[HN Gopher] Raspberry Pi 4 achieves Vulkan 1.1 conformance, gets...
___________________________________________________________________
Raspberry Pi 4 achieves Vulkan 1.1 conformance, gets GPU
performance boost
Author : rcarmo
Score : 280 points
Date : 2021-10-30 15:03 UTC (7 hours ago)
(HTM) web link (www.cnx-software.com)
(TXT) w3m dump (www.cnx-software.com)
| rixrax wrote:
| What's the test software / benchmark I should use on Linux
| nowadays to measure (and compare) shader and raw GPU performance?
| That would ideally run under both X and Wayland?
| arminiusreturns wrote:
| I have always tended towards Phoronixs test suite
| (https://www.phoronix-test-suite.com/) but Im sure there are a
| few specific to Vulkan around. Not sure about wayland.
| fulafel wrote:
| The application(s) you want to run is the best benchmark.
| rixrax wrote:
| Problem with this is that the application I have in mind
| doesn't provide anything but perceptual feedback. I'd rather
| have some cold numbers that are to some degree reproducible
| and would give at least rough idea of the performance of
| given HW+drivers+other-settings combination.
| handrous wrote:
| Just checked, looks like Vulcan under DRM on the Pi4 works, and
| at least some people in the Libretro ecosystem have already
| messed around with it, so this could benefit Lakka. Awesome.
| Maybe this'll mean getting to play with some decent CRT shaders
| on the Pi without an unacceptable performance hit, and/or getting
| to make better use of Retroarch's advanced input lag reduction
| features.
| mewse-hn wrote:
| If this could get playstation emulators using vulkan running at
| decent frame rates that would be really, really awesome
| willis936 wrote:
| CRT Royale running 60 fps in a sub 15 W machine would be
| impressive. 1080p would be nice, 1440p would be great, and 4K
| would be best. The pi4 can output 4K60, but I really doubt it
| can shove through that many simulated pixels.
| exabrial wrote:
| ugh, now if I could only buy a few haha
| boromi wrote:
| Seriously I had to pay 75$ to get one recently, which was
| painful but I needed it.
| juanse wrote:
| Same here but with 10 units. Almost 100$ each.
| GhettoComputers wrote:
| What for? I'm sure you have some older computers that run much
| better that you already have in your house. An APU would
| trounce it. Pi feel like netbooks of desktop computers, the new
| ones get extremely hot, I would expect it to require a heavy
| heatsink and constantly spinning fan if you tried this.
| lytedev wrote:
| Better power efficiency
| ncmncm wrote:
| I guess this means you can use Kompute (kompute.cc) on RPi 4,
| now?
| lucb1e wrote:
| (A GPGPU framework, to save others a click.)
| ncmncm wrote:
| Except not a framework, it's just a library. You can use it
| to do any of the stuff you would do with CUDA, about as fast,
| but portably. #include it to accelerate your game's physics
| engine, or whatever.
|
| It doesn't say so at kompute.cc, but I found that it depends
| on Vulkan 1.1.
| Factorium wrote:
| Could we see a portable Epic Games console, pugged directly into
| their store?
|
| Like the Steam Deck, but better, since developers will get 88% of
| revenue instead of 65% on Steam.
| LeoPanthera wrote:
| The Steam Deck is a generic PC. It's not locked to the Steam
| store. It's not even locked to the OS it comes with. You can
| install the Epic store, or any other store, on it right now. If
| you have one, anyway.
| fortyseven wrote:
| *slurp*
| smoldesu wrote:
| You can download EGS games on Linux just fine, so ostensibly
| you could build one of these right now. Of course, you probably
| wouldn't want to use ARM for a PC game console, but you're
| welcome to try it.
| pengaru wrote:
| in TFA: s/Iglia/Igalia/
| hesdeadjim wrote:
| If only there was a world where Apple would sell M1 chips
| separately from their walled garden.
| bla3 wrote:
| The Pi has much better performance per dollar, which is a
| metric that's important to some people too.
| tinus_hn wrote:
| Purchase dollar? Or energy dollar?
| rbanffy wrote:
| I don't think there's anything else on the planet that
| rivals the performance per watt of the M1 family.
|
| Also, the RPi's SoC is made in an older 28nm process
| (that's one of the reasons why it's cheaper).
| Rovanion wrote:
| They won't. Their margins on the services side are obscene so
| getting people into that ecosystem is worth much more than the
| sales of some processors.
| mirekrusin wrote:
| Why? They gave "it's possible" proof. They rip benefits of
| doing it first - all good. Now it's time for competition to
| pick it up, possibly improve on it or fade away Intel style.
| smoldesu wrote:
| Also accepted would be a world where they just add Vulkan
| support to their APUs already.
| WithinReason wrote:
| How about MoltenVK?
|
| https://github.com/KhronosGroup/MoltenVK
| smoldesu wrote:
| It's fine, but it's frankly silly that you're forced to
| translate a _free and open_ graphics API into a more
| proprietary one. Compare that to something like DXVK, which
| exists because Linux users cannot license DirectX on their
| systems. MoltenVK exists simply because Apple thought
| "let's not adopt the industry-wide standard for CG graphics
| on our newer machines". Again, not bad, but a bit of a
| sticky situation that is entirely predicated by technology
| politics, not what's _actually possible_ on these GPUs.
| tinus_hn wrote:
| Is what is possible with Metal possible with GL though?
| Both in performance and features? They didn't build Metal
| just to be contrarian.
| bzzzt wrote:
| Metal was released a year before Vulkan. Apple just
| didn't want to wait and decided to design their own
| better than OpenGL API.
| oynqr wrote:
| Mantle was released ~1 year before Metal.
| smoldesu wrote:
| DirectX was released a decade before Vulkan, that didn't
| stop manufacturers from including support for both so the
| user could decide for themselves.
| my123 wrote:
| A fully compliant Vulkan implementation for M1 would come
| with very surprising performance cliffs for a developer.
|
| One of them:
| https://github.com/KhronosGroup/MoltenVK/issues/1244
| WithinReason wrote:
| And also potential optimisations that are not possible in
| other GPUs:
|
| https://developer.apple.com/documentation/metal/gpu_feature
| s...
| monocasa wrote:
| That's pretty common for TBDRs. The tile is rendered into a
| fixed size on chip buffer, and the driver has to split the
| tile into multiple passes to fit all of the render target
| data for nutty amounts of data coming out of the shader.
| PowerVR works the same way (completely unsurprisingly).
| fulafel wrote:
| See this comment on that issue: https://github.com/KhronosG
| roup/MoltenVK/issues/1244#issueco...
| zamadatix wrote:
| It'd be surprising if an architecture had 0 such surprises
| and did everything Vulkan allows without any special
| performance considerations vs another architecture.
| ChuckNorris89 wrote:
| Well, number one, why would they? Apple makes money by getting
| consumers and locking them into their unicorns and rainbows
| ecosystem where everything is perfect which makes consumers
| comfortable spending boat loads of money, not by selling
| commodity hardware.
|
| Ecosystems with great UX and paid subscriptions plus a 30% cut
| on all transactions are far more profitable than the margins
| you make selling commodity hardware. Just ask famous phone
| manufacturers like Siemens, Nokia and Blackberry why that is.
| That's why SW dev salaries are much higher than HW dev salaries
| as the former generates way more revenue than the latter.
| That's why Apple doesn't roll out their own cloud datacenters
| and instead just gets Amazon, Microsoft and Google to compete
| against each other on pricing.
|
| Apple only rolls out their solutions when they have an impact
| on the final UX, like designing their own M1 silicon.
|
| And number two, selling chips comes with a lot of hassle like
| providing support to your partners like Intel and AMD do.
| Pretty sure they don't want to bother with that.
|
| Before they start selling chips I would rather they open
| iMessage to other platforms to eliminate the bubble color
| discrimination.
| rafamaddd wrote:
| > Before they start selling chips I would rather they open
| iMessage to other platforms to eliminate the bubble color
| discrimination.
|
| Outside of the countries where iOS is on par with Android (I
| think US, Canada and UK are the only ones, maybe also
| Australia) in terms of popularity, I don't know or have seen
| a single person using iMessage, of course there's a lot
| people using iphone outside of the mentioned countries, but
| absolutely nobody uses iMessage.
|
| The whole discrimination of the color bubble seems to only
| happen in those countries were iOS is the same or more
| popular than android and people is actually using iMessage.
| InvaderFizz wrote:
| It's worse than that in the US. While iOS is a bit over
| 50%, it's closing in on 90% for teens[0], where such
| discrimination is most likely to occur. These numbers also
| bode well for Apple's future market share as these teens
| grow into adults.
|
| 0: https://finance.yahoo.com/news/apple-i-phone-ownership-
| among...
| ChuckNorris89 wrote:
| It's getting similar in Europe for teens. I rarely see
| them on public transport with anything other than an
| iPhone.
| rimliu wrote:
| > but absolutely nobody uses iMessage
|
| Uhm, iMessage works transparently. I just use Messages app,
| if my recipient uses iPhone it get an iMessage, if they use
| something else, they get SMS.
| mcintyre1994 wrote:
| Their point is that most people don't use the Messages
| app to communicate with others. In the UK for example
| WhatsApp is massively dominant.
| rbanffy wrote:
| > Before they start selling chips I would rather they open
| iMessage to other platforms to eliminate the bubble color
| discrimination.
|
| When so many telcos charge outrageous prices for SMSs, it's a
| useful feature.
| IgorPartola wrote:
| I agree with you right up to how exactly does the M1 chip
| affect the final UX? A different keyboard, screen, touchpad,
| etc. all make a difference but why does the chip make a
| difference?
| masklinn wrote:
| > I agree with you right up to how exactly does the M1 chip
| affect the final UX?
|
| It allows apple to focus on what they want without being
| limited by and two their hardware provider's strategy.
| ArgyleSound wrote:
| Power efficiency for one.
| IgorPartola wrote:
| Was there nobody else who made power efficient chips?
| JiNCMG wrote:
| Not in the x86 arena. Every time Apple gets involved with
| a CPU developers (Motorola, IBM, Intel) their needs
| splits from the developers desires. This time they
| decided to go on their own (well after years of doing
| this for the iPhone). Note: They have been involved in
| the ARM CPU market since the days of the Newton.
| ChuckNorris89 wrote:
| _> how exactly does the M1 chip affect the final UX?_
|
| Everything runs faster, cooler, quieter and battery lasts
| longer. Is that not part of the product UX?
| IgorPartola wrote:
| That makes it sound like Intel, AMD, ARM, etc. we're
| trying to build chips that run hotter and less
| efficiently.
| bzzzt wrote:
| Seems like Intel really lost the plan there with every
| new generation having just a few percent better
| performance, trouble with moving to smaller nodes and the
| enormous regression from spectre/meltdown.
|
| The Apple chips are made for running macOS/iOS. Seems
| there are some hardware instructions that are tailor made
| for increasing the performance of Apple software so they
| can make sure everything is working toward a common goal.
| ChuckNorris89 wrote:
| The end users don't care what brand of chip is under the
| hood, or why the UX on Apple's implementation of Intel
| chips sucked, they just know the new device has much
| better UX overall due to the more powerful and more
| efficient chip and will upgrade for that.
| tinus_hn wrote:
| > Before they start selling chips I would rather they open
| iMessage to other platforms to eliminate the bubble color
| discrimination
|
| It's probably easier to just move to one of the 99% of
| countries where nobody uses iMessage.
| ChuckNorris89 wrote:
| I already do, in Europe, where everyone and their mom uses
| Facebook's WhatsApp for everything. While that evens the
| playing field, I'm not sure I'd call trading a walled
| garden for a spyware one a massive victory though.
| tinus_hn wrote:
| So who cares that a network nobody uses exists where only
| people that have an Apple device can login?
| ChuckNorris89 wrote:
| Apparently teens and even some adults in the US where
| they'll miss out on social activities or be mocked or
| ignored due to not being on iMessage.
|
| That doesn't affect me though as i don't live in the US
| and am too old for that kind of stuff but I do remember
| how easy it was to be mocked or bullied as a teen for not
| having the same stuff as the herd, even before
| smartphones were a thing.
| NelsonMinar wrote:
| Or alternately one where some Windows / Linux manufacturer
| could match Apple for all the innovations in the M1 Macbooks.
| I'm not an Apple fan but I'm envious of what they've
| accomplished and wish I could run Windows and Linux on similar
| hardware.
|
| Other folks are starting to get there but only from the mobile
| device direction, e.g. Tensor. Maybe I should look closer at
| what Microsoft has done with ARM Surface.
| smoldesu wrote:
| It doesn't help that Apple bought the entire manufacturing
| capacity for 5nm silicon from TSCM right before the chip
| shortage hit. I think the next few years are going to get
| very competitive though, and I'm excited to see how Intel and
| AMD respond.
| phkahler wrote:
| Apple has done that before. IIRC when the original iPod
| came out it used a new generation of HDD. Apple went to the
| drive manufacturer and said "we'll take all of them" and
| they agreed.
| taf2 wrote:
| How is Amazon able to product their arm chips for aws?
| Assuming those are not the 5nm?
| smoldesu wrote:
| There's still 5nm silicon for sale, but just not at TSCM
| (the largest semiconductor manufacturer in the world).
| Companies like Samsung are just now getting around to
| mass-producing 5nm, and afaik there were a few domestic
| Chinese manufacturers who claimed to be on the node too.
|
| As for Amazon specifically though, I've got no idea.
| They're a large enough company that they could buy out an
| entire fab or foundry if they wanted, AWS makes more than
| enough money to cover the costs.
| jeffbee wrote:
| The "walled garden" comes with a C and C++ toolchain, python,
| perl, awk, sed, and a Unix shell. It is not, in any way, a
| "walled garden" in a universe where words have shared meaning.
| aftbit wrote:
| Its a walled garden when you're not allowed to leave or bring
| your friends in, no matter how nice the stuff on the inside
| is.
| jeffbee wrote:
| And that analogy applies to macOS and the M1 CPU how,
| exactly?
| rimliu wrote:
| What does it even mean?
| monocasa wrote:
| I'm hoping Alyssa Rosenzweig's fantastic work documenting the
| M1 GPU will let us write native Vulkan drivers even for MacOS.
| I believe she's been focusing thus far on the user space
| visible interfaces, so a lot of that work should translate
| well.
| snvzz wrote:
| No worries. Competition is coming.
|
| https://www.phoronix.com/scan.php?page=news_item&px=SiFive-P...
|
| Should be roughly M1 performance, but on RISC-V.
| rafamaddd wrote:
| uffff
|
| who knows when that is coming and when are we going to be
| able to buy regular laptops from e.g. Lenovo, HP, Acer, etc
| with that.
|
| By the time that happens, Apple may already be on their
| third, fourth? generation on M1. Which is going to much much
| much faster than M1.
| phkahler wrote:
| M1 is WAY faster than a cortex A78.
| marcodiego wrote:
| Now combine this with Zink and boom! We get OpenGL 4.6 for free:
| https://www.phoronix.com/scan.php?page=news_item&px=Zink-Clo... .
|
| Vulkan is too low level, but AFAICS it is not something one use
| directly, instead a library which uses it as a back-end should be
| used.
| kcb wrote:
| I've always wondered how this would work. Surely if it was
| possible to reasonably implement OpenGL 4.6 on the PI GPU it
| would already be done through Mesa.
| my123 wrote:
| > Now combine this with Zink and boom! We get OpenGL 4.6 for
| free
|
| For the RPi4 specifically:
|
| That GPU has hardware limitations that make it unable of OpenGL
| 3.0. However, it supports GLES 3.2.
|
| If you want GL desktop minus the unsupported features by the
| hardware, you can set MESA_GL_VERSION_OVERRIDE=3.3 for example.
| That will however never be compliant.
|
| Vulkan has many extensions to allow it to work on hardware
| which doesn't support the full feature set. (by not
| implementing them, instead of having only version numbers)
| zamadatix wrote:
| The Pi hardware may not support multiple render targets or
| other features in hardware directly but Zink is not required
| to (and does not always) emit 1 Vulkan API call for each
| OpenGL API call. It is free to issue as many as are needed to
| properly emulate the OpenGL API in a conformant way. That
| being said I don't think this particularly compatibility is
| in Zink today but there is nothing preventing it from being
| possible just because the hardware couldn't create the render
| targets all in one shot.
| seba_dos1 wrote:
| > but Zink is not required to (and does not always) emit 1
| Vulkan API call for each OpenGL API call
|
| The OpenGL driver also doesn't have to emit 1 logical
| hardware operation for each OpenGL API call.
| zamadatix wrote:
| There is no hard technical requirement for hardware
| drivers but it's riskier to expose performance impacting
| emulation at that level vs the layered driver level
| (where Zink is). For instance imagine a case where the
| hardware supported 4 MRTs but the hardware driver
| emulation layer exposed 8 MRTs for OpenGL compatibility
| yet Zink needed to use 16 MRTs. Now you've got all sorts
| of translation happening where Zink is likely calling the
| lower emulation layer multiple times rather than just
| calling the hardware directly. Such emulation layers are
| expected in a layered driver, that's part of their actual
| intent, whereas base hardware drivers are meant to expose
| what the hardware is able to do natively and let you work
| around it otherwise.
| seba_dos1 wrote:
| You can already enjoy stuff like OpenGL 2.1 support on
| purely GLES 2.0 hardware this way - for instance on older
| Raspberry Pis. There's not much Zink will bring on the
| table that Gallium doesn't already when it comes to
| emulation of missing hardware features (at least not if
| you want them to actually perform in any reasonable way).
| jdc wrote:
| I wonder what specifically the GPU missing that OpenGL needs.
| my123 wrote:
| The OpenGL 3.0 spec mandates support for 8 render targets,
| the RPi4 GPU only has support for 4.
| salawat wrote:
| When you say render targets, do you mean drm buffers? Or
| on GPU output buffers?
|
| I'm not quite completely clueless, but I have the feeling
| that clarification on this point will nudge me in the
| right direction to understanding these things better.
| my123 wrote:
| GL_MAX_DRAW_BUFFERS
| ArtWomb wrote:
| Congrats! Huge effort. Full spec of Broadcomm GPU (24 GFLOPS)
|
| https://forums.raspberrypi.com/viewtopic.php?t=244519
| prox wrote:
| I wonder if a Raspberry GPU board (low cost graphics performance)
| is possible. For light Blender work and maybe simple games.
| my123 wrote:
| It's better even in GPU perf/$ to buy a Jetson Nano 2GB, the
| RPi4 GPU is really small (and not that well featured).
| amelius wrote:
| But you can only run one flavor of Linux on it, since NVidia
| keeps the specs closed.
| my123 wrote:
| Today on the Jetson Nanos, you can just use the Fedora
| stock image. (flashed to a microSD card)
|
| It's much better than what it was before. nouveau works
| ootb, including reclocking too.
|
| It's also to be noted that all Tegras have an open-source
| kernel mode GPU driver (nvgpu) even when using the
| proprietary stack. However, that driver isn't in an ideal
| state today.
| numpad0 wrote:
| FP32 GFLOPS, ballparks from random sources:
|
| - this: 24
|
| - Ryzen 5600g: 200(CPU)
|
| - Jetson nano: 235
|
| - GeForce GT1030: 1127
|
| - Ryzen 3rd IGP: 2100
|
| - Apple M1X: 5200
|
| - Apple M1 Pro: 10400
|
| - RTX3080: 35580
|
| 1030 can be had for $110 even at this height of GPU
| shortages, not that much more than a Nano. hmm
| prox wrote:
| Wow that's an interesting device! Thanks!
| krallja wrote:
| The Jetson Nano uses a very similar SoC to the Nintendo
| Switch, so you can expect similar performance.
| JustFinishedBSG wrote:
| It uses half a switch SoC GPU wise
| GhettoComputers wrote:
| All versions? Would be cool to use a hacked switch
| running linux instead of Jetson if the performance was
| that much better.
| my123 wrote:
| 921.6MHz is the GPU clock on Jetson Nano (at MAXN).
|
| For the Switch:
|
| > The GPU cores are clocked at 768 MHz when the device is
| docked, and in handheld mode, fluctuating between the
| following speeds: 307.2 MHz, 384 MHz, and 460 MHz
| StreamBright wrote:
| This is great for many reasons.
|
| https://www.reddit.com/r/MachineLearning/comments/ilcw2f/p_v...
| causi wrote:
| I wonder if we'll see any impacts from this on the Pi 4
| applications that are presently borderline when it comes to
| performance, like N64 emulation.
___________________________________________________________________
(page generated 2021-10-30 23:00 UTC)