vacancies advertise contact news tip The Vault
facebook rss twitter

Nvidia will source its HBM2 from both Samsung and SK hynix

by Mark Tyson on 21 September 2015, 12:06

Tags: NVIDIA (NASDAQ:NVDA), Samsung (005935.KS), SK hynix

Quick Link: HEXUS.net/qacusl

Add to My Vault: x

A couple of months ago we heard rumours that AMD had secured priority access to SK hynix's HBM2 chip output. The news sources explained that AMD had worked closely with SK hynix on HBM(2) development and the pair of companies has a long established relationship. Furthermore AMD was expected to strive to capture "as much of the initial production capacity as possible," which could have derailed Nvidia's Pascal GPU launch proper.

In a report published earlier today, Business Korea said that both SK hynix and Samsung Electronics are planning to mass produce second-gen High Bandwidth Memory (HBM) for Nvidia's Pascal GPU. Production will start in Q1 2016 following the pilot production and reliability tests expected to complete later this year.

While we know that Nvidia will pair up its Pascal GPU with HBM2 (called 3D memory in its roadmap slide), we don't have any roadmap from AMD indicating what it plans to do with the second generation of this high bandwidth memory. HBM2 offers significant enhancements over the first generation including higher density, higher performance, and lower power usage. More specifically, HBM2 is said to offer twice the bandwidth of HBM and offer a fourfold leap in density, with a maximum chip capacity of 32Gb compared to the 8Gb possible with HBM.

In addition to the HBM2 bandwidth boost, Nvidia says it will enhance the performance of Pascal GPUs with NVLink. This interconnect technology is said to allow data sharing "at rates 5 to 12 times faster than the traditional PCIe Gen3 interconnect".

The first shipments of HBM2 are said to be destined for use with GPUs, like Pascal, and also for servers, supercomputers, networks, and high-performance PCs. It is expected that HBM equipped graphics cards will be 'mainstream' by 2018.



HEXUS Forums :: 5 Comments

Login with Forum Account

Don't have an account? Register today!
yawn @ 2018………
lumireleon
yawn @ 2018………

It makes sense.

14/16nm stuff will need HBM for the high end, next year.

10nm parts will need HBM for the mid range. 2018 seems plausible if a little hopeful given the slips we have seen in recent processes ramping up.

Cheap cards will use DDR4 for low cost, or be replaced with integrated graphics and simply not exist.
I do wonder if AMD are being a bit tight-lipped on their future plans for HBM because they're waiting to see what supply will be like. With a single cube of HBM2 giving up to 4GB of VRAM with 256GB/s of bandwidth (slightly better than an R9 280X), high yields of HBM2 could make it viable for mid-range cards in the next generation, but the supply would need to be plentiful and relatively cheap to ensure the kind of volumes that AMD would want for a next-gen card with similar performance to a 380/280X.
Another foible is that use of HBM/HBM2 requires an interposer. The interposer needs to be at a minimum the area of the GPU die PLUS the width of one HBM stack added to one side of the die.
Even on an older process (with amortised setup costs) that means fabbing a pretty big die, and a big fixed cost to add to the manufacturing costs. That means HBM isn't going to be a good fit or low-margin lines.
scaryjim
I do wonder if AMD are being a bit tight-lipped on their future plans for HBM because they're waiting to see what supply will be like. With a single cube of HBM2 giving up to 4GB of VRAM with 256GB/s of bandwidth (slightly better than an R9 280X)…

Isn't the 256GB/s of bandwidth per stack though, I thought the Fury X has 512 GB/s of bandwidth using standard HBM so wouldn't a card with HBM2 have 4x as much bandwidth?