A processing unit (CPU, GPU or no matter) and RAM are generally separate issues constructed on separate chips. However what within the event that they were portion of the identical chip, all mixed collectively? That’s exactly what Samsung did to secure the sector’s first Excessive Bandwidth Memory (HBM) with constructed-in AI processing hardware referred to as HBM-PIM (for processing-in-reminiscence).
It took its HBM2 Aquabolt chips and added Programmable Computing Units (PCU) between the reminiscence banks. These are somewhat easy and operate on 16-bit floating level values with a dinky instruction space – they are able to switch files around and fabricate multiplications and additions.
However there are a range of PCUs and so that they literally sit down next to the concepts they are engaged on. Samsung managed to secure the PCUs working at 300 MHz, which works out to 1.2 TFLOPS processing energy per chip. And it kept the energy usage (per chip) the identical whereas transferring files at 2.4 Gbps per pin.
Per-chip energy usage could be the identical, nonetheless general procedure energy consumption drops by 71%. That is because a same old CPU would must switch files twice – be taught the enter then write the finish end result. With HBM-PIM the concepts doesn’t surely race any place.
It’s no longer factual energy saving, the reveal of PIM for machine studying and inference tasks researchers saw procedure performance better than double. That’s a steal-steal misfortune.
The HBM-PIM make is backwards love minded with customary HBM2 chips, so no novel hardware wishes to be developed – the tool factual wishes to say the PIM procedure to alter from customary mode to in-reminiscence processing mode.
There is one enviornment with this and it’s the PCUs mediate up region previously occupied by reminiscence banks. This cuts the total ability in half – the total procedure down to 4 gigabits. Samsung decided to cleave up the adaptation and blend 4 gigabit PIM chips with 8 gigabit customary HBM2 dies. The usage of four of every it created 6 gigabyte stacks.
There’s some extra unhealthy files – this would possibly occasionally likely presumably well be a whereas ahead of HBM-PIM lands in consumer hardware. For now Samsung has despatched out chips to be examined by partners developing AI accelerators and expects the make to be validated by July.
HBM-PIM shall be presented at the World Stable-Relate Circuits Digital Conference this week, so we are able to set a question to extra particulars then.