|
|
|
|
||
Samsung places processing-in-memory chip onto AMD MI100 GPUDecember 14, 2022 by Info News Go Samsung has constructed a claimed first-of-its-kind supercomputer containing AMD datacenter GPUs affixed with its processing-in-memory chips, which the corporate stated can considerably enhance the efficiency and vitality effectivity of coaching giant AI fashions. The supercomputer, disclosed Tuesday at an business occasion in South Korea, consists of 96 AMD Intuition MI100 GPUs, every of that are loaded with a processing-in-memory (PIM) chip, a brand new form of reminiscence know-how that reduces the quantity of information that should transfer between the CPU and DRAM. Choi Chang-kyu, the pinnacle of the AI Analysis Heart at Samsung Electronics Superior Institute of Expertise, reportedly stated the cluster was capable of prepare the Textual content-to-Take a look at Switch Transformer (T5) language mannequin developed by Google 2.5 occasions sooner whereas utilizing 2.7 occasions much less energy in comparison with the identical cluster configuration that did not use the PIM chips. “It’s the solely one in all its type on this planet,” Choi stated. Samsung has stated that its PIM tech has main implications for vitality consumption and the atmosphere, decreasing annual vitality use of a cluster by 2,100 Gigawatt hours and, consequently, chopping down 960,000 tons of carbon emissions. As at all times, we must always reserve judgement till these claims may be examined and verified independently, however the firm stated such a discount in energy is the same as the quantity of carbon it could take 16 billion city timber to soak up over a decade. One huge cause why the PIM-powered supercomputer has a lot horsepower is that every PIM chip makes use of high-bandwidth reminiscence (HBM), which the business is more and more turning to for dealing with high-performance computing and AI workloads. Nvidia and AMD have used HBM in datacenter GPUs for a number of generations now, and Intel plans to introduce HBM in a forthcoming variant of server processors branded Xeon Max and a high-end datacenter GPU. What makes Samsung HBM-PIM chips totally different from HBM implementations by different corporations is that every reminiscence financial institution on the PIM chip features a processing unit inside. This, in keeping with the South Korean electronics large, reduces bottlenecks related to shifting knowledge between the CPU and reminiscence by shifting among the computation contained in the reminiscence itself. Samsung hopes to spur adoption of its PIM chips within the business by creating software program that can enable organizations to make use of the tech in an built-in software program atmosphere. To do that, it is counting on SYCL, a royalty-free, cross-architecture programming abstraction layer that occurs to underpin Intel’s implementation of C++ for its oneAPI parallel programming mannequin. The corporate has been hyping up PIM for practically three years now, and one different manner it plans to take the tech to market is thru what it has referred to as the AXDIMM, brief for accelerated DIMM. We’ll know if Samsung finally ends up making any inroads if PIM begins showing in new supercomputers arrange by analysis labs, tutorial establishments, and different organizations over the following few years. ® https://infonewsgo.in/samsung-places-processing-in-memory-chip-onto-amd-mi100-gpu-the-register/ |
return to message board, top of board |
Msg # | Subject | Author | Recs | Date Posted |
1576 | Re: Samsung places processing-in-memory chip onto AMD MI100 GPU | doodaddy64 | 0 | 12/15/2022 9:32:47 AM |