Hbm 2 gets another upgrade-HBM2E

[cj09beira]

Sk hynix has just anounced a new hbm stack that will give up to 460GB/s per stack capacity wise it will offer up to 16GB, with mass production starting in 2020, this means sk hyninx might offer the fastest gbm offer in 2020 as samsung's HBM2E offerings which were announced back in march only go up to 410GB/s (also up to 16GB per stack) though without any timeline as to when it will be available 

Quote

Back in March, Samsung announced HBM2E memory stacks which offer 4.1GB/s of memory bandwidth, placing them behind SK Hynix's latest offerings. This memory was already 2x as fast as the HBM2 memory used on AMD's RX Vega 56, a feat which makes SK Hynix's HBM2E offerings appear all the more impressive

SK Hynix targets mass production of its HBM2E memory in 2020. This is when SK Hynix believe markets for HBM2E memory will to open up. SK Hynix expects customers from the GPU market, creators of machine learning accelerators and other AI chip makers. 

 

 

opinion:

Its good to see hbm getting some love specially now with actual timelines as to when we can see it used in new products, in our beloved gpus i wonder how they will be used as last time the fastest hbm chips were not used as they didn't provide enough bandwidth to either improve performance at the same stack count (enough bandwidth already) or enough performance with less stacks, with 460GB/s being fairly close to the objectives of hbm 3 a card like the radeon vii could achieve the same levels of performance with just 2 stacks of hbm, its also enough bandwidth to power a mid range class gpu with a single stack.

 

Source

[Sakkura]

Cool.

Quote

Using SK Hynix' new HBM2E memory, a Radeon VII like graphics card would offer 1840GB/s of bandwidth

 

[LukeSavenije]

well... that'll cost another kidney

[Sakkura]

2 minutes ago, LukeSavenije said:

well... that'll cost another kidney

If you can start measuring your GPU memory bandwidth in TB/s.... worth it.

[LukeSavenije]

1 minute ago, Sakkura said:

If you can start measuring your GPU memory bandwidth in TB/s.... worth it.

until then I'll have to look at my pesky 9ghz gddr5

[Taf the Ghost]

Interesting. Looks like HBM3 is still a good ways off.

[Princess Luna]

HBM2E instead of HBM2X? what world are we living in!!!1

[AlTech]

44 minutes ago, LukeSavenije said:

well... that'll cost another kidney

Actually unless I misunderstood this, it means some GPUs could be cheaper to make with HBM2E than HBM2 in situations where you need loads of VRAM but don't care about the speed too much. E.g. 16GB 2048 bit bus HBM2E vs 16GB 4096 bit bus HBM2. The HBM2E option would be cheaper in that scenario.

[MeatFeastMan]

Looks like Navi with HBM2E confirmed. Well, maybe..hopefully. Mouth is officially watering. 5950XT dual-gpu behemoth with 32gb HBM2E...come on AMD YOU KNOW YOU WANT TO!

 

We can only hope..can't come soon enough.

 

[Doobeedoo]

Interesting this may be for big Navi I's expect. 

[Humbug]

All set for big Navi lol. Probably $1000 GPU which I will admire on YouTube.

[rawrdaysgoby]

I am sure Nvda can go to the next node anytime they want and go with hmb2e for their next high end.... but will cost 1500$

[RejZoR]

HBM turned out to be a Rambus of graphics. Theoretically fast but essentially pointless as it's endlessly short in supply and stupid expensive. Maybe in 5 to 10 years time it might become a mainstream on high end cards. Anything below will remain on whatever version of GDDR.

[79wjd]

11 hours ago, MeatFeastMan said:

Looks like Navi with HBM2E confirmed. Well, maybe..hopefully. Mouth is officially watering. 5950XT dual-gpu behemoth with 32gb HBM2E...come on AMD YOU KNOW YOU WANT TO!

 

We can only hope..can't come soon enough.

 

I guess you want Navi to be DOA then.

 

HBM tends to be a really poor choice in the consumer GPU space when the aim is profitability.

[leadeater]

13 minutes ago, 79wjd said:

I guess you want Navi to be DOA then.

 

HBM tends to be a really poor choice in the consumer GPU space when the aim is profitability.

Better to get the engineers to work on better memory compression than increase part supply cost, package cost and reduce supply channels.

[CarlBar]

8 hours ago, RejZoR said:

HBM turned out to be a Rambus of graphics. Theoretically fast but essentially pointless as it's endlessly short in supply and stupid expensive. Maybe in 5 to 10 years time it might become a mainstream on high end cards. Anything below will remain on whatever version of GDDR.

 

I've pointed this out before tbut the big thing with HBM was that no one saw GDDR6 and to a lesser degree GDDR5x coming. So if you wanted high capacity and/or high speed memory your only expected option after hitting the limits of GDDR5, (which was imminent), was HBM. NVIDIA even used it on some of their previous generation flagship pro cards.

 

Technically on the capacity front it's still true, Assuming the memory controller can address the needed capacity you could build Radeon Instinct cards using these with upto 64Gb of capacity per GPU die, and whilst AMD would have to heavily modify the silicon, (which some customers might ask them for), there's space to squeeze up to 12 modules around each GPU die.thats 194Gb of capacity. GDDR6 is never going to be able to remotely touch that. But right now there's not a lot of demands for that much capacity either.

 

But either way the significant speed and capacity enhancements that have come along recently took a lot of the early wind out of HBM's sails which meant many had no reason to use it which pushed prices up which killed it as a mainstream product. Thats not a reflection on weather HBM is any good, just a reflection on mistiming, sooner or later GDDR is going to run out of growth room.

 

Rambus on the other hand was allways competing against an alternative standard that could keep pace with it long term.

[cj09beira]

10 hours ago, RejZoR said:

HBM turned out to be a Rambus of graphics. Theoretically fast but essentially pointless as it's endlessly short in supply and stupid expensive. Maybe in 5 to 10 years time it might become a mainstream on high end cards. Anything below will remain on whatever version of GDDR.

packaging tech needs to catch up thats it, we need embedded interposers so that hbm packaging costs go down, when that is ready hbm will be more appealing (which will come in handy for all other chiplet applications, there is a good chance that the next amd chiplet architecture will need it as they are already at the limits of organic substrates)

amd at least for the high end will probably stick to hbm, as its better, the hbm prices we know about are also all from the time of memory price fixing i wonder if steve can update us on that

[Sakkura]

The price ratio between HBM and GDDR is probably pretty similar today, BUT spending X% more on memory matters a lot less when memory in general costs half as much. So HBM is significantly more viable than it used to be. The combination of Vega and HBM2 just launched at a bad time for AMD.

[Unixsystem]

On 8/12/2019 at 6:55 PM, RejZoR said:

HBM turned out to be a Rambus of graphics. Theoretically fast but essentially pointless as it's endlessly short in supply and stupid expensive. Maybe in 5 to 10 years time it might become a mainstream on high end cards. Anything below will remain on whatever version of GDDR.

It definitely feels like a tech that came out a decade too early for widespread adoption. I'm mostly interested to see not whether or not it'll eventually gain widespread GPU adoption, and more if it'll get picked up for some sort of next level CPU caching system. 

[cj09beira]

5 hours ago, Waffles13 said:

It definitely feels like a tech that came out a decade too early for widespread adoption. I'm mostly interested to see not whether or not it'll eventually gain widespread GPU adoption, and more if it'll get picked up for some sort of next level CPU caching system. 

with intel going feverous and amd needing to keep up, things will be very interesting for the next decade

[Trik'Stari]

Don't tell AMD.

 

They'll spend even more time dicking around and continuing to fail to compete with Nvidia.

[Colonel_Gerdauf]

23 hours ago, Trik'Stari said:

Don't tell AMD.

 

They'll spend even more time dicking around and continuing to fail to compete with Nvidia.

When hasn't it? Not even the 200 series were particularly good, it only stuck around because of people blindly rooting for the underdog.

[thorhammerz]

1 hour ago, Colonel_Gerdauf said:

When hasn't it? Not even the 200 series were particularly good, it only stuck around because of people blindly rooting for the underdog.

Catering to tribalism is just part of the fun .

[DeScruff]

On 8/15/2019 at 6:40 PM, Colonel_Gerdauf said:

When hasn't it? Not even the 200 series were particularly good, it only stuck around because of people blindly rooting for the underdog.

Wasn't it because they were good for mining crypto currency but then the first Bitcoin bubble burst, and all the sudden you could get things like the R9 270x for $150?

(And if I recall the R9 270x was better, and cheaper then the sorta new 750ti which ranged from like $150-200 at that time.)

[Derangel]

On 8/15/2019 at 5:40 PM, Colonel_Gerdauf said:

When hasn't it? Not even the 200 series were particularly good, it only stuck around because of people blindly rooting for the underdog.

Outside of heat and power the 200-series cards were excellent. They stood up pretty well against Nvidia's 700 series and ended up being MUCH better long term purchases. Before that, I'm guessing you aren't old enough to remember the days when ATI/AMD and Nvidia were trading blows on a pretty consistent basis. ATI crushed Nvidia's FX series and the 5000-7000 cards were very competitive.