W3C home > Mailing lists > Public > public-gpu@w3.org > August 2019

Re: Some Feature requests.

From: Kai Ninomiya <kainino@google.com>
Date: Wed, 7 Aug 2019 13:49:52 -0700
Message-ID: <CANxMeyAKMr87OZjU6ZMrGEu8awvTYoK=+WteVbtE24xuS9UzzA@mail.gmail.com>
To: Kevin Rogovin <kevinrogovin@invisionapp.com>
Cc: Dean Jackson <dino@apple.com>, Doug Moen <doug@moens.org>, public-gpu <public-gpu@w3.org>
Just want to jump in and say that, in our experience, for GPU compute even
more so than graphics, the optimal code can be wildly different on
different GPU chipsets/architectures, and it's even more important that
apps reach peak performance. We should expect that frameworks (like
TensorFlow.js) will need to know which of several architecture-specific
hand-tuned implementations of each performance-critical operation (like
matmul or conv2d) to choose from for best performance.

This doesn't necessarily preclude us from hiding this information behind
some user action (like a permission prompt, or PWA installed app), but
sites that can't get that permission may resort to brute force tests as
mentioned by others.

On Wed, Aug 7, 2019 at 1:30 PM Kevin Rogovin <kevinrogovin@invisionapp.com>

> Hi,
> I am sorry, this is very counter-productive to the success of WebGPU whose
> goal is to close the performance gap for hardware accelerated graphics
> between native and web.
> The previous posts prove that that it is quite possible and feasible to
> identify (roughly) the GPU at the cost of slower start-up performance and
> batter life on mobile. GPU-intensive apps will NEED to do this (because the
> performance profile for different GPU's is literally all over the map
> across architectures for various specific loads), thus it will be done.
> Rather than making the developer's life more difficult along with making
> the user's experience worse, there is a pretty clear way forward: provide
> the GPU info directly. If hiding this is important to a user, just as many
> browsers allow for changing the browser ID, perhaps a browser can also
> provide an option to NOT provide the GPU (i.e. a WebGPU implementation
> would report "Generic GPU"). This would then give customers the control of
> identifying the hardware, and for those users which take that option who
> use a highly GPU intensive app, the downside that it the app will run
> perf-test-probing along with other probing on the device. In all honesty,
> compared to all the other shenanigans going on with browser tracking, this
> is by far the smallest potato.
> If we were in a world where the GPU architectures were quite similar, this
> would not be needed, but that is not the world we live in (in all honestly
> thankfully too, since variety in hardware is a good thing in my eyes).
> My 2 cents.
>  -Kevin
> On Wed, Aug 7, 2019 at 11:04 PM Dean Jackson <dino@apple.com> wrote:
>> > On 7 Aug 2019, at 14:05, Doug Moen <doug@moens.org> wrote:
>> >
>> >
>> >
>> > On Tue, Aug 6, 2019, at 10:35 PM, Myles C. Maxfield wrote:
>> >> We’ve heard this defeatist argument before and entirely disagree with
>> it.
>> >
>> > My point is that security designs by non-experts have a poor track
>> record.
>> > I think fingerprinting security for WebGPU should be designed by a web
>> browser fingerprinting security expert
>> We agree. This *is* advice from our Web Browser fingerprinting security
>> experts.
>> It's not a secret that there are low-level techniques to identify
>> hardware, and thus users. Their existence does not mean it is acceptable to
>> provide high-level ways to identify hardware/users.
>> To give an analogy (that will probably cause more distraction than help,
>> but whatever).... just because someone can throw a brick through your
>> window doesn't mean you should leave your front door unlocked. Maybe some
>> day you'll get brick-proof windows.
>> Dean
>> > , and I also think that you are trying to fix the problem at the wrong
>> level of abstraction. The security should be provided at a level above the
>> WebGPU API.

Received on Wednesday, 7 August 2019 20:50:29 UTC

This archive was generated by hypermail 2.4.0 : Friday, 17 January 2020 19:52:26 UTC