https://www.skhynix.com/eng/pr/pressReleaseView.do?seq=2086&offset=1
>With a forthcoming high-end graphics card of 384-bit I/Os, this DRAM processes up to 768GB(Gigabytes) of graphics data per second.
THANK YOU BASED JEDEC
HBM2 512GB/s 2 STACK IS FINISHED & BANKRUPT
>>60066583
>not wanting his ram on chip closer to the processor
ayy luddites.
GDDR is worse in every single technical metric than HMB besides cost, which are irrelevant in high $700 GPUs
>>60066583
>not wanting smaller lower power graphics cards
can't wait until normal vram becomes obsolete
Can HBM be mounted on a pcb without a interposer? 1 stack is gonna be 8gb soon, that's more than 8 gddr modules.
>>60066705
Price is a production metric, not a technical one.
>>60066786
Most likely not, its not really meant to be put on a PCB its meant to but up against the GPU/CPU its wired to on the same interposer.
You might be able to if you can solder a bare die that small onto a PCB but that would really ruin some of the advantages you get with HBM, latency being a big one.
>>60066786
HMC was on package, but that flopped.
>>60066786
Yes, ask Intel how EMIB werks.
>>60067112
Products with it would be a good starting point.
>>60067145
Kaby Lake-G is rumored to release around Q4 2017.
>>60066583
GDDR isn't just more power for the same bandwidth, it's also more die space. GDDR6 controllers will take up more area than even GDDR5X for the same channel widths.
GPUs are at the point again where additional logic can be more useful than raw bandwidth at increasing performance (tiling/binning rasterization, color compression, native fp16, etc.), so it's not exactly a pure win approach.
HBM is better from the GPU die perspective in every way, it's just the lingering high costs there that suck.
>>60067427
And all those GDDR dies take up more space on the card, more circuitry on the PCB, more delay going through the PCB.
768Gbps of GDDR6 is probably not even going to be equal to 512Gbps of HBM2.
At that point, the bandwidth is sufficient but what you really need is power efficiency and low latency past that point.
Also HB3 and.. HBC I think it's called are also coming.
>>60069335
jesus, so much wrong here...
> GDDR dies
I think you mean controllers, or perhaps PHYs.
> delay going through the PCB
PCB delay isn't an appreciable factor in memory speeds
> HBM3
coming ca. 2020
> 768 GB/s < 512 GB/s
I hope this is B8. Bandwidth is bandwidth, since both systems are half-duplex and have comparable latencies and effective duty cycles (~75% theoretical bandwidth achievable).
> HMC = Hybrid Memory Cube
pretty much DOA. it's designed for density and overall capacity, not power or bandwidth, so it was never a great choice for GPUs to begin with.
>>60069727
No I mean the dies.
The controller is on the GPU die itself. The SGRAM dies are on the PCB.
>PCB delay isn't an appreciable factor in memory speeds
You're right there. It's probably less than a tenth of a nanosecond. But HBM2 still has less latency as a total package.
>>60070067
>But HBM2 still has less latency as a total package.
not really.
latencies are still dominated by typical DRAM access costs.