Wednesday, June 27th 2018

Samsung Doubles its HBM2 Output, May Still Fall Short of Demand

Samsung has reportedly doubled its manufacturing output of HBM2 (high-bandwidth memory 2) stacks. Despite this, the company may still fall short of the demand for HBM2, according to HPC expert Glenn K Lockwood, Tweeting from the ISC 2018, the annual HPC industry event held between 24th to 28th June in Frankfurt, where Samsung was talking about its 2nd generation "Aquabolt" HBM2 memory, which is up to 8 times faster than GDDR5, with up to 307 GB/s bandwidth from a single stack.

While HBM2 is uncommon on consumer graphics cards (barring AMD's flagship Radeon RX Vega series, and NVIDIA's TITAN V), the memory type is in high demand with HPC accelerators that are mostly GPU-based, such as AMD Radeon Instinct series, and NVIDIA Tesla. The HPC industry itself is riding the gold-rush of AI research based on deep-learning and neural-nets. FPGAs, chips that you can purpose-build for your applications, are the other class of devices soaking up HBM2 inventories. The result of high demand, coupled with high DRAM prices could mean HBM2 could still be too expensive for mainstream client applications.
Source: Glenn K Lockwood (Twitter)
Add your own comment

2 Comments on Samsung Doubles its HBM2 Output, May Still Fall Short of Demand

#1
bug
consumer graphics cards (barring AMD's flagship Radeon RX Vega series, and NVIDIA's TITAN V)
Lol, good one :D
Posted on Reply
#2
looniam
i'm sure some high end enthusiasts will disagree but i am ok with HBM staying in the HPC space for awhile.

so yeah, thanks guys for footing the bill.
Posted on Reply