At the SC24 supercomputing conference held in November in Atlanta, Hafþór Júlíus Björnsson (the actor who played The Mountain in Game of Thrones), deadlifted a custom barbell weighed down by 453 kilograms (1000 pounds) of solid state drives. The data stored in those drives totaled just over 280 petabytes.
“Without question, this is the most data lifted by a human in history,” says Andy Higginbotham, senior director of business development at Phison Electronics.
Behind this publicity stunt is a real trend—to feed AI’s insatiable data appetite, memory drives are getting larger, with no end in sight. Phison recently announced the largest SSD memory drive to date, storing 128 terabytes of data, and piled hundreds of them into Björnsson’s barbell. Within a few weeks, Solidigm announced its own 123 Tb drive. Samsung and Western Digital also recently started carrying similar products.
The shift towards more AI workloads in data centers has led to very power-hungry chips, mostly GPUs. Since the overall power use in a data center is going up, people are looking for ways to use less power wherever possible. At the same time, large language models and other AI models require ever increasing amounts of memory.
“You can see where storage requirements are going,” says Roger Corell, senior director of AI and leadership marketing at Solidigm. “You look at a large language model just a couple years ago, you had a half a petabyte per rack or lower. And now there’s large language models that pair with between three and three and a half petabytes per rack. Storage efficiency to enable continued scaling of AI infrastructure is really, really important.”
Crucially, this new crop of solid-state drives takes up the same area in a computing rack and power budget as their roughly 32 Tb and 64 Tb predecessors—although they are slightly taller—meaning they can be swapped into data centers for an easy win.
3 Ways to Boost SSD Capacity
“There’s three kind of vectors you can innovate on to drive up capacity per SSD,” says Corell. “One is the bits per cell, and then two is how many cells can you pack in one layer, and how many layers can you stack of these memory cells.”
The number of bits per cell is literally how many different values can be stored in a single NAND flash floating-gate transistor. These transistors are packed onto a layer with ever-increasing densities. Adding more layers does make the device taller, but, Corell says, height is not a limiting factor in data centers today, so there is room to expand.
Solidigm was already making four-bit-per-cell devices, so it innovated by packing more cells per layer to go from the 60 Tb model to its 122 Tb device, mainly by using the smallest available NAND technology and reducing the size of non-NAND components. Phison went from thee-bit to four-bits NAND cells, improving along all three vectors to go from its 32 Tb to its 128 Tb drive.
Going to a higher number of bits per cell makes the write times somewhat slower, but this is still a win in a lot of applications, says Allyn Malventano, senior manager of technical marketing at Phison.
Maintaining the same power draw as smaller devices is a fine balancing act. “The drive developers are always going to have to do to figure out, okay, if we configure it this particular way, we maybe we can get some higher performance, but if we do that, it’s going to take some more power in order to do that right,” Malventano says. “So there’s a, basically a tuning operation that goes on for any particular capacity.”
The demand for higher capacity memory drives is not slowing down, either. “It’s never going to stop. There’s always going to be more demand for larger SSDs,” Higginbotham says. Solidigm’s Corell predicts there will be petabyte SSDs on the market before the decade is done.
Reference: https://ift.tt/EDV5fY8
No comments:
Post a Comment