With integrated gpu's becoming better and better on the processing end, the bottleneck is and has always been system ram to use as VRAM.
Is there any reason why we cannot have DRAM manufacturers make a GDDR5 module that fits right into the PCIE 16x slot?
If you guys have any better solutions to the IGPU bandwidth issue I'd like to know your thoughts.
there's no reason we have to live with modern gpu processing with ten year old vram bandwidth.
i have like ten of them.....that's not my issue here.
There's no reason to buy a GPU if you only use an office PC
>Is there any reason why we cannot have DRAM manufacturers make a GDDR5 module that fits right into the PCIE 16x slot?
You dont understand what latency or bandwidth is do you?
Just get a dedicated gpu at that point. Are you literally retarded?
i know full well what latency and bandwidth are.
if you're asking me on a technicality whether or not we can have gddr5 out of a pcie 16x slot then you need to address those issues yourself. from my lane address understanding there's absolutely no reason why this is not feasible.
if you think you're being cute and smart as to say hurr muh cpu will take all day to address vram in a pcie 16x slot you're kidding yourself.
>
nigger did you not read where i said have a fuck ton of gpu's?
this thread is about pajeet tier theoretical ideas of how we can get igpu's to get more bandwidth, not about how you just built your first pc last year.
Because a pci 3.0 x 16 and ddr3-1600 memory have comparable bandwidth. It doesnt matter if the gddr5 is faster it will be limited by the pci slot. So why just not use the system ram. Intel could put more L4 cache on their chips but that makes it more expensive. AMD does plan to release an APU with hbm directly on the die.
good, someone answered my question.
pcie 3 16x max is 32 GB/sec
now we're talking. someone knows their shit.
that's still twice as much than typical 10-15 GB/sec ram.
so the sooner we can get a pcie 4 standardized slot we can push 64 GB/sec. if we're not gonna see DDR4 gain bandwidth any faster we need to consider things like this.
why do you guys think M.2 is a thing? using PCIE bus instead of SATA.
Remember that the igpu also has a limit. It will only be able to effectively make use of so much memory. Professional applications it might make sense. AMD released a gpu with a m.2 slot so you could directly load data from an ssd. Its not exactly using it for calculation but fast storage which is neat.
now we're thinking instead of 'buy a titan x yo'.
i know how expensive dedicated gpu's are in laptops and in a few years i'd like to see standardized ways to see laptops have dedicated vram without having dedicated gpu's, something cheaper to work with. a middle ground solution as IGPU/APU get better and better and become more and more bottlenecked by vram.
current IGPU's are significantly bottlenecked by vram and this will only hinder the further progress. what good is a geforce 950m equivalent if it can only move 15 GB/sec instead of the 80 GB/sec the dedicated version could move?
marketing solutions to sell 'gaming capable' machines that don't cost 1000-2000 dollars will become a thing. instead of 4-500 dollar machines that can't play crysis yet have "i7" tier cpu's, we can have machines with i3 and i5 tier cpu's that can actually play metro 2033 at 1366x768 instead.
just have 128MB of L4 cache and quad channel memory controllers, that's more than enough bandwidth for anything an iGPU could ever handle.
i was thinking about that after seeing the huge cache on the new Zen chips.
but it's gotta be cost effective cache. if the market can bear it, that would be a great solution for the next gen of APU's.
It's been done before, this is a 4MB AGP VRAM card
there's no reason it can't be done again then.
we're a long way's away from AGP though.
>but it's gotta be cost effective cache
the Intel chips that have been in the 15" MacBook Pro have had 128MB of L3 cache since Haswell so it doesn't seem like it's that expensive. The Xbox One has 32MB of SRAM on-chip and it was pretty cheap even at launch. Xbone also has a quad channel memory controller making even it's old DDR3 pretty fast.
I wonder how it would work if motherboards somehow could have 1-2GB of ddr5 or hbm of some sort integrated for the igpu to work
That could make APUs actually viable
lmao how many anti static bags have you been sniffing
The chips with the iris graphics at the moment are way more expensive than their regular integrated graphics.
expensive due to manufacturing costs or due to lack of competition?
You can't seriously think PCI-E has the performance of the DIMM slots connected to the IMC, can you?
i already noted where it has 32 GB/sec.
did you even read?
now this guy gets it.
none
no shit? that's not in question here.
Bandwidth is not the only performance metric for computer I/O.
no shit? but it's the limiting factor in an IGPU
Go ask retarded questions somewhere else.
like Sup Forums
>Is there any reason why we cannot have DRAM manufacturers make a GDDR5 module that fits right into the PCIE 16x slot?
That's be even slower than right now user...
Dual channel DDR4-2400 (pretty mediocre/average now) is ~38GB/s.
PCI-e 3 at 16x is only 15.75GB/s.
>If you guys have any better solutions to the IGPU bandwidth issue I'd like to know your thoughts.
AMD and Intel are already way ahead of you.
AMD will (in Late2017/2018, with Zen APUs) have HBM for the iGPU.
Intel has eDRAM cache for igpu which speeds up a small portion of the ram - and will likely up thier game in the future to compete with AMD.
my data was bad then, thought 16x performance was double the 16 GB/sec.
dual channel theoreticals are that high but real world throughput isn't that high due to how dual and quad channel works.
wait for zen+hbm
Pretty much this.
There isn't any way out of it user (well Intel will probably use HMC, but it's the same deal)
PCI-e adds too much latency and doesn't have the bandwidth until PCI-e 4.0 (and by then HBM2 will be mass production available, HBM1 will be cheap and HMC will likely be in sampling).
The current lot of iGPU's aren't being held back by memory, they're being held back by being low spec (which can't be addressed until faster memory is available)
Bristol Ridge (AMD) only has 8 ROPs @ 1108mhz* for ~9.2GP/s
GT4e (Intel) only has 9 ROPs @ 1150mhz* for 10.3GP/s
>*best stock clocks
A 9600GT does 10.4GP/s... A midrange GPU from 2009 is beating them and it
The PS3 RSX (GeForce 7800GS-ish) did 8.8GP/s...