how is it a sub par GPU given it targets a specific segment (looking at it's price point, die area and memory & power envelope) with its configuration? You're upset that they didn't aim for a halo GPU and I can understand that, but how does this completely rule out a mid to high end offering from them? the 9000 series is reminiscent of nv10 versus vega10 GPUs like the 56, 64, even the Radeon 7; achieving _far_ more performance for less power and hardware.

> Nvidia does not care about the ISA of the CPU at all. That’s kinda my point. They’re stuck communicating over PCI-E instead of being a first-class co-processor over AMBA.

Sure, but do you need a discrete video card if you’re gaming on an ARM SoC? And we’ve seen from the struggles of x86 iGPUs that graphics APIs pretty much have to choose whether they’re going to optimize for dedicated VRAM or shared memory, cuz it has inescapable implications for how you structure a game engine. ARM APIs will probably continue optimizing for shared memory, so PCI-E GPUs will always be second-class citizens.

I've been able to use cuda accelerated cycles rendering in blender with my nearly 12 year old gt 750 for a decade. We aren't even talking RT cores, though they still have a solid lead yes. AMD didn't get the capability till basically the covid chip shortage and crypto bubble. When everything was unobtanium. Likewise, go talk to anyone that edits video semi professionally. Accelerated timeline rendering via cuda in premiere was massive. AMD and now Intel are supporting both finally. But are only roughly a decade late. And software is still maturing for them. I'm looking at upgrading to a battle mage card since they can support my workflows. Gaming and 3d modeling/raytracing. 2 years ago that wouldn't have been a possibility. Nvidia made a massively good investment and position with cuda.

Yes, even for applications other than gaming. There are legitimate mad lads out there running steam games with discrete video cards on Raspberry Pi's and LLMs. Not to mention there are non soc arm machines. And soc Intel machines. Sometimes getting the integrated graphics on any of these SOCs working is a much harder prospect than getting a discrete one.

Data centers are already paid for, so they're being built. But if they can't go online due to power costs. Then that will burst the bubble. As for AI use... Sadly there are a bunch of people using it. And while the drop off rate of people trying it and ditching it is steep, there's actually a readoption curve. Which never fucking happens. So everyone is betting on the next model being better, and more people giving it all a second chance... Which are two open questions. But no power means no new model and no readoptions. This no profit. Those other steps can fail, but without power it all fails.

Again I don't know anyone in that space, nor do I know numbers on gamers vs pros, so I cannot comment further.

Ok doomer

You need to do you, you /S

They are in cahoots with the RAM cartels to push gaming onto their cloud services so that competitors like AMD don't just pick them up. Trying to make everything into a service is just a side benefit, although I'm sure they realize 16 bit SNES games are still fun and that people will just be driven to less powerful entertainment platforms.

>AMD didn't make a 5090 equivalent so I won't buy their mid-tier card Is there a name for this thinking?

> Data centers are already paid for, so they’re being built. no they are not… they are contracted in paper by openAI, for example, who has no way of paying for them other than "trust me bro" > But if they can’t go online due to power costs it's not just the cost… the infrastructure to produce all this additional power does not exist… another issue with these massive bubble >As for AI use… Sadly there are a bunch of people using it. And while the drop off rate of people trying it and ditching it is steep, there’s actually a readoption curve. Which never fucking happens. Served by the current infra, why do we need the next 200 gwatt for? to make memes faster? >So everyone is betting on the next model being better, and more people giving it all a second chance… Which are two open questions. by everybody you mean those ped.ling the bubble… we already saw the decline in the newer models and know it´s matematically impossible to get rid of the slop or get to "gen ai" through LLMs

Friend plugged his screen into the onboard graphics card output and hasn't figured out how to setup pass-through.

You need a discrete video card on ARM exactly as much as you do on x86. The GPU is independent of the CPU architecture, so if you don't like x86 iGPU performance expect to not like ARM iGPU performance either. Of course you can dump a full desktop dGPU into the same package as your CPU, like Apple did with the M4 Max, but again this can be done on x86 as well. None of that is dependent on the ISA.