Another Opensource GPU Kickstarter

Did you ask if it will support DSUB connections and audio over DVI-HDMI converters?
 
Any previous attempt at creating an open-source GPU has been outperformed by software rendering in both features and performance. With AVX-512 likely arriving in mainstream CPUs at some point, and Xeon Phi here now, why bother trying to create your own hardware with wide vectors?

It is mind boggling that they're asking 200,000$ for 2D only and 1,000,000$ for unified shaders. I can understand their passion for hardware, but this really isn't the most cost-effective way to achieve the actual goal of running 3D applications. It's not even worthy of being called a graphics accelerator.
 
Any previous attempt at creating an open-source GPU has been outperformed by software rendering in both features and performance. With AVX-512 likely arriving in mainstream CPUs at some point, and Xeon Phi here now, why bother trying to create your own hardware with wide vectors?

It is mind boggling that they're asking 200,000$ for 2D only and 1,000,000$ for unified shaders. I can understand their passion for hardware, but this really isn't the most cost-effective way to achieve the actual goal of running 3D applications. It's not even worthy of being called a graphics accelerator.

If you are a chinese chipmaker producing bottom-of-the-barrel-minimum-order-100k-units ARM SOCs for tablets, this is probably a steal. Only the $1 million version can do GL ES 2.0, though. And power optimization is probably not on their list, at all. But for a $99 tablet, who cares? What is the licensing cost for, e.g., a Mali400 as integrated by Allwinner in their lowest-cost SOCs?
 
If you are a chinese chipmaker producing bottom-of-the-barrel-minimum-order-100k-units ARM SOCs for tablets, this is probably a steal. Only the $1 million version can do GL ES 2.0, though. And power optimization is probably not on their list, at all. But for a $99 tablet, who cares? What is the licensing cost for, e.g., a Mali400 as integrated by Allwinner in their lowest-cost SOCs?

ARM likely licenses those Mali-400s as hard macros available for fabs companies like Allwinner utilizes. Here you'd be on your own with the physical design portion. Even if you use synthesis tools I don't think it's totally trivial.

I also have serious doubts that these guys can come anywhere close to being competitive in the mobile space, starting with a 1997-era design with only a year and $1m. The GPU could be totally free for SoC makers but if it's a total joke it won't matter.

It's kind of moot though, I see no chance of them getting that $1m. In fact, I'd be surprised if they got the $200k and really, really surprised if they got the $400k.
 
Last edited by a moderator:
If you are a chinese chipmaker producing bottom-of-the-barrel-minimum-order-100k-units ARM SOCs for tablets, this is probably a steal. Only the $1 million version can do GL ES 2.0, though. And power optimization is probably not on their list, at all. But for a $99 tablet, who cares?
How much would it cost to extend NEON to 512 bit and add gather support?
 
How much would it cost to extend NEON to 512 bit and add gather support?
Do you have access to the HDL code or just the RTL or just a hard-macro? I don't know, ask ARM, the kings of low-power, but I doubt they are going that route (yet). But I sure would like that see a NEON-512 with all the nifty vector instructions. MIPS, errr, IMG said mips64 has vector-complete vector extensions, I assumed they had been reading your posts with interest this past years.;)
OT: Some time ago and inspired by your ideas I whipped out the calculator on the calxeda-SOC based products available, but they weren't there yet in FLOPS/Watt, being Cortex-A9 based and just quad-core. Maybe octo-cores and A15/A5x with FMA will be there as a possible solution for render farms. (For example: http://www.boston.co.uk/solutions/viridis/viridis-2u.aspx but with twice the cores and twice the FLOPS/clock and the same power consumption).


ARM likely licenses those Mali-400s as hard macros available for fabs companies like Allwinner utilizes. Here you'd be on your own with the physical design portion. Even if you use synthesis tools I don't think it's totally trivial.
"(Synthesized in an evening open source GPU die area minus GPU Hard-macro die area) times $/mm2" in $/die is the magic number to compare to the price of licensing an ARM/IMG/... GPU. It comes with AXI interfaces so there's a path to lowest-effort no-fscks-given ARM integration.

I also have serious doubts that these guys can come anywhere close to being competitive in the mobile space, starting with a 1997-era design with only a year and $1m. The GPU could be totally free for SoC makers but if it's a total joke it won't matter.

The unified shaders $1m project would be a new design, the 1997 compatible part is just VGA and bitblt as far as I can see. They could just design a GCN compatible core for giggles. And competitive is relative, to paraphrase Jawed's signature: "Can it play Angry Birds?" (At least Angry Birds version 1.0 used mostly bitblt (glCopySubImage or sth like that, iirc) I once traced it with the adreno profiler on my nexus one).
To be clear, I agree with "total joke" part wrt. to performance, performance/watts, features and software driver support. But I'm not sure on the $/SOC part.

It's kind of moot though, I see no chance of them getting that $1m. In fact, I'd be surprised if they got the $200k and really, really surprised if they got the $400k.

Agreed, their only change is a chinese sugardaddy willing to win the race to the bottom. IIRC Allwinner already uses an OpenRISC CPU at less than 100MHz for system task (power management, iirc), instead of licensing a Cortex-M for pennies. Googled: yep http://linux-sunxi.org/AR100

Edit: forgot: ETC1 licensing royalties. The patent is still valid, isn't it?
 
Neon aka the powervr chip ?
ps: what bit are modern gpu's it used to be a selling point "64bit graphics accelerator 128 bit ect"
 
Do you have access to the HDL code or just the RTL or just a hard-macro? I don't know, ask ARM, the kings of low-power, but I doubt they are going that route (yet). But I sure would like that see a NEON-512 with all the nifty vector instructions. MIPS, errr, IMG said mips64 has vector-complete vector extensions, I assumed they had been reading your posts with interest this past years.;)

Cheap Chinese SoCs use hard macro ARM cores. The much bigger players like Samsung and nVidia who license RTL and handle the physical implementation themselves (or through some other party) don't modify the behavior of the chip. I'm not sure if the license allows it (I've been told it's not a problem) but it's a big hit having to re-validate all that.

The unified shaders $1m project would be a new design, the 1997 compatible part is just VGA and bitblt as far as I can see. They could just design a GCN compatible core for giggles. And competitive is relative, to paraphrase Jawed's signature: "Can it play Angry Birds?" (At least Angry Birds version 1.0 used mostly bitblt (glCopySubImage or sth like that, iirc) I once traced it with the adreno profiler on my nexus one).
To be clear, I agree with "total joke" part wrt. to performance, performance/watts, features and software driver support. But I'm not sure on the $/SOC part.

The 1997 era 3D accelerator is what you get with the $400k goal. Playing Angry Birds alone is not enough for a phone chip these days. These chips aren't just a total race to the bottom where you can completely throw out 3D performance to save maybe a few dozens of cents. Allwinner, Rockchip, Amlogic, Ingenic, MediaTek, etc etc have been taking 3D a lot more seriously than that. They're also not getting ARM11s put in phones anymore.

People do care even for $99 tablets, if the competitor's $100 tablet is dramatically better.
 
The question is: will it be performance-competitive with the Riva TNT2 Ultra?

BTW, people play lots of 3D games in today's $99 tablets. Today, the "weakest" tablets have at least a Mali 400MP2 at 400MHz, or a Vivante GC1000. They are all excellent devices to play GTA3, Shadowgun, Dead Trigger, etc.
 
Theres a huge demand for non desktop, non console, non mobile/tablet graphics chips. Virtually everything is getting guis and touchscreen and cheap RTL that enables basic effects like swooshes and spinny bitmaps etc. likely has a large market.

Look at the new FTDI chip that being used in the Gameduino 2 as an example, its essentially a scan-line vector accelerator but great for the market its intended for. Its low power but also easy to integrate which for many markets is key. Your washing machine vendor probably doesn't want to have to learn shader languages and hire a 3D renderer just to give a modern tablet like interface. The HW guys probably also don't want to have to hook up a 32bit 400Mhz bus just to display a icon of your cloth spinning around...

I suspect we shall see lots of micro-gpu appearing in the next few years. They won't be designed to play the latest games but if they can produce some swish ui, maybe play games of the quake/play station era they will do well I suspect.

Deano
 
Back
Top