/tech/ - Technology★

>>830508 (OP)

>What is preventing this from being a thing?

we don't have AGP's 6MB/s main memory bottleneck any more so there's no point

Graphics memory is a lot more timing-sensitive than main memory

The Vega graphic cards allows to use the computer's RAM in the case the 8GB are insufficient. That's not perfect, but it's a good idea if you have a lot.

>>830616

Literally the same case with Intel iGPUs and AMD APUs released for the past 5 years at least

Part of the reason integrated graphics are so shit is because generic SDRAM is slower for graphical workloads than GDDR memory

>>830626

Also, we're stuck using PCIe, which has to squeeze through the on-die RAM controllers in each CPU. We should just have one single bus (HyperTransport, RapidIO, Infiniband, etc.) used for everything (CPUs, RAM, storage, expansion cards, external I/O) with a unified fabric topology.

>>830633

That's what AMD is trying to do with HSA I think

>>830646

Yeah, it just pisses me off this didn't happen yonks ago. All of the buses I mentioned were ready back in the early 2000s (before inferior garbage like AGP 8X, PCI-X, PCIe, SATA, GBe, USB 2, etc., which you'll notice were mostly Intel's doing), and were already implemented in CPUs & mobos (particularly AMD's use of HyperTransport), so they could've just neatly phased it in as the replacement for legacy standards. SGI had even perfectly demonstrated exactly how such an architecture should be used, with NUMAflex on their awesome modular Origin/Onyx systems.

Replaceable memory only makes sense when manufacturing process of said memory has certain fault probability. Back in 90's and early 2000's memory was too expensive to produce, therefore they gave users an option to buy modules separately, today 8GiB of ram cost $20 for board manufacturer and is stable aff, be it graphics card or fagpple notebook, they solder it down and it works perfectly till next planned obsolescence model rotation.

>>830658

The difference between the 3GB & 6GB GTX 1060 is still about $50

>>830580

I think this anon is probably right, latency is yucky and user control harms latency.

>>830677

Extra dosh probably pays for VHDL monkeys, kernel contributors and Shareblue-tier "underdog" marketing on forums.

>>830586

The physical space wasted on those slots and trace routing for them could be used to add another 8GB nowadays. If you have to ask idiotic questions like the OP is doing you're not doing anything that needs that much anyway.

>>830895

>implying PCB footprint is a factor in the amount of VRAM on GPUs

>>831030

Only a few years ago one geforce card came out that tried to use the max allowed ATX card size (~1ft) and people found out the hard way their standard ATX cases actually weren't

bump

>>831064

Pfffffffffffft

>>830577

>AGP's 6MB/s

You're showing your age anon... or you're just completely retarded.

AGP 1.0 266MB/s

AGP 3.0 2133MB/s

There's no point to it.