Graphcore Blog

For those that are interested, I was able to attend a talk by the guy that started Graphcore. His experienced approach to customized chip design for ML/AI is a really great thing to tap into. Some of the themes were: maximize power usage as its a limiting compute factor, remove off-chip memory, and scale horizontally at the individual processor level rather than "sharing resources across cores". https://www.graphcore.ai/

That hardware advance may be closer than expected: https://www.graphcore.ai

Companies like https://myrtle.ai/ and https://www.graphcore.ai/ are popping up.

Yes, there are several TPUs on the market now including https://www.sophon.ai/ https://www.graphcore.ai/ https://www.intel.ai/nervana-nnp/ and probably a lot more.

I went to a talk from the CTO of Graphcore (https://www.graphcore.ai/) on Monday. They are designing chips targeted at machine learning. As I understood it, their architecture comprises - lots of "tiles" - small processing cores with collocated memory (essentially DSPs) - very high bandwidth (90TB/s!) switching fabric to move data between tiles - "Bulk Synchronous Parallel" operation, meaning that the tiles do their work and then the switching fabric moves the data, and then we repeat. The key challenge he pointed to was power - both in terms of getting energy in (modern CPUs/GPUs take similar current to your car starter motor!) and also getting the heat out. Logic gates take a lot more power than RAM, so he argued that collocating small chunks of RAM right next to your processing core was much better from a power perspective (meaning you could then pack yet more into your chip) as well as obviously being better from a performance perspective. https://www.youtube.com/watch?v=Gh-Tff7DdzU isn't quite the presentation I saw, but it has quite a lot of overlap. Hope that helps!

One of the smartest projects I've ever seen was a tool that took the human-readable tables of the HEVC and AV1 specs, used them as input to https://en.wikipedia.org/wiki/OMeta parser-generator, and then output both HEVC parsers in a variety of languages and also auto-fuzzers for test coverage. Ended up at https://www.graphcore.ai/posts/graphcore-open-sources-argon-streams-av… Personally I've also written a parser-generator for XML in C# to overcome some of the odd limitations of Microsoft's one when used in AOT contexts. Hand-rolling is easy if the grammar is small. The larger it gets (and video codecs are huge!) the more you want something with automatic consistency.

Official Graphcore release: https://www.graphcore.ai/posts/graphcore-joins-softbank-group-to-build…

You can download their SDK without a login from their website here: https://www.graphcore.ai/downloads

>"each Bow IPU delivers up to 350 teraFLOPS of AI compute, an impressive 40% leap forward in performance and up to 16% more power efficiency compared to the previous generation IPU." Related: Colossus MK2 GC200 IPU: https://www.graphcore.ai/products/ipu

Lots of cool stuff in the hardware AI space https://sambanova.ai/ - Enterprise AI and dataflow-as-a-service for established moodels https://www.cerebras.net/ - AI accelerator, trying to compete with NVidia https://www.graphcore.ai/ - Another AI accelerator company - UK based https://femtosense.ai/ - Sparse NNs on very low power chips - cool hardware and software challenges https://sima.ai/ - ML accelerators for embedded applications https://ambiq.com/ - Not AI, but low power chips for wireless using some fancy tech that reduces energy leakage There are dozens more, these are just the ones I've heard of.

I agree. I have high hope for Graphcore. They stated "We will be participating in MLPerf in 2021, starting with the first training submission in the spring". https://www.graphcore.ai/posts/graphcore-sets-new-ai-performance-stand… (I would ignore all benchmarks in that post except the sentence quoted above however.)

Have you heard of the billion dollar unicorn from England, United Kingdom called Graphcore? https://m.youtube.com/watch?v=_zvU0uwIafQ https://www.graphcore.ai/products/mk2/ipu-machine-ipu-pod

Im sceptical that a good single CPU search can compete with massive parallel HW, like this one: https://www.graphcore.ai/posts/introducing-second-generation-ipu-syste…

Thank you for a blog, very much enjoyed reading it! By any chance do you know how Graphcore generates beautiful graph plots ? (seems like this are graph plots) link: https://www.graphcore.ai/hs-fs/hubfs/Graphcore%20open%20sources%20popl…

What does IPU stand for in this context? The only use I know of is Image Processing Unit — silicon that is custom developed for cameras. https://en.wikipedia.org/wiki/Image_processor This is clearly not that. Edit: Intelligence Processing Unit. https://www.graphcore.ai/products/ I assume that the Bionic neural engine in Apple's A11 and later is an example of this?

Perhaps Gephi? It was used for those Graphcore neural network visualizations (millions of edges and nodes) and they look stunning. [1] https://gephi.org/features/ [2] https://www.graphcore.ai/posts/what-does-machine-learning-look-like

Here is some explanation: https://www.graphcore.ai/blog/graph-computing-for-machine-intelligence…

From the last thread on this (https://news.ycombinator.com/item?id=32610780): - https://sambanova.ai/ (Enterprise AI and dataflow-as-a-service for established models) - https://www.cerebras.net/ (AI accelerator, trying to compete with Nvidia) - https://www.graphcore.ai/ (Another AI accelerator company, UK based) - https://femtosense.ai/ (Sparse NNs on very low power chips, cool hardware and software challenges) - https://sima.ai/ (ML accelerators for embedded applications) - https://ambiq.com/ (Not AI, but low power chips for wireless using some fancy tech that reduces energy leakage) - https://www.esperanto.ai/ (RISC-V based Tensor computes chip, founded by Intel Hybrid Parallel Computing Vice President Dave Ditzel) - https://www.furiosa.ai/ (AI accelerator company which show good results in MLPerf benchmark) - https://groq.com/ (From the team that built the original TPU at Google) - https://lightmatter.co/ (Light tubes instead of copper) - https://www.untether.ai/