Friday, July 15th 2016

SK Hynix to Ship HBM2 Memory by Q3-2016

Korean memory and NAND flash giant SK Hynix announced that it will have HBM2 memory ready for order within Q3-2016 (July-September). The company will ship 4 gigabyte HBM2 stacks in the 4 Hi-stack (4-die stack) form-factor, in two speeds - 2.00 Gbps (256 GB/s per stack), bearing model number H5VR32ESM4H-20C; and 1.60 Gbps (204 GB/s per stack), bearing model number H5VR32ESM4H-12C. With four such stacks, graphics cards over a 4096-bit HBM2 interface, graphics cards with 16 GB of total memory can be built.

Source: SK Hynix Q3 Catalog
Add your own comment

77 Comments on SK Hynix to Ship HBM2 Memory by Q3-2016

#1
P4-630
The Way It's Meant to be Played
So a Titan will come with 16GB HBM2.
Posted on Reply
#2
Chaitanya
Is this the reason for delay of flagship GPUs from either of the two manufacturers?
Posted on Reply
#3
TheGuruStud
Chaitanya said:
Is this the reason for delay of flagship GPUs from either of the two manufacturers?
That and yields *cough* nvidia basically paper launched 10xx series *cough*

No doubt there's tons of lost dies on GP102.
Posted on Reply
#4
Prima.Vera
Meaning we wont see the flagships in stores until next year most likely...
Posted on Reply
#6
laszlo
this mean amd will rush to buy&use otherwise nv will have access sooner than them...
Posted on Reply
#7
Nokiron
P4-630 said:
So a Titan will come with 16GB HBM2.
laszlo said:
this mean amd will rush to buy&use otherwise nv will have access sooner than them...
Remember that Nvidia showed Tesla P100 with HBM2 sourced from Samsung, a company that has mass-produced HBM2 since January.

It's probably still rare to get hold of, but Samsung is way ahead SK Hynix on this one.
Posted on Reply
#8
ZeppMan217
laszlo said:
this mean amd will rush to buy&use otherwise nv will have access sooner than them...
They could bide their time, let Nvidia roll out their $1500+ Titan Pee, and then smash them with something sub $1000 yet nearly as powerful.
Posted on Reply
#9
Aquinus
Resident Wat-man
...am I the only person who would be more interested in seeing this on a CPU?
Posted on Reply
#10
Jermelescu
Aquinus said:
...am I the only person who would be more interested in seeing this on a CPU?
I thought about it for 15 secs or so, but I can't find any reason why would someone do that. Or are you talking about APUs?
Posted on Reply
#11
P4-630
The Way It's Meant to be Played
Aquinus said:
...am I the only person who would be more interested in seeing this on a CPU?
Instead of system ram and gone with the memory slots on the motherboard?
Posted on Reply
#12
Aquinus
Resident Wat-man
Jermelescu said:
I thought about it for 15 secs or so, but I can't find any reason why would someone do that. Or are you talking about APUs?
Either. Eliminating the need for external memory has huge space saving benefits. Having memory closer to the CPU also will reduce latency compared to off-chip memory. To me, that's the natural evolution of the SoC.
P4-630 said:
Instead of system ram and gone with the memory slots on the motherboard?
Sounds great for SoCs and low profile systems likes HTPCs and laptops where space is at a premium.

Edit: Consider something like an Intel Xeon D-1567 with on-board extended memory. You could fit such a thing on a motherboard smaller than Mini-ITX. You probably could do it with a motherboard that's twice the size of a Raspberry Pi because you're eliminating the need to run any DDR3/4 wires which adds quite a bit of space and complexity to the motherboard. Moving this complexity to an interposer simplifies a lot of things on the motherboard and for the connection between CPU and mothgerboard. It also would reduce the number contacts on the CPU by a significant amount considering modern DDR is still utilizing parallel data and address buses which are bulky. If you consider that maybe 200-250 of those 288 pins on a DDR4 module go to the CPU, you have two channels so that's 400-500 contacts for DRAM alone. You take a skt2011(or -3) CPU and quad-channel memory brings that number up to a whopping 800-1000 pins for DRAM alone. People might not realize it but I/O takes up most of the pins of microprocessors these days when I/O is external to the processor.
Posted on Reply
#13
Caring1
Nokiron said:
Remember that Nvidia showed Tesla P100 with HBM2 sourced from Samsung, a company that has mass-produced HBM2 since January.

It's probably still rare to get hold of, but Samsung is way ahead SK Hynix on this one.
This leads me to believe nVidia will continue to use Samsung as their source, and AMD may have two flagship cards coming due to the two differing types of memory being made.
Posted on Reply
#14
Nokiron
Caring1 said:
This leads me to believe nVidia will continue to use Samsung as their source, and AMD may have two flagship cards coming due to the two differing types of memory being made.
I agree, especially since AMD has priority access to SK Hynix HBM2.
Posted on Reply
#15
Captain_Tom
TheGuruStud said:
That and yields *cough* nvidia basically paper launched 10xx series *cough*

No doubt there's tons of lost dies on GP102.
Exactly.

1) It is obvious AMD needs better yields considering undelclocked mid-range card are all they can produce in large volumes right now.

2) Nvidia on the other hand is just doing paper launches of paper launches (literally) so they can make it look like they have actually released cards.
Posted on Reply
#16
Fluffmeister
Meh, Pascal cards are in stock and widely available here.
Posted on Reply
#17
Dj-ElectriC
Captain_Tom said:
Exactly.

1) It is obvious AMD needs better yields considering undelclocked mid-range card are all they can produce in large volumes right now.

2) Nvidia on the other hand is just doing paper launches of paper launches (literally) so they can make it look like they have actually released cards.
GTX 1070s and GTX 1080s are in stock in most places. GTX 1060s will be in stock next week in most places.
Posted on Reply
#18
Nokiron
Captain_Tom said:
Exactly.

1) It is obvious AMD needs better yields considering undelclocked mid-range card are all they can produce in large volumes right now.

2) Nvidia on the other hand is just doing paper launches of paper launches (literally) so they can make it look like they have actually released cards.
Well, that really depends on the market. In the nordics, retailers are selling ten GTX 1080 for every RX 480.
Posted on Reply
#19
Captain_Tom
Nokiron said:
Well, that really depends on the market. In the nordics, retailers are selling ten GTX 1080 for every RX 480.
Link?
Posted on Reply
#20
RejZoR
Imagine AMD's APU with CPU+GPU+VRAM+RAM crammed together on same interposer. I think that's what AMD is actually targeting for their APU's. This would bring a new era for laptops. Imagine all the extra space left for other stuff like bigger battery, better cooling etc.
Posted on Reply
#21
laszlo
RejZoR said:
Imagine AMD's APU with CPU+GPU+VRAM+RAM crammed together on same interposer. I think that's what AMD is actually targeting for their APU's. This would bring a new era for laptops. Imagine all the extra space left for other stuff like bigger battery, better cooling etc.
why not imagine a single chip(not like now close&connected to each other) which have all that a current pc need to run (( mb with cpu+gpu+storage+ext. connections+audio(don't know if needed as gpu will have it)+......all the rest )) for sure will be made but maybe not in my lifetime....
Posted on Reply
#22
Casecutter
Aquinus said:
...am I the only person who would be more interested in seeing this on a CPU?
I though the consoles would be the brightest place a APU with HMB on an interposer.

Fluffmeister said:
Meh, Pascal cards are in stock and widely available here.
Only place I see good amounts of "in stock" is England, In the USA it seem to gotten better there's some 1070 EVGA, Gigabyte, MSI, while it looks like the Zotac boat came-in if your into those... Most nicer customs (EVGA, Gigabyte, MSI) are like $460, while some perhaps some lower grades like the "GIGABYTE GeForce GTX 1070 DirectX 12 GV-N1070G1 GAMING-8GD 8GB" at $430. There's a "ZOTAC GeForce GTX 1070 AMP! Edition, ZT-P10700C-10P, 8GB GDDR5 IceStorm Cooling, Metal Wraparound Carbon ExoArmor exterior, Ultra-wide 100mm Fans, Spectra Lighting, PowerBoost, FREEZE Fan Stop"... that's a month-full version is $440.

Captain_Tom said:
Exactly.

1) It is obvious AMD needs better yields considering undelclocked mid-range card are all they can produce in large volumes right now.

2) Nvidia on the other hand is just doing paper launches of paper launches (literally) so they can make it look like they have actually released cards.
Honestly the whole FAB situation is better for gamers, and while hard to watch AMD/GloFo get something going, consider what would be happening if TSMC was still the only game in town? Every company is vying to get their production in the schedule at TSMC. What if more folks (those at GloFo/Samsung) where trying to get 16nmFinFet wafers from them? TSMC would be charging whatever they want (which I'm sure they're doing now) and the lead-times longer. The situation is ugly... as face it the market was done with buying 28nm graphic cards. On the other hand something had to get moving even if it not perfect, or just a dribble. The market has been constipated for to long, this is at least that feeling thing are at least about to dump.
Posted on Reply
#23
Caring1
Aquinus said:
...am I the only person who would be more interested in seeing this on a CPU?
It would be interesting to see what affect it has on die size, heat and other factors, replacing eDRAM with HBM2 on Die could be done, but it would be costly.
Posted on Reply
#24
ShurikN
RejZoR said:
Imagine AMD's APU with CPU+GPU+VRAM+RAM crammed together on same interposer. I think that's what AMD is actually targeting for their APU's. This would bring a new era for laptops. Imagine all the extra space left for other stuff like bigger battery, better cooling etc.
Been talking about this from the first second HBM was anounced by AMD for Fury.
Posted on Reply
#25
Loosenut
Aquinus said:
...am I the only person who would be more interested in seeing this on a CPU?
I do see this as a possibility for APU's in consoles though
Posted on Reply
Add your own comment