Graphene FET on diamond for high-frequency electronics (2021) | Hacker News Reader

Graphene FET on diamond for high-frequency electronics (2021)

arxiv.org

35 pointsnynx22 comments

Am I the only person here who wants to put ~60 thousand of these on a chip and make a 54 GHz 6502 for a ridiculously fast Apple II?

Could be a Commodore 64 or Atari too ;-)

Would it work though? The transistors are quite large compared to today's processor's transistors, and at that high frequency the routing delay for the signal to travel around transistors is quite significant. This seems more useful for a low-transistor-count high frequency device.

It’s more useful for the intended uses. OTOH, it’s always fun to misuse technology.

There would need to be multiple carefully positioned clock entries or a fully asynchronous design rather than just a copy of the NMOS layout, but it’d still be a lot of fun.

Why 60000? Do you mean both the CPU + the RAM and peripherals chips?

Wow. I was off by an order of magnitude. With 50K you can probably do a complete Apple //e!

Yes, you had me puzzled for a bit :)

Well... If the whole Apple II runs at 54GHz it'll be a lot easier to interface to the external world.

A 54GHz 6502 would die of boredom waiting for the RAM.

10000 wait state memory :)

With the current prices of graphics cards, it may be more economical to make them out of diamonds...

I know you joke, but with the existing costs of fab manufacturing I wouldn't expect that using a diamond substrate would take up that much of the cost.

New diamond manufacturing processes include chemical vapor deposition, which would make integrating with fab processes pretty feasible. I'd guess it depends on weighing the performance graphene can offer vs the cost increase over silicon.

> New diamond manufacturing processes include chemical vapor deposition, which would make integrating with fab processes pretty feasible.

That is honestly pretty freaking cool.

Afaik synthetic diamonds are nowhere nearly as expensive as natural ones. Diamonds are just carbon after all.

So far as I know, no one can grow diamonds as big as wafers so I think it's sort of moot. Some other process like vapor deposition would probably be used rather than starting from a giant block of diamond.

Nice BUT. The actual frequency/clock rate of a system is NOT defined by the transistor Ft value - it defined by the parasitics of the circuit it's placed into. If you build an IC with something like this, the claimed Ft does NOT define or allow the building of something far beyond the frequency or clock of what we have now. That is the fundamental flaw of both the PR by groups announcing such things and of non-EEs reading about them.

I wonder what the hashrate would be if this was used for a bitcoin ASIC…

Always look at performance per dollar, not absolute performance.

Well not quite.

In this case rewards are highly nonlinear, not proportional to processing work at all.

The value measure is how good performance is relative to competitors, with a vastly greater than linear reward if it is

I expect energy use (and other performance costs) do not matter much once your winning proof-of-work competitions due to higher performance (as apposed to simple scaling of hardware)

Given this is an experimental device it’s probably only the fab cost that is a barrier

Interesting, I was under the impression that for mining cryptocurrencies 10x the amount of machines was equivalent to 1 machine at 10x the performance.

So I guess this means that cryptocurrencies are akin to a winner-takes-all situation. Raising the question: is that how we want our monetary systems to work?

Well it's not exactly "winner takes all". It's more like: The current winner takes a little bit, for a while, until the next winner emerges (usually through development or commoditization of new technology). There was a situation like that in Bitcoin when ASICs first started out when everyone else was still doing GPUs and few people had FPGAs. BTC survived.

Yes its interesting both because of the series of winner-take-all competitions, but also because while cost-of-processing X computations is linear with scaling, the processing threshold X is not a constant, but a variable based on competition.

You are incentivized to increase costs/performance wherever that will allow you to win competitions.

What costs do that? Any cost that increases your speed faster than linear - i.e. transitions to new hardware tech.

Note that if the cost of the transition is achievable but extremely high, that actually makes the transition more attractive. It is a barrier to followers who won't be rewarded as much for their investment as the first to do so will be.

A hyper competitive market.

it's cool, but the problem with previous graphene transistors is that they didn't really have a bandgap, meaning there wasn't much of a difference between on and off, making them unsuitable for use in computers. It looks like that could be the case here too.

Looks like a market opportunity for de beers...

Reply on news.ycombinator.com