Ravi A Giri

Technologist: Principal Compute Architect @ Azure; Ex-Sr. PE @ Intel. Tweets barely represent my own opinions, let alone my employers

[Ravi A Giri @ravigiri]

@QuinnyPig I assumed MS Build would be a 'target rich' env. for you :-) ; I would humbly recommend this paper, it's worth a read: https://t.co/Fep9mTkkqi

[Ravi A Giri @ravigiri]

@danluu There are unlimited tokens effectively, but not for personal projects as far as I know. There are thresholds that trigger email notifications + follow-up's - if you have justification it'll be allowed - it's all too easy to consume a crazy amount of tokens and not be aware of it

[Tim Prickett Morgan @TDaytonPM]

You don’t need to wait for Ultra Ethernet to build a better Ethernet, and Microsoft and OpenAI proved it. $MSFT $NVDA $AMD $AVGO

https://t.co/yAWL1EfADv

[Jessica Lessin @Jessicalessin]

Good lord. Half-ish of the cloud backlog at Microsoft, Oracle, Google and Amazon is OpenAI and Anthropic????

[Ravi A Giri @ravigiri]

@FelixCLC_ I've found Codex with GPT5.4 somewhat better than Opus. For Perf Engg, they both definitely need a lot more context, tools, etc. vs. what they're capable of for say, Python out of the box. Not yet great for generating asm but much improved from even a few months ago

[Ravi A Giri @ravigiri]

@QuinnyPig You're on the right t(rack)! I'm sure anyone hosting the latest 100kW+ 'accelerator' racks have similar problems if there is any interruption to the cooling ...

[Ravi A Giri @ravigiri]

@vikramskr All I'll say is that for consideration as a head node, the criteria has more to do with interconnects, reliability etc. and less to do with ISA etc

[Ravi A Giri @ravigiri]

@MarcJBrooker Would it be fair to read it as 'the cost of any activity that is 'programmatically verifiable' will trend to near-zero? It would be great to hear your thoughts on what that could do to adoption of formal methods in SW

[Ravi A Giri @ravigiri]

@soft_fox_lad I used to diss the Intel intrinsics guide, sw dev. guide, the tools etc. (we ended up building detailed internal wikis and tools for specific use cases)... until I had to work with other arch's and now I *wish* the rest were half as good when it comes to public docs *and* tools.

[Ravi A Giri @ravigiri]

@gunnarmorling @ijuma I would suggest looking at zlib-chromium and/or zlib-ng, most optimizations that used to be only in zlib-cloudflare should now be in those libs, for both x86-64 and aarch64.

[Ravi A Giri @ravigiri]

@lemire I assume that the choice of N3 vs. V3 does not necessarily limit the possibility that Google has better and wide(r) vector width units? Perhaps @jonmasters could shed light on this.

[Satya Nadella @satyanadella]

Our newest AI accelerator Maia 200 is now online in Azure.

Designed for industry-leading inference efficiency, it delivers 30% better performance per dollar than current systems.

And with 10+ PFLOPS FP4 throughput, ~5 PFLOPS FP8, and 216GB HBM3e with 7TB/s of memory bandwidth it's optimized for large-scale AI workloads.

It joins our broader portfolio of CPUs, GPUs, and custom accelerators, giving customers more options to run advanced AI workloads faster and more cost-effectively on Azure.

[Ravi A Giri @ravigiri]

@sallywf It's simply a guess, perhaps ill informed. Consider the fact that when Nvidia releases new HW, it is generally accompanied by optimized kernels (technology demonstrators?), we will have to see how the nvcc pipeline evolves

[Ravi A Giri @ravigiri]

@geofflangdale @medawsonjr defining the initial config space (knobs and valid ranges of values) is part of the DoE aspect and requires expertise to limit it to the most 'valuable', the process/algos for arriving at the 'local maxima' is more akin to auto tuning

[Ravi A Giri @ravigiri]

@geofflangdale @medawsonjr ConcertIO (and internal Intel solutions that pre-dated it) use(d) classical ML techniques (Bayesian, Genetic algos) to 'automate' DoE - i.e, they could identify the most performant config but without context, black-box effectively. It's useful but not for everyone.

[Ravi A Giri @ravigiri]

@medawsonjr Interesting post, words to live by: "Try to achieve the best possible result, even if it’s worse than the theoretical optimum".

[Ravi A Giri @ravigiri]

@brendangregg Wish you all the best Brendan! it was a pleasure to have met you in person and a privilege to work/interact, however briefly :-)

[Ravi A Giri @ravigiri]

@emollick Most use cases seem to focus on replacing an existing process or 'retrofitting' AI into existing workflows for faster/better/cheaper/scalable results, are there good examples of truly new use cases?