Samsung Unveils New AI Memory
In the normal scheme of things microprocessors perform functions on data. Since those functions can have more than one processing step, the processor sends the interim data to a memory bank until it is needed for the next calculation, at which point it is sent back to the processor. While this is standard practice for typical systems, there is a great deal of traffic between the processor and the memory block, and depending on the type of calculations, the sequential nature of how the data is processed, the traffic, and the time between when the data is stored and when it is returned for more computations, all contribute to the amount of data that can be processed in a given timeframe.
Samsung, the leader in the memory market, has come up with a solution that should increase the throughput of existing systems, without adding to the processing load. This is accomplished by adding a ‘mini’-processor (PCU or Programmable Computing Unit) to each bank of the memory (typically 32 per memory chip), in the memory itself, which can be programmed to do some of the data processing that would normally have to be done by the main processor. As these ‘mini-processors are programmable, they can be set to perform functions based on the system’s application, making the device ‘intelligent memory’ or they can be turned off, allowing the memory to operate in ‘normal’ mode. The PCUs utilize standard memory commands for the PCUs, so Samsung says early adopters would not have to make any hardware or software changes when using the new memory configuration.
Of course, everything comes at a price, and in this case we do not mean the actual cost of the chip, but in real estate. Since the mini-processors are part of the memory chip itself, they take up some of the real estate that would normally be used for memory. In fact the PCUs take up about half of the memory space of a standard 8Gb die so Samsung offsets some of that loss by combining both regular memory and PCU memory, returning about half of what was lost. All in, Samsung says that the new chip, which is currently being tested by customers (expected for commercial release in 2H), provides 4x higher processing bandwidth than an off-chip solution, reduces power consumption by 70%, and does not need and changes to conventional memory controllers and their command protocols. Not addressed thus far is the cost and the ability to keep such stacked chips cool given the internal processing, but if the device lives up to its promotional material, we expect the heat challenge will be taken care of, while the cost will likely depend on how quickly Samsung wants the new devices to become widespread.
Samsung, the leader in the memory market, has come up with a solution that should increase the throughput of existing systems, without adding to the processing load. This is accomplished by adding a ‘mini’-processor (PCU or Programmable Computing Unit) to each bank of the memory (typically 32 per memory chip), in the memory itself, which can be programmed to do some of the data processing that would normally have to be done by the main processor. As these ‘mini-processors are programmable, they can be set to perform functions based on the system’s application, making the device ‘intelligent memory’ or they can be turned off, allowing the memory to operate in ‘normal’ mode. The PCUs utilize standard memory commands for the PCUs, so Samsung says early adopters would not have to make any hardware or software changes when using the new memory configuration.
Of course, everything comes at a price, and in this case we do not mean the actual cost of the chip, but in real estate. Since the mini-processors are part of the memory chip itself, they take up some of the real estate that would normally be used for memory. In fact the PCUs take up about half of the memory space of a standard 8Gb die so Samsung offsets some of that loss by combining both regular memory and PCU memory, returning about half of what was lost. All in, Samsung says that the new chip, which is currently being tested by customers (expected for commercial release in 2H), provides 4x higher processing bandwidth than an off-chip solution, reduces power consumption by 70%, and does not need and changes to conventional memory controllers and their command protocols. Not addressed thus far is the cost and the ability to keep such stacked chips cool given the internal processing, but if the device lives up to its promotional material, we expect the heat challenge will be taken care of, while the cost will likely depend on how quickly Samsung wants the new devices to become widespread.
RSS Feed