Hacker News Reader

Comment by pcwalton | Hacker News Reader

I'm happy to see vendors paying attention to having working graphics hardware.

It's really sad how "you can actually use the 3D functionality in the chip" is a selling point that differentiates these small ARM boards. OpenGL ES 2.0 will be 10 years old next year. There's no excuse to not have a working implementation of it accessible.

Imagine if you had to shop around to find a PC where the GPU actually worked without crashing (ODROID XU4 fails, Google "black screen bug"). Or if being able to play high-definition video was something you couldn't rely on, despite the hardware being there (again, ODROID XU4 fails). Or if the 64-bit mode fully implemented in silicon was inaccessible in practice without losing other hardware functionality (Raspberry Pi 3). This is sadly the situation we're in with these ARM boards, and I don't see it improving unless we demand working drivers.

View on HN

How is it not a selling point? GPUs aren't cheap, someone has to develop the IP, drivers and support tools.

I totally get a 2D accelerator being standard but 3D is complex enough and has such a steep curve for building content/using it correctly that I don't see why you'd include it by default.

[edit]

There's also a only a few ways to do 3D reasonably on embedded at low power and you're probably going to run afoul of all the patents in the space if you try to build your own IP.

I believe pcwalton is saying that it is sad that "you can actually use the GPU" is a selling point in the market for embedded devices that have a GPU, not that having a GPU shouldn't be a selling point. Many other boards come with processors with integrated GPUs but they're unusable because a driver is unavailable or some IP is under lockdown by an upstream vendor.

Yes, but that's what I was getting at. From a tape-out perspective it costs more to have a second run w/ no-GPU vs a power gated GPU all in one die(which is also why PoP memory exists).

The IP is locked out because there's an extra cost to the business to use it.

Also embedded/mobile GPUs haven't been around nearly as long as OpenGL ES. Heck the Adreno 2xx which was one of the first real-ish GPUs was released in 2009 and even then you'd be lucky to get any sort of reasonable fill rate out of it. These type of GPUs really haven't come into their own until the last 3-4 years.

That's not the case on any of the models I've been complaining about (ODROID XU4, Raspberry Pi), nor the PocketCHIP (according to this). They have various solutions available to access the GPU. It's just that the drivers don't/didn't work.

> Heck the Adreno 2xx which was one of the first real-ish GPUs was released in 2009 and even then you'd be lucky to get any sort of reasonable fill rate out of it.

Adreno 200 is one of the worst offenders! There was no excuse for not being able to call glTexSubImage2D() without crashing the driver (for example). Qualcomm had the money to create a working driver; they just didn't.

Yeah.

Sure, tape-out may be cheaper with one chip line, but we aren't talking about chips where the feature is there in silicon but intentionally disabled (say, for the other more expensive variant). We're talking about a feature that the chips and systems are advertised as having, which doesn't work as it is intended to because the driver stack is broken. Where the vendor specifically says '....supports OpenGL ES x.x', but then you find it sort of does but not really and it's broken and won't ever be fixed.

It's even more frustrating because there's a whole open-source community that would be willing to do the work of writing the drivers even, but then of course the vendors won't give them the data they need because of IP concerns. It just really sucks.

OpenGL was never an advertised feature of the C.H.I.P as far as I can tell. Economies of scale mean that it's cheaper to reuse a tablet SoC and just ignore the GPU entirely than it is to use a chip where all the functionality is available. The GPU probably works fine in its intended application, low-end Android tablets.

Well the ODROID and RPi have two different IP stacks, Broadcom is a bit better since they own the IP but I wouldn't be surprised to see Arm charging for Mali support.

Each vendor has their share of driver/hw problems, some are just more widely published than others :). That bug was gnarly on the 200, but it was fixed. If the GL feature wasn't included in Android's HWUI renderer all of the major vendors had at least one thing wrong.

> I wouldn't be surprised to see Arm charging for Mali support.

That's a problem, if so! We wouldn't accept "we won't give you working drivers for the hardware you bought unless you pay us" in the PC world; we shouldn't accept it in the embedded world either.

It's not so bad to not have a GPU at all (though given that ARM owns Mali and you're licensing from ARM anyway, it seems prudent to have one if you can). The real problem is not being able to use the hardware that is there due to driver issues.

Reply on news.ycombinator.com