendations, if one works, it's a win (for the company).
Ironically, this suggestion is probably driven by real-world behavior. Whether your bulb fails due to age (and you'll probably need replacements for your other similarly-aged bulbs soon), or an electrical problem caused the burn (and you might need another replacement soon), or you just want to stock up, having your exact bulb model recommended also saves you a little time on having to check it yourself. :P
I think a lot of "wrong" recommendations are actually right for other users, which seems to be just what you're saying -- they annoy one user but result in more sales for others. From the company's perspective, these recommendation systems are working as expected.
On the bright side, that also means in their best interest to improve accuracy (and drive more sales for more users, instead of just annoying them). Hopefully new tools like Nvidia's announcement result in fewer annoying ads.
- frequently click through on Facebook ads because they're often SaaS products I'd be interested in (and have found a bunch of cool tools I use now)
- think my Spotify suggestions are spot-on. I used to use Google Music which had comparable suggestion qualities at the time, but I feel like Spotify has gotten significantly better at suggestions over the past year-ish.
- think YouTube is the shining example of controllable recommendation systems. Looking at my front page right now, I'm interested in all 8 of the videos above the fold, and almost every video below it -- probably a direct result from actively guiding/curating what videos get recommended to me. My "to watch" queue is hundreds of videos long since I almost always add more and more videos until I get the time to sit down and watch a chunk, which usually turns into a positive feedback loop of more good videos getting recommended.
All three of the above recommendation systems make me enjoy the related product more than I would without them, and probably also directly lead to more revenue for the company (sales on FB, keeping my Spotify sub, and seeing more YT ads).
On the other end of the spectrum, posts on FB and Quora are two examples where recommendation systems seem to make products significantly worse, so I guess it's hit or miss depending on whether what you want out of each product aligns with how the recommendation systems are set up.
Youtube is the worst. I have all the time to open videos (and think about it!) in private tab so that my recommendations don't get messed up. And then of course on phone you don't have that luxury so you have to go clean history manually. And now they added two interstitials before you can actually view video in private.
Shining example indeed
That we just remember the bad examples when den recommendation was bad but not when it was good because we didn't make the connection.
I'm asking myself that for quite some time.
Do you use a tracking blocker. That could also be a reason why you get bad recommendations.
Seriously though - When amazon is showing me all these recommendations, they are charging someone to show these to me. That means I am not the customer of those recommendations. They make money when I browse and I wonder if they make more money doing that than when I actually purchase. Meanwhile to the people selling - they pay to advertise with amazon, and they pay (via a %) to sell as well.
"It is difficult to get a man to understand something, when his salary depends on his not understanding it."
-- Upton Sinclair
In other words the reason might be that there is an incentive to suggest you chart hits. The abilities of the recommendation have probably not much to do with it.
Of course they do. Their real customers, record labels, pay handsomely for that service.
The release radar playlist may have more chart hits, but this one is not super smart. If you listen to a popular artist you will get all its new songs and shitty remix forever, including its chart hits.
That is just part of the game and it is called "machine learning". They get to know you better. ;)
Annoying but only part of the game/training.
Most users won't care what hardware their PyTorch model runs on in the cloud. All that matters for them is dollars per training epoch (or cents per inference). This could be a steal for an alternate hardware vendor.
[0] https://web.archive.org/web/20201109023551/https://www.digit...
Great ideas are much harder come by. And while you can try and buy your way to the best ideas, it doesn't seem to work all that well.
>So there is a back and forth here and all of the other players that we compete with, they are also economic animals. They don't have any magic elixir or magic algorithm that we don't have, right. We all kind of are doing the same thing and we're all providing great service and value for the marketplace. So I know this a little handy way to give an answer but there is an ebbs and flows around competition. The business continues to be very profitable for us on a net basis. On a gross basis it's incredibly profitable but as I said in my remarks, we have paid, put that in quotes, not in our financials we've provided back to our retail customers about $950 million of price improvement this year.
https://seekingalpha.com/article/4386120-virtu-financial-inc...
Top quantitative firms like RenTec don't rely critically on speed or compute power, but rather on high quality data, which they've also said publicly.
over 2 terabytes per second of memory bandwidth."
[0] https://www.nvidia.com/en-us/data-center/dgx-station-a100/
There are processors that are designed only for DNN inference but A100 is not one of them.
$US 200K for startling performance this time.
Compiler bugs a plenty: the compiler can enter infinite loops just trying to compile OpenCL 2.0 code, taking down your program. If you ever come across such a bug, you're now in guess-or-check mode to figure out exactly what grammar you did to bork the OpenCL compiler.
Oh, and the OpenCL compiler is in the device driver. As soon as your customers update to Radeon 19.x.x.whatever, then you have a new OpenCL compiler with new bugs and/or regressions. The entire concept of tying the COMPILER to the device driver is insane. Or you get support tickets along the lines of "I get an infinite loop on Radeon 18.x.x.y drivers", and now you have to have if(deviceDriver == blah) scattered across your code to avoid those situations.
In practice, you end up staying on OpenCL 1.2 which is stable and has fewer bugs... and has functional debugger and profiler. But now you're missing roughly 8-years worth of features that's been added to GPUs over the last decade.
----------
ROCm OpenCL is decent, but that's ROCm. At that point, you might as well be using HIP, since HIP is just a way easier programming language to use.
Ultimately, I think if you're serious about AMD GPU coding, you should move onto ROCm. Either ROCm/OpenCL, or ROCm/HIP.
ROCm is statically compiled: the compiler is Clang/LLVM and completely compiled on your own workstation. If you distribute the executable, it works. Optimization flags work, there's a GDB interface to debug code. Like, you have a reasonable development environment.
So long as your card supports ROCm (admittingly: not many cards are supported, but... AMDPro OpenCL tooling is pretty poor)
Also, no Windows support whatsoever...
A more detailed explanation (ROCm section) https://timdettmers.com/2020/09/07/which-gpu-for-deep-learni...
AMD has HIP, which is closer the Cuda but HIP seems less developed.
But I believe the basic problem is AMD doesn't make sufficiently high end GPUs to compete with Nvidia in ML.
Personally, I don't care what happens in the cloud, just what I can buy. I would note Nvidia does have competition in the cloud from Google's TPUs and I assume any large cloud vendor is going to negotiate with Nvidia. While I'd love AMD to be cost-effective for ML somewhere, it seems they aren't 'cause that's not what they're targeting.
Tensor Processing Units are too specialized: they can't traverse a linked list, they can't traverse trees. They're good at one thing and one thing only: matrix multiplication.
GPUs are still bandwidth-optimized and are good at matrix multiplication (but not as good as tensor units). But GPUs can traverse trees and new data-structures. Ex: BVH trees for raytracing, or linked lists... or whatever else you need. Its a general computer, a weird... terrible latency computer with HUGE bandwidth... but that's still useful in many compute applications.
--------------
Matrix multiplication is the cornerstone of many scientific problems. But you still need software to manipulate the data into the correct "form", so that the matrix multiplication units can then process the data.
Its in this "preprocessing" or "postprocessing" phase where GPUs do best. You can implement bitonic sort for highly-parallel sorting / searching. You can perform GPU-accelerated join networks for SQL. Etc. etc.
And even then, NVidia's A100 have incredibly good matrix multiplication units. So you're really not losing much anyway.
On the other hand, Google has good reason to hold back TPUs from general public, and instead only offer them on Cloud. That also contributes its limited use.
What is the reason? Just competitive advantage for running Google level AI/ML workloads?
If I had an acquaintance who's a recovering drug addict, then I would never ever recommend him to try out a new drug, because that is obviously against his best interests. But for an ethical void such as a company blindly pursuing profits, bombarding him with free samples for highly addictive drugs would be an effective strategy for increasing short-term profits.
You'll have some kind of MAB in production, but in order to measure individual incremental impacts, you still need a proper experiment.
Instead of focusing on what I just bought, they need to focus on predicting what I'll want to buy next. Just because I bought a coffee maker doesn't mean I'm turning into a coffee maker collector.
Yeah the actual humans with ears are me and you. They use play-time as the basis for recommendation as well as content-similarity, and put effort into de-biasing this data for, e.g., position bias. This is the same for YouTube etc.
Spotify is one of the best examples of a modern large-scale recommender system, for me.
Both of those are low probability, but they're probably higher probability than the probability that you are interested a product chosen uniformly at random.
Why would you assume I use discover weekly to discover music?
I just use the Radio feature, which plays endless amount of songs based on the song you started with.
The radio feature is global to all playback situations so I assumed that couldn't be it.
[1] https://docs.microsoft.com/en-us/windows/win32/direct3d12/gp...
1. Technically, TPU is less generalized, making it publicly appealing requires too much engineering effort, which has very high risk of not paying off. Case in comparison: How AMD is not able to capitialize on GPU in Deep Learning despite more uniform architecture.
2. TPUs does have some advantage, that Google want to keep at its own possession. For example, Google would not want FB to easily copy TPUs. FB indeed very likely benefit from TPUs, as they are running similar business.
They are. Its called "Use ROCm". Tensorflow support, PyTorch support, etc. etc.
Yeah, its limited to Linux, its limited to a few cards. But within those restrictions, ROCm does work.
Both with AMD MI100 providing the bulk of their compute. Frontier seems like it was given development boards of MI100, because AMD is talking about how they already ported some code over to the MI100 and tested it.
Nvidia has quite a head start. You're not just talking about some simple driver support either. You're talking about runtime compilation/JIT(to target various flavors of HW), tooling support, library optimizations, API stability and maintenance... AMD can catch up, but unless they come up with a new approach it's going to take a long time and a lot of smart people to do so.
I think they will. AMD has the challenger mindset. They rose from the ashes and now actually compete with Intel and they can tackle NVIDIA as well.
Same with bogus results in Google search. It would be a mistake to fixate on a fail case at the expense of seeing what it gets right.
One thing that can be said about Amazon is how data-driven it is. Even an obvious "improvement" to a system would require analysis to back it up as an improvement. For example, it might seem obvious to filter out lower quality user-created answers in the product FAQ, but answers with poor grammar might actually boost sales because shoppers trust the answer more.
Also, as we descend deeper into ML/AI and black boxes, the deeper we get into effects from afar. There's no real place to write if (user.sex == M) then weigh('tampons', -1) as it was a constellation of factors that cascaded into a man seeing tampons like that time he purchased something related for his girlfriend. The next rung in line is the business of mind-reading.
Not that that justifies the practice.
And, yes, I do think Facebook should take responsibility for the content of their ads and products sold through them. If the New York Times only sold ads for scams, and all their articles were lies, we would no longer trust the New York Times
Facebook is like Vegas strip junk ads. It may look nice, but you’re going to have a bad time
Facebook and Google would have to innovate instead of rent seek.
Ads/marketing budgets create employment that would otherwise not be there.
Ads provide monetization for apps you use on your phone. Without monetization you wouldn't have the proliferation of the ads, and not everyone is willing to pay directly.
It creates jobs, supports businesses, helps otherwise unsustainable content. It helps the small businesses grow their market. It supports economy.
You are also sort of missing the point. We are trying to make ads relavant for everyone, but even if it's relavant, you don't click on everything you see. This is true because of how you as a person do things. You don't buy the first car you see on the street, you do research, you spend days etc. Just think of that, and you'll see why why 1% ctr is actually not a terrible thing as you made it sound like. if you clicked on everything you see, you wouldn't be able to do _anything_.
Disclaimer: Current FB and ex-Google employee working on ads for 8 years. And no, my pay doesn't depend on me saying these, i could just work _anywhere_ i wanted, literally.
Actual money.
In the absence of ads, we'd have people paying money to honest app makers for the utility of the app. And along with that, the accountability, honesty, and incentive alignment that comes from a straightforward exchange. Instead, we have app makers selling users, their data, and their attention to the various companies/intermediaries/entities involved -- without ever telling the user how the bread is buttered.
Yeah, when I'm buying a toaster, I do research. But that has nothing to do with modern ad-tech and its various tentacles whatsoever. Actually, nowadays, I find this whole ecosystem actively prevents me from being proactive about my consumer choice, because every single thing I see when I google "best toaster 2020" is some type of eldritch symbiosis of Fb Goog Az (choose any; none are good), a 1 cent-paid-per-letter 3rd world content farmer, instant click bidding based on my browser, age, gender, location, prior 2 week consumption pattern, political view.
It's really tough to compete with free, and I understand that there are significant barriers to paying directly for content/utility/etc. I know, the system we have set up is ambiguous and complicated and subtle. But maybe those ads/marketing budget dollars could be used to actually like, I don't know: improve people's lives? fix these systems? reverse course? Convince people to consume less so that we don't collapse the biosphere?
It's so tiring to see the most intelligent people alive say, well, this is quite clearly a massive problem that might literally destabilize the order of our entire society on one hand, but on the other, people just have to know how bad they want the Newest Garbage On Sale This Upcoming Black Friday...
Yes, it is a lie and a scam. One of the most epic of recent years.
How do you think you'd get to know your neighborhood burger joint without some form of advertising (either word of mouth, or through real paid advertising). how would you know the burger joint somewhere else?
Short of a micropayment solution that pays out proportional to the value you get (flattr? maybe), ads is the only viable option. But that only solves the content creator pov. It doesn't help the advertiser - the business that need your money to survive.
I don’t think so; and the counter example is gaming. We would see free to download products with upsells catered to whales. I’m not sure if that would be a better product but I don’t think that business model is honest either
[Replying to parent because I can't reply directly to poster]
It looks like you have no idea that there are games worth paying for. Not on mobile phones though. And definitely not free to play games. They are designed to take your money, not entertain you.
That, and the seemingly utter stupidity of ad engines. Like, yes I bought that new power tool last week. Stop showing me the ad for the tool from the same store I already bought it from.
I'm pretty much convinced that apart from the people working in the ad business, nobody actually profits off ads. Certainly it's close to impossible to prove that ads are effective, and people who sell ads to companies are good at cherry picking and suggestive correlations. Then in the end everyone tries to buy the same eyeballs by paying the same ad companies, and it all averages out to nothing except you spent a bunch of money.
Yes. I hate this too, but think of it as a bug, not the actual intention. We would love to know when you wouldn't buy a product as much as we would love to know when you actually would. But we don't, always, and end up having to make approximations. We don't always know what you actually bought, we just know you bought something. Advertisers don't even always tell the value of stuff you bought, something tht'd have benefited them to share from ROI pov etc.
Whether something creates jobs or not doesn't make it a net good. Would you suggest Ransomware is good? Because it also creates jobs. As do Nigerian email scams and Ponzi schemes. "Creating Jobs" doesn't mean something is good for society.
> You are also sort of missing the point. We are trying to make ads relavant for everyone, but even if it's relavant, you don't click on everything you see.
Apparently you are missing the point here. Whether you see it or not, Facebook and google's marketplaces by nature serve advertisers, not me. Advertisers don't give a damn what advertising is "most relevant" to me. What they care about is which demographics are most profitable to their brands.
> We are trying to make ads relavant for everyone
So long as Google and Facebook put the advertisers in control of who sees their advertising, any assertion that the system is designed to serve us is bullshit. People don't see advertising that is relevant, they see advertisements from the people who pay google/ Facebook to see their message. Those two things are not equivalent and never will be.
But think of it this way. You are getting a service, you are paying either directly or indirectly through ads.
Facebook wouldn't be a thing if it were a paid product from get go. It'd probably even lose a major part of its user base if it became a paid product overnight.
Like i said, i would love if i can pay for a product i am benefiting from, but it's not always possible.
But they don't put advertisers in control of who sees their ads.
If you look at audience sizes on FB, and then run some direct response ads on that audience, you'll notice that you only ever reach maybe 10% of that audience.
This is because what FB/Goog are good at is figuring out which ads are likely to get someone to click and/or convert, and show only those ads.
The dirty secret is that those people might have converted anyway.
One can measure this with an attribution model, but the trouble is that the two biggest players Google and Facebook have very little incentive to co-operate, so all attribution models are extremely biased.
tl;dr the advertisers set boundaries on who should see the ad, but they don't control who the ads get served to.
Or you sell the 1st elsewhere like Craigslist.
Either way, the numbers show those ads (remarketing ads in industry speak) are insanely effective.
Clearly, nobody is interested in fixing it then?
"It's a developing area" is a pretty old adage by now, and is worn out. Modern tech exists for 25+ years now. Do something. Most of the people I've spoken to hate ads.
Not fixing ancient bugs isn't helping the situation.
If everyone spends money on ads, a single actor choosing to not spend money on ads would (probably) see a loss. And in this case, you have a centralized actor (ad companies) spending a lot of resources telling everyone this fact.
But if everyone decided to not spend money on ads, everyone (except ad companies of course) would see a win.
>Come now, this is only an enhanced form of prisoners dilemma, where the prison guards do a side hustle helping prisoners rat each other out, in exchange for a "very small" commission. Everyone spending money on advertising is a strong Nash equilibrium, but it is not the optimal solution.
Imagine you are making your own craft beer in your basement and you are sitting on two crates of beer. You want to sell your beer but you don't tell anyone that you are selling beer because that would be advertising and according to you advertising is not an optimal solution. Even if we assume you are the only company on the planet and have no competitors, your business is still in trouble and about to make losses and close down.
>If everyone spends money on ads, a single actor choosing to not spend money on ads would (probably) see a loss. And in this case, you have a centralized actor (ad companies) spending a lot of resources telling everyone this fact.
Now we assume that you tell someone that you are selling beer which is basically what advertising is. People know that you sell beer now. They can now make a decision to buy your beer. If people like beer that's what they are going to do. If they hate beer they can still decide to not buy your beer.
Yeah, if you did not advertise then you would indeed see losses. Simply because nobody is aware of your products, meaning they are unable to buy your products. But again, no second party is involved, so you don't need a centralized actor to decide to do advertising.
>But if everyone decided to not spend money on ads, everyone (except ad companies of course) would see a win.
Since we are the only company around we are "everyone" and if we decide to not spend money on ads then we would see losses.
So now that I have proven that your hypothesis is not correct we can actually talk about the ad market in general.
Advertising is providing value for companies but the amount of value advertising can provide is not based on how many resources you spend on advertising, rather it is dependent on the size of the market. A big market with lots of consumers can make more money off of more advertising but there is a certain point beyond which you end up spending more on advertising than the market needs. On a planet with 1000 consumers and two companies the best case would be if both companies put out 500 ads. So clearly the optimum amount of advertising is not 0. However, a big company can put out 1000 ads and thereby displace its competitor's ads. It's best to have some ads but not too many.
The problem discussed today is the invasive and often clueless nature of online ads. They just spam you and nothing else. Adtech companies gather mountains of data and at the end of the day what do they have to show for it? "Oh, we see you bought a gaming machine. How about you buy these other five ones?". lol.
Seriously. This is sloppy work.
My conclusion is that they don't utilize the private data (that they gather in some very legally questionable ways) and thus they won't lose anything if they're robbed of all the personal data. And the world will be better for it.
Over all businesses over long periods of time, maybe. But a business can definitely be dumb. And a smart business can definitely do a dumb thing.
What I'm getting at is that some things are better for giant tech companies and corporations and worse for regular people and some things are better for regular people but worse for giant corporations and tech companies.
I'm not even saying that I hate that this is the case, I understand our reality. The lack of creativity, the lack of imagination on this from anyone at all, but especially our best and brightest -- that is the worst part about all of this.
The absolute depth of monoculture on these issues is "oh for sure, it's messed up, but like, fixing it is too hard because of how messed up it is. Better double down before this whole thing implodes!"
Here's the difference: If I am paying for a thing, I have a choice. I am never presented with the choice over whether I want to get tracked online. Or between apps. Usually tracking is invisible and completely obfuscated in such a way that even if you want to know who is tracking me and what is getting tracked you can't.
> How do you think you'd get to know your neighborhood burger joint without some form of advertising
It's in my neighborhood, I see it when I drive by, friends recommend it. Sometimes I do a web search. I don't think I've ever found a restaurant (grocery store, pub, etc etc) due to an advertisement. About the closest I get is when the local paper runs their people's choice awards for local businesses. (and I know, the local paper gets revenue from advertising)
The only choice you have is to stop paying for the thing. In reality it's more likely that you'd end up paying with money and with the data that's being collected. Businesses always want to grow revenue so at some point collecting data again or serving you ads in a paid product constitutes low hanging fruit.
Look at Samsung and the ads they force on you after you paid thousands on their TVs, look at Amazon who crams some ads in movies and shows you already pay for with Prime, look at Google who still collects info on you even if you pay for YouTube Premium.
This isn't about paying with money or your data. You may get something for your money at first, until you don't anymore.
TV manufacturers were forced to add AD revenue because they reduced their purchase price to levels that were less than sustainable so there is a trade off being made there.
The area is ripe for disruption. But sadly, that's the world we have to live in. I am pretty sure google would have prefered if you paid for the services you get (i suggest you sign up for google one, if you use gmail/drive/etc). But until that's ubiquitous ads is what we have.
re: neighborhood burger joint - web search implies someone is providing you this for free. or through ads. or you pay.
I certainly would and have. I highly suspect that's the majority (or at least a significant percentage) of money made in the gaming industry.
Very, very different markets.