[TechSpot] HBM4 could double memory bandwidth to 2048-bit

[Space Lynx]

HBM4 could double memory bandwidth to 2048-bit

HBM2 doubled potential speeds to 256 GB/s per stack and maximum capacity to 8 GB. In 2018, HBM2 received a minor update called HBM2E, which further increased...

www.techspot.com

Both SK Hynix and Samsung believe they will be able to achieve a "100% yield" with HBM4 when they begin to manufacture it. Only time will tell if the reports hold water, so take the news with a grain of salt.


100% yield, impressive claims they are so confident

 

[TumbleGeorge]

Imagine devices with terabytes of HBM4E.

 

[AusWolf]

100% yield, impressive claims they are so confident

Still not enough. I want 110%!

 

[Kapone33]

HBM4 could double memory bandwidth to 2048-bit

HBM2 doubled potential speeds to 256 GB/s per stack and maximum capacity to 8 GB. In 2018, HBM2 received a minor update called HBM2E, which further increased...

www.techspot.com

Both SK Hynix and Samsung believe they will be able to achieve a "100% yield" with HBM4 when they begin to manufacture it. Only time will tell if the reports hold water, so take the news with a grain of salt.


100% yield, impressive claims they are so confident

I wonder what an RDNA card would perform like if it's memory structure was the same as Vega but HBM4? Watercooling would be a no brainer.

 

[TumbleGeorge]

In fact, in terms of overall data processing speed, there is no basis for comparison. Even if only two HBM4 chips were used in the video cards, if the article is correct, it would surpass anything GDDR7 would offer even if GDDR7 used a 512 bit wide bus. Now how much latency HBM4 will have and how much it will be able to mask from lower level caches is anyone's guess. How well would games be tuned to efficiently use the data stream input and output from HBM4?

 

[R0H1T]

Has latency ever been a major problem for GPU's as such?

 

[Wirko]

What about manufacturing costs? HBM3 requires silicon interposers, which are the most expensive means of 2.5D integration. With the pin density doubled, fabs may need to switch to interposers made at finer nodes and/or with more layers.

Imagine devices with terabytes of HBM4E.

Ha! Never again will a flagship gaming GPU be as cheap as twenty hundred dollars.

 

[TumbleGeorge]

Please don't paint money in this beautiful topic.

 

[pk67]

What about manufacturing costs? HBM3 requires silicon interposers, which are the most expensive means of 2.5D integration. With the pin density doubled, fabs may need to switch to interposers made at finer nodes and/or with more layers.


Ha! Never again will a flagship gaming GPU be as cheap as twenty hundred dollars.

They will not be used with gaming GPUs - instead they will go to AI accelerators which are used in 24/7 regime.

 

[Vya Domus]

Watercooling would be a no brainer.

HBM actually makes cooling more trivial because for one the chips themselves are running at lower frequencies than GDDR modules and because they are usually cooled by the same coldplate that cools the GPU die itself cooling is more effective anyway.

 

[Kapone33]

HBM actually makes cooling more trivial because for one the chips themselves are running at lower frequencies than GDDR modules and because they are usually cooled by the same coldplate that cools the GPU die itself cooling is more effective anyway.

Exactly, when I did that with my Vega 64 it translated to all of my Temps being the same but he GPU running in the high 40s low 50s in Gaming.

 

[Wirko]

Please don't paint money in this beautiful topic.

They will not be used with gaming GPUs - instead they will go to AI accelerators which are used in 24/7 regime.

Agreed with both, no dollares here (and it's much too early anyway).

 

[Space Lynx]

I just hope we see HBM4 on 8xxx series of AMD cards, probably won't be out for 30-40 months or so... so the timing might just be right.

 

[AusWolf]

I just hope we see HBM4 on 8xxx series of AMD cards, probably won't be out for 30-40 months or so... so the timing might just be right.

I somehow doubt it. HBM is expensive, and doesn't bring much to the table in the consumer space. If you remember previous AMD cards with HBM, they weren't faster than the competition. Besides, AMD's focus on chiplet GPUs is an entirely different direction. MCM dies with HBM chips would be too much to put into one package, imo.

 

[LabRat 891]

I'm confused...

HBM 1 on Fury was 4096-bit, and Vega 10+20s' HBM2 were 2048-bit.

 

[TumbleGeorge]

I'm confused...

HBM 1 on Fury was 4096-bit, and Vega 10+20s' HBM2 were 2048-bit.

Fury HBM chips was with capacity 1GB each and 4 of it in use. Ordinary Vega was with two chips HBM2 with 4GB capacity each.

 

[LabRat 891]

Fury HBM chips was with capacity 1GB each and 4 of it in use. Ordinary Vega was with two chips HBM2 with 4GB capacity each.

That makes sense! Thank you.

This is 'new' because it is 2048-bit to each HBMstack. Got it.