Tweet

This Thread may be Removed Anytime!

Twitter may remove this content at anytime! Save it as PDF for later use!

More from @draecomino

James Wang

@draecomino

31 May

@LidoFinance

ETH, with a little help from @LidoFinance and @CurveFinance, can generate 12% yield. But where does this yield come from?

Let's break down the ponzu recipe. 🧫👇

The Ethereum blockchain gives rewards to computers that validate transactions. If you hold ETH, you can validate transactions. The easiest way to do this is to use a service like Lido. The yield is currently ~6%. lido.fi

Normally when you stake your ETH, your ETH is locked up. Lido gives you staked ETH (stETH) tokens in return. This makes you ETH liquid and allows you to do stuff with them.

Read 6 tweets

James Wang

@draecomino

25 Aug 20

The public cloud makes it easy for anyone to start a software company—but at a cost—your margins now belong to AWS. Thread:👇

There are three ways of paying for software infrastructure:
1. have your customer pay for it (cheap)
2. build your own data center (somewhat costly)
3. rent from a public cloud (very costly)

The cheapest infra is no infra. This is the classic enterprise software model: the customer buys your sw to run on their own hw. Selling pure sw yields the highest margins in industry: 90%+.

Read 8 tweets

James Wang

@draecomino

13 May 20

1/ How James Cameron’s Terminator 2 predicted modern AI chips and sparked the debate on AI safety. An appreciation thread.👇

2/ This is the chip that powers the T-800. Based on its appearance and commentary from chief architect Miles Dyson, the movie makes three predictions about future processors: 1) neural net acceleration 2) multi-core design 3) 3D fabrication.

Let’s look at these claims.

3/ Among the many technologies Cameron could have picked for Terminator, neural network processor was spot on. Neural net is the breakout technology of the past decade. As of 2020, there are ~100 companies building neural net processors with annual revenues exceeding $5 billion.

Read 8 tweets

James Wang

@draecomino

15 Apr 20

1/ Apple's upcoming ARM MacBooks isn't just going to save them some money and run a bit faster. It marks the beginning of the end of the x86 era and Intel's four decade empire.

Thread⬇️

2/ In the computer industry, victories are won through standards and scale. Intel invented the x86 standard. And by winning in the largest market of the 90s—PCs, it moved upmarket and eclipsed all server CPU vendors in a decade.

3/ Today smartphones are by far the largest computer market. At $600B, it’s larger than PCs and servers combined. Smartphones use ARM CPUs while PCs and servers use x86. Through the forces of standards and scale, it ought to displace x86 CPUs in PCs and servers. Why hasn’t it?

Read 11 tweets

James Wang

@draecomino

16 Jan 20

ARK's Big Ideas 2020 deck is here—a year of research packed into 80 slides covering AI, robotics, autonomous, genomics, bitcoin, and more.

Download: ark-invest.com/big-ideas-2020

Here are 5 slides that really hit it home. Thread:

Deep learning continues to improve at astounding rates. In 2017 it costed $1,000 to train ResNet50 on the cloud, now it costs $10.

Thanks to AI accelerators and 'unlimited' compute in the cloud, AI algorithms are eating up 10x the amount of FLOPs per year.

They promised us flying cars and instead we got something even better—drones!

Drones are to flying cars as smartphones are to mainframes—they are cheaper, smaller, and serve more markets.

Everything from delivering Starbucks to moving people can be made 10x cheaper with drones.

Read 6 tweets

James Wang

@draecomino

20 Aug 19

9/ Neural nets can consume GBs of memory. GPUs only have MBs of on-chip memory. So GPUs store neural nets on external memory soldered next to it on the PCB.

The problem is external memory is 10-100x slower & more power hungry vs. on-chip memory. They are also very expensive.

10/ Large models like Google’s Neural Machine Translation don’t even fit in one GPU’s external memory. Often they have to be split up across dozens of GPUs/servers. This increases latency by another 10-100x.

Ideally the whole model fits on a single chip—that's Cerebras' WSE.

11/ Cerebras’ Wafer Scale Engine (WSE) is *one chip* holding 400,000 cores and 18GB of memory. Neural network training happens on one piece of silicon rather than spread across dozens of boards, servers, interconnects. If it works, one chip can replace a rack of GPU servers.

Read 8 tweets

Share this page!

James Wang

Try unrolling a thread yourself!

More from @draecomino

James Wang

James Wang

James Wang

James Wang

James Wang

James Wang

Did Thread Reader help you today?

Like this author's thread?