THIS KILLS THE HBM MEME

no matter how fast your computer memory is it wont make you a better person

hbm is infused with cannabis, which is why its high bandwidth memory.

HBM3 when?

if HBM was cheaper and OCed better... hnnng

Please help a brainlet. Why HBMeme like design isn't used for RAM?

fujitsu.com/global/about/resources/news/press-releases/2018/0822-02.html

>THIS KILLS THE HBM MEME
That's why HBM is featured in every HPC product worth mentioning.

Why is HBM superior?

HBM is the future of ultra-high bandwidth and capacities (Protip: They aren't for silly gayming). GDDRx is already at an dead-end with GDDR6 likely being the swan song of the standard.

It doesn't scale down easily. You can't just remove tracing and chips and call it a day. You have completely redesign the controller(s) and imposer if you want to scale down for lesser SKUs.
That's why Radeon VII and Vega 56/64 simply cannot be sold with binned "8GiB/4GiB versions". They pretty much kills HBM's viability for mainstream and lower-end SKUs.
This is the real reason why it only exists on high-end and professional-tier SKUs.

>Fujitsu
>Tofu controller
mildly chuckled

Attached: tofu.png (619x398, 171K)

I don't know the exact rules of what is/isn't HBM, but AMD's HBM cards have the memory on a die in the same package as the GPU itself, i.e. the two pieces of silicon are nestled directly next to each other. The bandwidth is pretty wild.

I wish AMD made high-end APUs with HBM for laptops.
I don't understand what's stopping them, especially now with them switching to a chiplet-based design.
A lot of laptop OEMs are already soldering RAM onto the motherboard so this would be a massive imporvement for those types of boards since they don't have to make traces for memory chips and allows them to make the boards even smaller.
Also APUs have this habit of being bandwidth-starved due to the fact that DDR4 wasn't really made with graphics processors in mind.

Attached: 1550504216523.jpg (750x784, 173K)

It's fast as fuck.

>netwrok on chip

>I wish AMD made high-end APUs with HBM for laptops.
right, nothing stops them now with chiplets design. i just wish AMD would work on power consumption/management for laptops where Intel still have a massive lead. not peak power performance, which is good (not GREAT, yet), but idle consumption.

1TB/s HBM2 only eats 40W while 672GB/s GDDR6 eats like 60W, not to mention HBM is much smaller

Attached: 155525.png (388x514, 36K)

>[ ] CUDA
Into the trash it goes

based

Raja was a mistake

How is that not the definition of a flawed design?

Infinity fabric is a packet based internal bus aswell...

Lower energy per bit, lower footprint, higher density and higher peak bandwidth all around.

>I wish AMD made high-end APUs with HBM for laptops.
No market.

>67% less bandwidth than HBM2
>100% more power consumption
>200% more time spent on PCB layout
DOA

>I don't understand what's stopping them
Cost. Also nobody'd buy it because muh Intlel brand loyalty

HBM off-die, i.e. without an interposer would defeat the whole point of using it in the first place, but it's certainly possible

*off-package

Latency would be awful

Reminder that DDR PHY is absolutely fucking enormous. A 384bit memory bus is 6 64bit PHY.
Reminder that using GDDR for a GPU means a more complex PCB.
Reminder that this DDR PHY still leaks more than an HBM PHY.
Reminder that it takes more transistors.
Reminder that it takes more die area.
Reminder that the only limiting factor to HBM's clock speeds is thermals and total package power which continually lowers as DRAM slices get put into production on smaller nodes.

Attached: 05065715663l.jpg (1358x789, 397K)

t. fragile hbm+gpu mcms