SK Hynix 4 Gigabyte HBM2 Stack Availability Paves the way for Vega

[btarunr]

SK Hynix, the principal supplier of high bandwidth memory (HBM) stacks to AMD since its pioneering Radeon R9 Fury series, released an updated HBM2 specifications document that pins availability of 4 gigabyte (32 Gbit) HBM2 stacks to the first quarter of 2017. This paves the way for mass-production and market availability of AMD's upcoming Radeon "Vega" graphics chip, which feature two such 4 GB HBM2 stacks, making up 8 GB of total memory.

SK Hynix H5VR32ESM4H-H1K memory stack, which will be integrated by AMD onto its Vega10 multi-chip module (MCM), ticks at 1.60 Gbps (per pin), with a cumulative bandwidth of 204.8 GB/s per stack. With two such stacks, Vega10 could feature 409.6 GB/s of memory bandwidth, assuming AMD uses the reference clock speeds for these stacks.



View at TechPowerUp Main Site

 

[thesmokingman]

Wow!

 

[cdawall]

MCM referencing the memory and gpu or dual GPU's?

 

[Patriot]

MCM referencing the memory and gpu or dual GPU's?

Referencing the interposer with gpu and 2 stacks of hbm2 on top.

 

[cdawall]

Referencing the interposer with gpu and 2 stacks of hbm2 on top.

Thanks you. I was about to be mildly pissed if the Vega 10 gpu was just a pair of 480's glued together with a pool of hbm. That would be the biggest let down this week

 

[RejZoR]

There must be a typing error. Those bandwidth numbers don't make any sense. Wasn't HBM1 at 512 GB/s already? And even if you move the period around, 4096 GB/s doesn't sound right. That's 4 TB/s. Or am I missing something here?

 

[ppn]

clocked to 1.6. Down from expected 2.0Ghz.

 

[FootLong]

Fury had 4 stacks of HBM and Vega has 2.

There must be a typing error. Those bandwidth numbers don't make any sense. Wasn't HBM1 at 512 GB/s already? And even if you move the period around, 4096 GB/s doesn't sound right. That's 4 TB/s. Or am I missing something here?

 

[TheoneandonlyMrK]

There must be a typing error. Those bandwidth numbers don't make any sense. Wasn't HBM1 at 512 GB/s already? And even if you move the period around, 4096 GB/s doesn't sound right. That's 4 TB/s. Or am I missing something here?

Fury had In total from four double stacked ram chips at 1Gb/s per pin. =512
Vegas two quad stacked memory chips at 1.6Gb/s per pin =400 does seem low but it's likely a cost saving thing having two.

 

[Steevo]

Its not like the HBM bandwidth gave them a huge performance lead with Fury..... HBM was a bandaid to other cache issues and their proof of concept to ready them for one day integrating HBM on die without a interposer.

 

[Kanan]

That 410~ GB/s has to be enough. But maybe they'll run it with higher frequencies.

 

[FordGT90Concept]

Its not like the HBM bandwidth gave them a huge performance lead with Fury..... HBM was a bandaid to other cache issues and their proof of concept to ready them for one day integrating HBM on die without a interposer.

I don't think that's possible or at least not practical. They would have to cut way back on the amount of memory if they were to embed it in the chip.


Fury had four stacks of HBM not because they needed the bandwidth, but shipping a high end card without 4 GiB of RAM wouldn't sell. Even at 4 GiB they likely lost some potential buyers. HBM2 having a lot higher density means they don't need many stacks. They lose bandwidth because of it but they make up for that by providing at least 8 GiB of VRAM.

I eventually expect a GPU that has four stacks (1 TB/s, 16 GiB). It is very clear that won't be available in the initial offerings.