Samsung's New HBM2 Memory Thinks for Itself: 1.2 TFLOPS of Embedded Processing Power
Today, Samsung announced that its new HBM2-based memory has an integrated AI processor that can push out (up to) 1.2 TFLOPS of embedded computing power, allowing the memory chip itself to perform operations that are usually reserved for CPUs, GPUs, ASICs, or FPGAs.
The new HBM-PIM (processing-in-memory) chips inject an AI engine inside each memory bank, thus offloading processing operations to the HBM itself. The new class of memory is designed to alleviate the burden of moving data between memory and processors, which is often more expensive in terms of power consumption and time than the actual compute operations.
[...] As with most in-memory processing techniques, we expect this tech will press the boundaries of the memory chips' cooling limitations, especially given that HBM chips are typically deployed in stacks that aren't exactly conducive to easy cooling. Samsung's presentation did not cover how HBM-PIM addresses those challenges.
HBM: High Bandwidth Memory.
ASIC: Application-Specific Integrated Circuit.
FPGA: Field-Programmable Gate Array.
(Score: 0) by Anonymous Coward on Thursday February 18 2021, @12:47AM (11 children)
Sound like neuron chip, flipping von neumann model.
(Score: 2) by takyon on Thursday February 18 2021, @12:59AM (9 children)
It sounds more mundane and awful than that. This is not some neuromorphic memristor thing, but it does move an AI accelerator closer to memory.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 3, Touché) by Arik on Thursday February 18 2021, @01:12AM (7 children)
Could you translate that to something meaningful?
If laughter is the best medicine, who are the best doctors?
(Score: 2) by takyon on Thursday February 18 2021, @01:26AM
Machine learning accelerator.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 1, Funny) by Anonymous Coward on Thursday February 18 2021, @01:35AM
It begins to learn at a geometric rate. It becomes self-aware at 2:14 a.m. Eastern time, August 29th.
(Score: 1, Informative) by Anonymous Coward on Thursday February 18 2021, @05:07AM (4 children)
The modules are FIMDRAMs that have a built-in PCU that is capable of performing FP16 SIMD operations at 300MHz with bank-level parallelism.
(Score: 0) by Anonymous Coward on Thursday February 18 2021, @05:32AM (1 child)
Sheesh, no need for all the techie jargon, and acronyms and shite, that just confuses matters.
How about you just say it accelerates the operations that are most commonly used in AI-related computations? Or an "AI accelerator" for short?
(Score: 2, Insightful) by Arik on Thursday February 18 2021, @10:27AM
If laughter is the best medicine, who are the best doctors?
(Score: 2) by Arik on Thursday February 18 2021, @10:25AM (1 child)
If laughter is the best medicine, who are the best doctors?
(Score: 0) by Anonymous Coward on Thursday February 18 2021, @10:52AM
You're welcome. Anyone interested can feel free to ask if they need any of the alphabet soup explained. Hopefully after at least attempting to read the Wikipedia articles so I don't have to start all the way at zero.
(Score: 0) by Anonymous Coward on Thursday February 18 2021, @01:35AM
"It sounds more mundane"
It kinda does. Basically a smaller computing node.
(Score: 1) by fustakrakich on Thursday February 18 2021, @03:10AM
I thought it was a work visa
La politica e i criminali sono la stessa cosa..
(Score: 1, Insightful) by Anonymous Coward on Thursday February 18 2021, @12:56AM (2 children)
http://cva.stanford.edu/classes/cs99s//papers/myer-sutherland-design-of-display-processors.pdf [stanford.edu]
(Score: 1, Insightful) by Anonymous Coward on Thursday February 18 2021, @01:38AM (1 child)
When I was a kid, a Tektronix salesman got us a 24bpps "loaner". We had a 300bps acoustic modem to the Perkin-Elmer at the company.
300bps * ATT phone charges meant I only got so much time playing on the "big" computer.
Figured out how to half-way do space-invaders on the local Tektronix "monitor".
---
Served me well the rest of my career. "Its the mainframe" ... "Its the LAN" ... "Its the Internet" ... "Its the Browser" ... "Its the Cloud" ... "Its the RAM" ...
Cycle is centralize/decentralize ... once you know what's hip, you know what jobs to look for if you want to catch the next wave.
(Score: 2) by Muad'Dave on Thursday February 18 2021, @12:50PM
Too cool! My first job out of college was working for Concurrent Computer, which was PE's computing division. Do you remember what model you used?
I still remember OS/32 commands and Assembler instructions.
(Score: 0) by Anonymous Coward on Thursday February 18 2021, @01:07AM
If this thing gets wide use, we'll have to refine the definition of computing "nodes".
(Score: 5, Funny) by crm114 on Thursday February 18 2021, @01:13AM (1 child)
user: give me access to address 0x1225817701, for 128bits
HAL^HMemory : I'm sorry, Dave. I'm afraid I can't do that.
(Score: 1, Funny) by Anonymous Coward on Thursday February 18 2021, @01:43AM
You see this? It's my Kill-O-Zap assault rifle. You see that? That's your memory bank.
What was it you said again?
(Score: 2) by legont on Thursday February 18 2021, @02:57AM
Now memory itself can decide if it does or does not want to share it's, well, memory with the requester. I imagine AI of the particular memory bit checking with bio-metrics of the user before responding to the request.
"Wealth is the relentless enemy of understanding" - John Kenneth Galbraith.
(Score: 2) by inertnet on Thursday February 18 2021, @08:14AM
No matter what you store into them, in due time all memory cells will contain 42.