"In addition to increasing capacity and speed, the improvements in energy efficiency are noteworthy. With HBM3, the core voltage is 1.1V, compared to HBM2E’s 1.2V core voltage. HBM3 also reduces the I/O signaling to 400mV versus 1.2V for HBM2E. There will be further improvements in future generations, as well."
AI already has led to a rethinking of computer architectures, in which the conventional von Neumann structure is replaced by near-compute and at-memory floorplans. But novel layouts aren’t enough to achieve the power reductions and speed increases required for deep learning networks. The industry also is updating the standards for floating-point (FP) arithmetic. https://semiengineering.com/will-floating-point-8-solve-ai-m...
Software engineers looking for a job should consider the chip industry, which is not just about hardware engineers. Plenty of software engineer openings https://semiengineering.com/jobs/ and many are remote
Are you speaking from experience? Almost all of the jobs at the page you linked are either not in the U.S. or require hardware education or experience. Most are not remote. Could you please describe the actual steps and timeline for, say, a frontend or backend web developer without an engineering degree to transition into one of those jobs?
Do you have strong CS fundamentals (generally acquired through university education) ? Because that stuff really matters if you want to work on stuff related to hardware
Most of the listings require past experience with hardware stuff, not just CS fundamentals. I'm seeing very little interest in hiring someone like an experienced React developer who just got laid off from FAANG. There's a big difference between being theoretically capable of doing a job, and having any reasonable chance of getting hired for that job.
why did you say CS fundamentals when it's clear that it should be computer architecture and instruction sets and compiler engineering, which are a very narrow and specific parts of CS and in fact are never the parts that people refer to (DS & algo) when they say CS fundamentals.
They're pretty closely tied together. If you have a strong background in algorithms you should be able to understand ISA's and compilers more easily. In the end its just algorithms all the way down
Software engineering in the chip industry always seemed like a completely and totally different thing to me. Knowledge of physics and nanoscale architecture and calculus and all kinds of stuff that I don't need to know just to create an online form on my webiste for people to fill out.
I don't know, maybe I have the wrong view, but it always seemed like that to me. I just do business software development because all it has is addition, subtraction, division, and muliplication, and percentages every once in a while. No way could I create a program for a hardcore physics app. Because you have to know physics. To some extent, anyways. Anything more than basic math, forget it. And I think most people at organizations like facebook, netflix, etc are more like me. Maybe I'm wrong, but that would be my bet.
Dealing with the exponential increase in data is driving some massive architectural changes. This Univ of Penn research is interesting. Chip companies are currently working a number of strategies. Related: https://semiengineering.com/ic-architectures-shift-as-oems-n...
not just the chip shortage. The rise of RISC-V coincides with a couple of other events in the industry. The first is the slowing of Moore’s Law, meaning that increases in total processing power no longer comes along with each new fabrication node. The second is the meteoric rise in machine learning, demanding massive increases in processing power. https://semiengineering.com/why-risc-v-is-succeeding/
Language models and image generation make fun demos, but do we have transformative use cases that'll actually require large ML compute in the future ?
Voice recognition and translation are the only ones that comes to mind, yet don't require that much power.
Hmm,
Is there any discussion of how RISC-V designs could be incorporated into a GPU or TPU that could train deep learning systems? Your link doesn't say anything about that but it's an interesting question.
