SK Hynix Begins HBM Production in Q1 2015 – Could Feature in AMD R9 390X/380X Fiji GPU

Author Photo
Nov 25, 2014
Share Tweet Submit

SK Hynix has officially listed HBM memory for Graphic Cards on its official catalog for Q1 2015. This is huge news, because this means that HBM memory is now readily available for GPU manufacturers such as Nvidia and AMD to utilize in their GPUs. Looks like the specs I speculated in my analysis a long time ago appear to have been true after all; regardless of the amount of flames I got for em. So far the highest SK Hynix will go for GPU HBM is a 4-Hi HBM Stack (4 Memory stacks plus 1 logic stack).

HBMA screenshot from the public catalog which can be found here.

Related AMD Radeon Pro Duo Benchmark Results Leaked – Dual Fiji GPUs Tested Against GTX 980 Ti and Fury X, 60% Performance Lead at 4K

HBM memory for next gen GPUs and AMD Pirate Islands gets listing at SK Hynix

Basically this memory is available for use both for AMD and Nvidia, however I would be willing to wager that it will appear first in Pirate Islands, specifically Fiji. On the off chance it doesn’t appear in Pirates Islands, it will probably then debut in an Nvidia GPUs. In my opinion, the chances of that happening (Green Debut) are very slim because if you recall the leak about AMD’s Fiji GPU that occurred a while back, the frequency listed was 1Ghz, which nets you 128 GB/s of bandwidth on an HBM (as opposed to an insanely low amount on GDDR5 ) – ergo, everything fits snuggly. The chips are created on the JESD235 standard and were developed in collaboration with JEDEC and AMD.

GDDR5 Memory Versus HBM Comparison Table GDDR5 2-Hi HBM 'Stacked DRAM'* 4-Hi HBM 'Stacked DRAM'*
I/O 32 512 1024
Max Bandwidth Per Pin 7 Gbps 1 Gbps 1 Gbps
Max Bandwidth 28 GBps 64 GBps 128 GBps
Voltage 1.35 - 1.65 ~1.2 ~1.2
Command Input Single DualDual
Layers 1 2 + 1 4 + 1

In this type of HBM a single logic circuit is planted directly over or under a stack of wafers. Basically, the chips are stacked and Silicon punched through TSVs (Through Silicon Vias) till the Control Die. In this case, it means that the DRAMs stacked like are controlled by a single logic circuit and henceforth referred to as a ‘Chip-on-Wafer’ design. Since we are using multiple conductive pillars punching through the structure, very low voltage is required to send signals through the stacks. As we overcome the learning curve, we can up the frequency to 2Ghz to get an effective 256GB/s bandwidth on just one 4-Hi Stack!

Related Official AMD Radeon Pro Duo Specifications Revealed – 16.38 TFLOPs Compute Card With 1.5x Performance Increase Over Titan X

The currently leaked specs of the AMD R9 390X/380X Fiji XT core stand at 4096 stream processors and a memory clock of 1250 Ghz. This means an effective bandwidth of 160 GB/s per chip. Since the specs say that it has a total memory bandwdith of 640 GB/s, we can easily deduce that the card will make use of four, 4-Hi Stacks (for a mind blowing bus width of approximately 4096 bits ?!? that’s almost over 9000). However, that is a huge number and I feel very uneasy even admitting about the possibility of a 4096 bit bus. Anyways, here you go folks, the first tangible proof that HBM is indeed on its way to GPUs. Now give me 4-Hi HBM Stack on a high end GPU with a Price Point of not more than $650 we’ll talk.

Share Tweet Submit