The Korean giant is taking advantage of its lead in the chip world to announce mass production for ever more agile graphic memories
A few weeks after we began to fantasize about the improvements they would bring Memories of the GDDR6, With effective bandwidths of 450 gigabytes per second for mid-range cards and almost 1,000 gigabytes per second at the seemingly top end - comes an announcement designed to remind us that the whole GDDR world is still a remnant of the past, trying to catch up with the pace of The new star in the city.
The new star is called HBM, or HBM2 to be precise, and has become more relevant than ever at the CES 2018 exhibition with the announcement of The Kaby Lake-G processors Which include cores Radeon Vega Structured and advanced memories of this kind, as well as announcing processors Vega Graphics for the mobile market - when it will not be excessive to declare that soon a large part of the graphics cards for the high market and maybe even the medium market will contain this standard, which is already capable of reaching a monstrous bandwidth of terabytes per second.
Samsung is not satisfied with the latest state of the art, and announces the start of mass production for a new generation of HBM2 chips, which bring even more impressive 2.4 gigabit per second performance per pin in the ultra-wide 1,024 bit interface of this standard (compared with 2 max) Gigabit per second to date), under a modest basic workload of 1.2 volts and with an increased storage density that allows for a total volume of 8 gigabytes per chassis Memories. The end result of all this is the effective bandwidth of 307 gigabytes per second per package, when we already know that it is possible to combine two to four Enclosures Those with a single processing core - which means we could soon have graphics processing cores with volume memory Dedicated 32 gigabytes, and an impressive bandwidth of over 1.2 terabytes per second.
All of this is certainly impressive, although it is worth noting that this is only an intermediate development, on the way to the arrival of the HBM3 generation of memories - perhaps as early as next year. These memories are expected to be another huge leap forward, with support for larger overall volumes of up to 64GB (!), A "much lower" operating voltage than the 1.2 volts of the HBM2 generation and especially doubling the effective bandwidth per Such a memory structure of 512 GB per second, twice the base of HBM2 and 2 percent more than the innovative memories of סמסונג. With an array of four buildings Memories On the chip we will reach an imaginary bandwidth of 2 terabytes per second (!), And if we see support for arrays of six or eight structures, this figure could reach 4 terabytes per second - insane.