A New Golden Age for Computer Architecture (2020) [video] | Hacker News Reader

A New Golden Age for Computer Architecture (2020) [video]

youtube.com

100 pointspjmlp25 comments

A New Golden Age for Computer Architecture - https://news.ycombinator.com/item?id=19237702 - Feb 2019 (8 comments)

A New Golden Age for Computer Architecture - https://news.ycombinator.com/item?id=19023734 - Jan 2019 (84 comments)

Also related:

John Hennessy and David Patterson Turing Award Lectures (2018) [video] - https://news.ycombinator.com/item?id=18118957 - Oct 2018 (5 comments)

Hennessy and Patterson win Turing Award - https://news.ycombinator.com/item?id=16637803 - March 2018 (32 comments)

The innovation we are all waiting on is an SSD which has the speed and performance of RAM such that they become one in the same. With some chip yields as low as 30% we are approaching the limit of shrinking transistors and the next 10 years of innovation will have to be architectural. The idea of photon computers is also a direction we will see more research into and might be about to get a 1mhz light processor by 2030

Are memristors still 5 years away or did HP bury those for business reasons?

Yeah what's going on? I bought a chunk of HP stock 10 years ago on the assumption they'd be revolutionizing swaths of the computing industry and it never materialized.

> HP ... revolutionizing swaths of the computing industry

Heh... Yeah, I don't HP in it's current incarnation can revolutionize anything.

That was sort of hyperbole, but the question stands :)

I think if you have a bet like that it’d make more sense to buy a sector ETF.

> The innovation we are all waiting on is an SSD which has the speed and performance of RAM

I've often wondered why no one offers a product that consists of DRAM with an integrated battery. Every portable device, when it goes into sleep mode, constitutes an existence proof that this could work. You could even integrate this with a regular SSD to provide persistence when the battery dies.

This has existed in various forms through the years: https://en.wikipedia.org/wiki/RAM_drive (section: Dedicated hardware RAM drives)

It's non-existence (or at least extreme difficulty to locate)for current gen tech suggests that it doesn't improve performance enough to be worth the tradeoffs, but I'm not an expert in that at all.

They exist mostly for servers. The capacity is much smaller since DRAM is much more expensive.

http://www.ddrdrive.com/

A company called Violin Memory used to make one for servers:

http://www.storagesearch.com/ssd-23.html

You can create persistent ramdisk with tmpfs.

Add some batteries and a cronjob with rsync and you have the product you are looking for.

This has been done for decades. For various reasons it hasn't become mainstream. Cost of DRAM is one of those reasons.

there even was NVDIMM

In theory, that would be MRAM - SRAM speeds, Non-volatile. Except as should be expected - there are details that have to be worked out in depth to ever allow such a substitution.

Persistent storage is not really a bottleneck. Single-threaded performance is. This is the innovation we are waiting for.

I thought modern computers are primarily cache limited and even running an os entirely on ram (on a system with no hdd/sdd) isn’t that much faster. I would be curious to see some data.

I'm halfway through the talk and he keeps mentioning points that are strangely reminiscent of things that usually come up when people like Mike Acton talk about Data-Oriented-Design, and writing code for the PS3's PPE-s

Nowadays he is working on the DOTS/Burst compiler team at Unity.

As I've said before, I just find it hard to believe that the very foreseeable problems lurking at the intersection of speculation, cache and protected mode were really such an unexpected shock to the industry https://youtu.be/kFT54hO1X8M?t=1195 . I have a jaded suspicion that in fact a fair few people had some idea, but they said nothing and didn't look closely because it would have been no fun, and in some cases not career-enhancing, to be the Jeremiah calling for major CPU deoptimisations. OTOH the very security-researcher heroes whose reputation is built on collecting trophies didn't seem to say much either, so maybe everyone really was completely blindsided?

Yes they were shocked. And it is simply because even the most fundamental laws of security (eg security is not scheduled via obscurity) have flown and fly over the heads of architects. They don’t understand software beyond the smallest of kernels which dominate execution time. They don’t understand that attackers need one weakness to exploit a system. They don’t know how to objectively analyze systems to prioritize their weaknesses (eg seeing how the coexistence and sharing between mutually untrusting software is a potential threat).

That’s why the biggest computer architecture breakdown came from outside the comp arch community.

I have anecdata that there were people who raised those exact concerns in those exact companies and were brushed off. If the companies were honestly surprised, it's because they purposely turned a deaf ear to the problem for the sake of greater revenue. But nothing really happened to those companies, so maybe it was ultimately a strategy that worked?

I can't help be reminded of the big US banks and sub-prime.

Reply on news.ycombinator.com