Pi 4 - full specification of VideoCore 6

[pi-tastic]

so its not like a Nvidia Geforce 210 Graphics card?

[jamesh]

so its not like a Nvidia Geforce 210 Graphics card?

In what way? You have to phrase your question so we can at least understand what you are getting at.

[pi-tastic]

what does performance of this GPU compare to?
what was it even made for? a cellphone?

[pi-tastic]

like back in the day the Intel Mac mini came out with GMA950 graphics, it was not a fantastic gpu but it could play World of Warcraft decently @ 1024x768 witch made it ok for many people. including people wanting a mac on a budget.

Maybe the PI4 could run WOW Classic with its GPU that would be pretty fun!

[fruitoftheloom]

what does performance of this GPU compare to?
what was it even made for? a cellphone?

VideoCore is a low-power mobile multimedia processor originally developed by Alphamosaic Ltd and now owned by Broadcom.

https://en.wikipedia.org/wiki/VideoCore

[graphicw]

It is definitely better than a Radeon 9800. That is what I have been able to determine so far. It can play Doom 3 at 720p resolution fairly competently and is likely to see a little improvement as the drivers continue to be worked on. If you want to visualize a comparison, think of it as equivalent to a gaming rig of 10 to 12 years ago. Still very competent for such an efficient system even in 2019.

[pi-tastic]

It is definitely better than a Radeon 9800. That is what I have been able to determine so far. It can play Doom 3 at 720p resolution fairly competently and is likely to see a little improvement as the drivers continue to be worked on. If you want to visualize a comparison, think of it as equivalent to a gaming rig of 10 to 12 years ago. Still very competent for such an efficient system even in 2019.

wow i remember that card!
that was the days of counter-strike

[graphicw]

Yes, it is quite an old one. It's hard to quantify the performance of the BCM2711 against PC's and dedicated graphics cards. It was designed to be low power (as in power efficient) and was not designed to be comparable to a modern gaming rig several times it's price. The best comparison I can give is of a gaming rig of anywhere from 10 to possibly even 15 years ago. Not bad considering the very low cost. It's the most realistic comparison I can think of for now. We may be in for a surprise and find it capable of running more than I have yet found it capable of running. Time will tell as the graphics drivers are worked on. One important thing to keep in mind when comparing to a PC gaming rig is that the GPU is OpenGL 2.1 compliant and that is all it ever will be. OpenGL ES is a different animal and more of a mobile specific compliance standard. OpenGL and OpenGL ES are two entirely different standards. In comparing PC graphics cards, look at those that are OpenGL 2.1 compliant and below. In making these comparisons, we are comparing a system using 15 watts or less of power against systems using 400 to 800 watts of power.

[ejolson]

This is worth a read!

https://www.raspberrypi.org/blog/vc4-an ... an-update/

Amazing! i mean, basically atm raspbian and any other software running on pi4 barely use the gpu, the amount of performance and feature will be added in next months will be massive, for what i understood!

this coupled with the 64bit kernel make me happy i choose a pi4 with 4gb of ram, over a second hand 3b+

keep up the good work devs!!!

Where are people getting the idea that the GPU is barely being used? The driver has been there, in Raspbian, since the Pi4 launch! By default, the desktop uses it! It uses the GPU!! This blog post outlines some bug fixes and improvements that have been made, making it a bit faster and more robust.

So, NO, there won't be massive improvement over the next few months.

There will be incremental improvements as the driver is optimised and bug fixed.

The ARM cores on the Pi 4B are more than 25 times faster than the original Raspberry Pi. On the other hand, the VC4 of the original Raspberry Pi was rated for 24 single-precision GFLOPS while the new one is 32. To me a 4/3-factor speed improvement appears to be very small compared to 25.

The CPU in the original Pi was similar in speed to early Intel Pentium processors which led to the statement

Overall real-world performance for Raspberry Pi 1 Model A, A+, B, B+, Raspberry Pi Zero/Zero W, and CM1 is similar to that of a 300MHz Pentium 2, but with much better graphics.

https://www.raspberrypi.org/documentati ... erformance

Note that the documentation used to say swankier graphics.

Nowadays the balance of performance in the Pi 4B is much different: The A72 CPUs seem swankier than the graphics.

[pi-tastic]

wow thanks for that info i am impressed, the best 35 bucks anyone could ever spend!

[graphicw]

Amazing! i mean, basically atm raspbian and any other software running on pi4 barely use the gpu, the amount of performance and feature will be added in next months will be massive, for what i understood!

this coupled with the 64bit kernel make me happy i choose a pi4 with 4gb of ram, over a second hand 3b+

keep up the good work devs!!!

Where are people getting the idea that the GPU is barely being used? The driver has been there, in Raspbian, since the Pi4 launch! By default, the desktop uses it! It uses the GPU!! This blog post outlines some bug fixes and improvements that have been made, making it a bit faster and more robust.

So, NO, there won't be massive improvement over the next few months.

There will be incremental improvements as the driver is optimised and bug fixed.

The ARM cores on the Pi 4B are more than 25 times faster than the original Raspberry Pi. On the other hand, the VC4 of the original Raspberry Pi was rated for 24 single-precision GFLOPS while the new one is 32. To me a 4/3-factor speed improvement appears to be very small compared to 25.

The CPU in the original Pi was similar in speed to early Intel Pentium processors which led to the statement

Overall real-world performance for Raspberry Pi 1 Model A, A+, B, B+, Raspberry Pi Zero/Zero W, and CM1 is similar to that of a 300MHz Pentium 2, but with much better graphics.

https://www.raspberrypi.org/documentati ... erformance

Note that the documentation used to say swankier graphics.

Nowadays the balance of performance in the Pi 4B is much different: The A-72 CPUs seem swankier than the graphics.

The documentation says the RPI 3 had a GPU capable of 1 Gpixel/s of compute. I can tell you the RPI 4 is capable of in the neighborhood of 2.2 to 2.5 Gpixel/s at best guess from testing. Safe to say the GPU power is at the very least two and half times greater than RPI 3 and possibly more since it is a impossible to gauge true real world performance from Gpixel/s alone. There are cases where hardware with lesser specs in those areas can beat hardware of greater specs on paper due to more efficient use of compute resources. Too many unknown variables to say for certain. I can say with certainty that the RPI 4 is at least 2.5x more capable on GPU compute than RPI 3 and probably even considerably more.

[ejolson]

I can say with certainty that the RPI 4 is at least 2.5x more capable on GPU compute than RPI 3 and probably even considerably more.

That sees like a more reasonable estimate. Mine was based only on theoretical floating point operations or second. At the same time, a 2.5-fold improvement in GPU capability between the original Pi and the present model is much less than the 25-fold improvement that happened for the CPU.

I think this is why people suspect there is hidden magic in the VC6 GPU that will be revealed in the future through an updated software driver.

Performance improvements often result from software updates, even when the original performed well. For example, improvements to the NVIDIA CUDA libraries moving from version 9.1.85 to 10.1.168 recently led to a surprising 1.77-fold performance increase for a program that I wrote on the same hardware.

[jamesh]

The way I look at it - the 3d performance on the Pi4 at 4k is about the same as the Pi3 atr 1080p· 4k is 4x the number of pixels of 1080p, which makes the GPU about 4x faster.

As for the relative speed increases of the ARM cores against the GPU, that easy to explain. We buy the CPU IP from ARM, who amortise the cost of developing new CPU';s aver a huge customer base. That's not possible with the GPU, so much less dev resource available.

[Akane]

Cannot give any schedules, Japan is a particularly long winded process requiring lots of documentation changes plus changes to the silk screening on the board. Check out Roger's blog post on the subject.

https://www.raspberrypi.org/blog/compli ... untry-yet/

Sorry for off-topic, but I want Raspberry Pi to either:

1. turn off wi-fi and bluetooth by software in some country of which use of them is restricted by law
2. add an ability to display the compliance mark on command line

1 is easy and permitted by law, as some frequency channels of wi-fi is turned off in some country by raspi-config even now.

We know that RPi4 already passed the compliance test in Japan https://twitter.com/meganetaaan/status/ ... 8270340096 , and you are just waiting for printing silk on the PCB.
If 2 is realized, then the use of RPi4 in Japan becomes legal just from now even if the compliance mark is not printed on the PCB, as the Wireless Telegraphy Acts of Japan states.

These are just my personal opinions and not to mean to force you.

[ejolson]

These are just my personal opinions and not to mean to force you.

How hard would it be to write a frontend for your QPU assembler so GPU codes could be written in a higher-level language similar to AMD HIP or NVIDIA CUDA?

[Gavinmc42]

...AMD HIP or NVIDIA CUDA

Call it OpenCL?

I think for the Pi4 Mesa OpenGL, SPIR-V, NIR, LLVM are involved.
https://www.khronos.org/spir/
OpenCL is just one part of this alphabet soup.
More is known about the QPU assembler code now as it is exposed in the Mesa source.

[jcyr]

These are just my personal opinions and not to mean to force you.

How hard would it be to write a frontend for your QPU assembler so GPU codes could be written in a higher-level language similar to AMD HIP or NVIDIA CUDA?

Hard... what you're asking for is a compiler.

[jcyr]

...AMD HIP or NVIDIA CUDA

Call it OpenCL?

I think for the Pi4 Mesa OpenGL, SPIR-V, NIR, LLVM are involved.
https://www.khronos.org/spir/
OpenCL is just one part of this alphabet soup.
More is known about the QPU assembler code now as it is exposed in the Mesa source.

More, but incomplete. Always bothered me that Broadcom doesn't release public specifications. It's not like they've got any leading edge GPU tech to hide... AMD and Nvidia are way ahead on the compute engine front.

[riccetto80]

Where are people getting the idea that the GPU is barely being used? The driver has been there, in Raspbian, since the Pi4 launch! By default, the desktop uses it! It uses the GPU!! This blog post outlines some bug fixes and improvements that have been made, making it a bit faster and more robust.

So, NO, there won't be massive improvement over the next few months.

There will be incremental improvements as the driver is optimised and bug fixed.

Well, this is what i understood, seems there is being added stuff, implemented new opengl version and adding stuff in the never implementation, not seems just bug fixing and minor improvements...

example:

At present, the V3D driver exposes OpenGL ES 3.0 and OpenGL 2.1. As I mentioned above, the VideoCore VI GPU can do OpenGL ES 3.2, but it can’t do OpenGL 3.0, so future feature work will focus on OpenGL ES.

move from opengl es 3.0 to 3.2 dont seems a minor improvement.

Also, vlc and other software wast able to use the gpu to decode videos, even chrome in the first release of raspbian didint had acceleration.
why? if the driver for the gpu was fully developed why didint work from the start?

but if you say so, i mistake.

maybe if you can explain better for noobs like me, please!

[6by9]

Shifting up versions of GL or GLES generally only exposes a few extra bells and whistles features. It doesn't add any significant extra processing performance.
It also relies on the calling app wishing to use said features. In the case of X it will generally be targetted at the lowest common platform level so that it can run on as much hardware as possible, therefore new bells and whistles are ignored.

With regard VLC and Chrome, you need to differentiate what you mean by GPU.
Generally speaking the video codecs and 3D/composition engines are totally distinct blocks with no interaction beyond the output of one being consumed by the input of the other.
The changes in the (3D) GPU drivers will have no impact on video decode. There are changes being made to make the passing of images between the two more efficient, but that is a problem in Chrome or VLC needing to support whatever acceleration APIs are supported by the platform.

[piglet]

More, but incomplete. Always bothered me that Broadcom doesn't release public specifications. It's not like they've got any leading edge GPU tech to hide... AMD and Nvidia are way ahead on the compute engine front.

I thought the thinking here was that this would run the risk of other manufacturers/ip scalpers trawling the specs/code for imagined (or real) IP infractions. You know, like "I have the IP on squares and rectangles, you owe me $bn" idiocy we see from time to time.

[jamesh]

More, but incomplete. Always bothered me that Broadcom doesn't release public specifications. It's not like they've got any leading edge GPU tech to hide... AMD and Nvidia are way ahead on the compute engine front.

I thought the thinking here was that this would run the risk of other manufacturers/ip scalpers trawling the specs/code for imagined (or real) IP infractions. You know, like "I have the IP on squares and rectangles, you owe me $bn" idiocy we see from time to time.

Quite. AFAIK, the IP for VC6 is purely Brcm, with no infringements. But that won't stop patent trolls trying, so Brcm don't give them the opportunity.

[Gavinmc42]

Yep IP craziness.
Saw MS have opened up exFAT.
Already complaints about how messy it is

If BCM bring out a Vulkan blob for VC6 would we celebrate or complain?

[jamesh]

Yep IP craziness.
Saw MS have opened up exFAT.
Already complaints about how messy it is

If BCM bring out a Vulkan blob for VC6 would we celebrate or complain?

Given past experience, complain. Because people are crazy..

[clicky]

If BCM bring out a Vulkan blob for VC6 would we celebrate or complain?

It depends whom you are including in 'we' Eh...