Thread by @jwangARK

1/ Congrats on @graphcoreai on its $30m round. It’s still early days, but this could be another ARM in the making!
bloomberg.com/news/articles/…

2/ Today GPUs have a monopoly position on deep learning training. But that’s set to change: DL ASICs could boost performance by 10-100x.

3/ Three notable teams are working on DL chips: 1) Google TPU team 2) Intel’s Nervana team 3) Graphcore

4/ Google’s TPU is the most far along. It’s already on gen2 and deployed in Google’ datacenters for training and inference.

5/ Intel’s Nervana is headed by @NaveenGRao of ex-QCOM fame. The chip uses mixed precision and ships later this year.

@NaveenGRao 6/ Which brings us to Graphcore, probably the most radical of the 3 designs, or in the words of investor @sirajkhaliq, “the most elegant”.

7/ For starters, Graphcore’s Colossus chip has *no external memory*. All memory is on-chip, making up 75% of the die.

8/ Moving memory on die improves bandwidth and latency by ~100x—a huge boon to deep learning apps, which are often memory starved.

9/ Using so much area for RAM means less area for logic. But since the chip doesn’t do graphics or HPC, it can do more with less.

10/ Unlike chips from Google and Nervana that were built on 28nm, Graphcore is 16nm out the gate. Claim is “well beyond Volta perf" (!!)

11/ The Graphcore team has done this before. They are the folks who founded Icera, a software modem company that NVIDIA acquired in 2011.

12/ The investors in this round are mighty impressive. They include DeepMind’s @demishassabis and OpenAI’s @gdb and @ilyasut.

@demishassabis @gdb @ilyasut 13/ Stepping back, Graphcore is another showcase of UK semiconductor talent. It could grow up to be the next IMG tech or ARM. Best of luck!