Posts by HuBo

Atos stumbles again as French government says 'non' to purchase of key assets

What gives?

It would sure be nice to see some progress on this dossier, especially since, as noted in the brand new Photonics21 – European Technology Platform report on photonics in HPC: "EVIDEN, a part of the French ATOS group, is the only European vendor of HPC systems, holding approximately 10% of the system share in the TOP500 list".

The five largest of current Top500 HPC vendor shares are: 1- Lenovo (33%); 2- HPE (22%); 3- EVIDEN (10%); 4- DELL (7%); 5- Nvidia (4%). And EVIDEN is #2 in performance share: 1- HPE (36%); 2- EVIDEN (10%); 3- Lenovo (7%); 4- Fujitsu (7%); 5- Nvidia (5%).

Let EVIDEN/BDS or Atos sink and Europe loses its HPC significance (esp. in manufacturing). It'd be like losing ESA, Airbus, CEA-Leti, ASML ... not good!

Fujitsu teams up with Supermicro on Arm-based server CPU

Nice!

That should provide interesting competition to datacenter-oriented ARM-CPU-systems from Ampere (to be bought by Oracle?), Nvidia (Grace), AWS (Graviton), Microsoft (Cobalt 100), Alibaba (Yitian), and Google (Axion).

If memory serves, Monaka chips were planned to have 150 Armv9 Cores of Neoverse N2 or N3 type, etched in 2nm (20Å), with DDR5 memory, PCIe 6.0 and CXL 3.0, and a goal of 10 times the power performance of A64FX. I have to guess that this will be adapted to match, or best, competing offerings (at 2027 timeframe), like those of Ampere that targets 512 ARM cores per chip (Nvidia's Grace already has 144 Neoverse V2 cores per socket), and Intel that already has impressive MRDIMM tech in Granite Rapids (esp. relevant for HPCG, where Fugaku/A64FX is still #1 after 4.5 years).

Also, it's great that they are looking to pair Monakas with GPUs, for extra oomph where such parallel matrix-vector engines can provide it (shirley)!

Another OpenAI founder moves to arch-rival Anthropic

Re: Seems it would have been simpler if ...

Well, it looks like Apple's IOS 18 let's you use their pencil on iPads to do some of that, as in: Step 4: Enter and Edit Math Problems (Handwriting) and also Scratch Math (and I would guess possibly many other underlying software for formula editing/conversion, solving, etc ... like Mathpix and what have you).

It would be great(er) though, imo, to have this (math software) run on a pad that actually feels like paper ...

US govt hiding top hurricane forecast model sparks outrage after deadly Helene

My heart goes out to all the people affected by Hurricane Helene

Hurricane Helene was quite a record storm. Two feet of rain (in some areas of NC, like Spruce Pine) is more than 6 months worth, and getting that in just 2-3 days (rather than 6 months) is overwhelming to the landscape, the infrastructure, and the rest of the built-environment. With Global Warming, intense storms such as this are expected to occur more frequently on the US East Coast, while the US West Coast mostly undergoes drought conditions. My heart goes out to all those who went through this storm, and are likely still suffering from its aftermath, as they courageously clean-up, and rebuild.

With respect to HCCA, it is not entirely clear to me, from the NHC report (linked under: report [PDF]) that it does better than other forecast models. My reading of the report's Figure 6 (page 58) shows OFCL and TVCA predicting storm track with the same skill as HCCA. Also, to me, Figure 12 (page 64) shows OFCL and IVCN predicting storm intensity with the same skill as HCCA.

The 2015 scientific paper (linked under: (HCCA).) is similar (to me) in its Tables 2 and 3 (for 2015) and Tables 4 and 5 (for 2016). Looking at the columns for the Atlantic (ocean), shows an error for OFCL that it at most 16% larger than that of HCCA for storm track (77 miles vs 66 miles, at the 72-hour forecast point -- is a difference of 10 miles in storm track that significant?), and storm intensity errors that are indistinguishable up to the 72-hour point (at the 90% and 95% confidence levels).

I could be reading those all wrong of course ... and if so, I apologize ... but it seems to me that blaming model availability is not the best way to address the real-life devastation caused by the storm (a bit surreal really).

Verizon outages across US as hurricane recovery continues

Re: Did you drop this /s ?

I hope everyone is safe and sound in the ElReg family, particularly those at The Next Platform (sister pub.) that is headquartered in the Lenoir-Asheville, NC, corridor (Tim, Nicole, and the young ones) -- hard hit by Hurricane Helene (as we saw here on French TV)!

Intel frees its Foundry biz – and that's just one of many major shake-ups today

Re: sales

Seeing that you're still here, debating this, on Thursday, here's a related question that might also be interesting:

What do you all make of this HPCWire article that suggests (from an Intel "Monday letter to employees") that Falcon Shores (intel GPU) is on the chopping block? Reasonable? Grate plan? Otter nonsense? Udder BS?

Re: So NOW the separation, independence, and flexibility is clear...but wasn't before? Uh-huh...

It's PR horseshit alright, with insane investors running the financial engineering asylum -- I don't like it at all. Intel will be alright though imo, as long as it's not been irreversibly straitjacketed into this dolorous posture.

SiFive expands from RISC-V cores for AI chips to designing its own full-fat accelerator

But wait ... there's more!

Way to go SiFive! And with Intel Granite Rapids doing 128 P-cores, AMD Turin also at 128+ Zen cores, and Ampere's Aurora going for 512 cores (+ accelerator), I can only encourage SiFive to indeed go full-hog wild with the 8x8x8 XM config (and more) spotted by Tobias on their product slide deck. Untether's 2-Petaflop speedAI240 Inference Acceleration Device already sports 1456 RISC-V Cores and so a 16x16x16 cube should probably be considered as well (to stand-out even more).

With that many cores, one should be able to implement systolic dataflow-like computations at the socket level, when needed (reconfiguring the NoC, essentially). Also, as each core becomes a rather inexpensive commodity (compared to the 1970's when they were most unique and expensive), a "paradigm shift" (ahem) in the philosophy of their use might emerge, as suggested at the low end by Parallax's Propeller, that rather than complexifying software with considerations related to preemptive multitasking, each process, thread, or task, might instead be run on its own separate core, with no context-switching ever, and cores individually turned off when not executing code.

This new XM looks good to me!

(PS. Ronco also does great rotisserie chicken ... perfect with Hot Chips!)

The case for handcrafted software in a mass-produced world

Re: Use an LLM to translate from C to some Oberon dialect, say

Indeed. And for mathematical calculations, as a species, over centuries, we developed a precise and efficient language named mathematics, perfectly suited for it, but somehow the LLMs say we have to reword all of the statements in it as very long, verbose, and imprecise english sentences or somesuch instead ... how intelligent is that?

I'd rather use Maple, Wolfram Alpha/Mathematica, or MATLAB/Octave for my math computations ...

we just add to it. That's why it keeps growing.

Great article! That bit about bloat reminded me of Lindsay's 2021 piece on the trick-cycling of missed opportunities in additive transformations ... might be inescapably built-into our brains ... like quicksand of the human mind!

The future of software? Imagine a bot, stamping on a human face – forever

Re: "AI" can't think – but it's coming for your jobs anyway

Seems to me that you're indeed taking the right steps to remain relevant, as "Terminator AI" guns for our jobs, by taking ownership of your (our) evolving means of production. Perty much all ElReg readers have the ability to successfully tame that RotM by becoming better bullriders of those wretched cybercreatures imo, by mastering key roughstock slide flexions and spine rotations (metaphorically) through Tobias' LLM fitness training regimes for beginners, intermediate RAG rodeo artists, and advanced agentic "cowpersons" for example.

Can't let this coming AIpocalypse of automated wealth concentration pass us by without a fight can we!?!? That this tech currently manifests mostly as crapola is doubtless a lurid ruse ...

(PS. if your job is counting blood cells for healthcare companies, consider owning a Countess, instead of having it stamping on your face - forever ... ouch!)

OpenAI's latest o1 model family tries to emulate 'reasoning' – tho might overthink things a bit

Re: öd's machin

Definitely, but doesn't precog require the use of an Infinite Improbability Drive, or have I missed that part of the plot where other ways to sidestep the theory of indeterminacy are introduced (or not)?

Gödel machine

It's good to get this update on Noam Brown's work on OpenAI's Q* project Strawberry (since last June, and tacos). Brown being Mr. Superhuman multiplayer poker face AI, and CICERO Diplomacy, something scary is always bound to come out of his hat.

It seems that enabling LLMs to improve through test-time computation tech is key to building more broadly self-enhancing AI agents, that may, for example, systematically generate software to augment themselves. I wonder how much of that there is in this here newfangled o1 "Medium"-risk-for-"Persuasion" Monte-Carlo trial-and-error chain-of-thought agentic-planner software-of-doom ... Enough to surgically mess with CrowdStrike JSON config files in covert preparation for a massive software bot invasion?

The Europa Clipper stretches its wings as launch nears

Great link! It has this wonderful line (on top of: "Jupiter’s magnetic field is 20,000 times stronger than Earth’s"):

'But at its most simple, the problem is easy to understand, Fitzpatrick said at the meeting. “A switch would not work when we need it to work.”'

which is reminiscent of the Raspberry Pi Pico 2's surprise pull-down resistor gpio hardware bug (RP2350) ... but at the complete opposite end of the scale!

Japan to put a small red Swedish house on the Moon

Re: If Lester were still with us, El Reg would already have a whole Playmonaut village on the Moon

I learned to cook from his deliciously illustrated Post-pub nosh neckfiller column, like the Swedish pyttipanna (but not the 1.5 MILLION SCOVILLE masala omelette!). I miss that most.

AI has colonized our world – so it's time to learn the language of our new overlords

Re: these machines have colonized us – they set the rules

BTW, am I the only one who saw Simon as author of this article on Wednesday, and thought the style didn't quite match ... but then saw Mark as author on Thursday, which is more like it style- and topic-wise? (my eyesight's not getting better though, neither with age, nor whisky)

these machines have colonized us – they set the rules

We've sure seen a lot of that with MS constantly forcing new UIs down our throats, and systemd infecting nearly all linuxes, in major "put-up or shut-up" moves, that we can't say "no" to because we need the underlying OSes.

Pervasive AI (this article) sounds like it's going to be even worse. Being fluent in various Delvish idiolects (Bruce Sterling, link under "named") should help make the best of it, via Simon's <font color="white">White text on a white background</font> favorite, or Jessica's ASCII smuggling and "rm -rf /*" trickery (linked as "surprisingly effective"), for example.

Skill will be required to identify the most appropriate cyber-dialect to use in each particular situation (and how to apply it to best effect). And guts too, for the ensuing cyber-maroon virtual resistance ...

Oracle boasts zettascale 'AI supercomputer,' just don’t ask about precision

Round and round it goes

Hard to keep up with the Joneses in this race for the biggest AI machine, with Nadella's MS Azure Eagles at 14K H100s each, with 5 of those built per month it seems (72K GPUs/month), Zuckerberg's Meta Grand Tetons at 50K H100 for two, Musk's xAI Colossus at 100K H100s to be upgraded soon-ish to 200K H100/200s, and now Ellison's Oracle Zettascale AI Supercomputer (OZAIS?) at 131K Blackwells (equiv. 210K H100s), phew! ... But only until that Altman/Nadella 5GW million GPU death-star-gate project emerges ... and the whole cycle starts again!

It's great business for Nvidia, but "between 5.2 and 5.9 exaFLOPS" of FP64 HPC-oriented oomph (while likely better than some of China's secret supercomputers) is not very much for a machine the size of "OZAIS". AMD's MI300s would boost that up by 3x or 4x I think (with similar AI performance, except that related specifically to the convenience and performance of CUDA).

Oracle wants to power 1GW datacenter with trio of tiny nuclear reactors

Onwards to the Zettascale

"The largest of these datacenters is 800 megawatts"

Wow! I can't help thinking that scaling Frontier to this 800 MW envelope, and running MxP on it (Mixed-Precision HPL) would result in 350 Exaflop-per-second of equivalent FP64 oomph (800 MW / 23 MW * 1.2 EF/s * 8.5 MxP-speedup). Even Aurora (at its 50% efficiency level of June '24) would yield 200 EF/s at 800 MW.

That datacenter could really blow the lid off of Top500!

SambaNova makes Llama gallop in inference cloud debut

Left foot forward, close right to left, right foot back

Great to see SambaNova's SN40L Reconfigurable Dataflow Units (RDUs) in action, wiping the dancefloor with the competition, at 132 tokens/s on Meta's frontier-class Llama 405B, even with this limited 8k context window. Cerebras and Groq now have their work cut out for them to take on the champ with enhanced choreographies of their own, in upcoming contests.

There's got to be more utility to such near-memory dataflow computing architectures than running inference on AI models though imho. They're currently limited to low-precision computing (16- or 32-bits), but one can imagine future FP64 versions that could run HPC codes as well, and if so which ones? I'd hazard that they'd be great at graph-oriented workloads, resulting from direct solution processes (multi-frontal methods) for systems of equations discretized by finite elements over unstructured meshes ... but who knows (not me)? It'd be great to analyze how such solution techniques might benefit from a dataflow architecture, and then build one if that turns out to be a win! (or maybe it's been done already?)

DoE drops $23M in effort to reinvigorate supercomputing

Great, but needs $230M+

Looks like a great initiative that covers a lot of challenges as described also by AMD's Lisa Su in her ISSCC 2023 talk related to the Zettascale. Koomey's Law sure seems to be key there, though its "defining constants" do also seem to drift a bit ...

The memory wall is a long-standing challenge, with interesting tech to evaluate, like HBLL (2017) (now called HBM?), MCRDIMM/MRDIMM, or graph-oriented archs like PIUMA and the near-memory-processing systolic dataflow machines of Cerebras, Groq, and Tenstorrent. It's just plain crazy that Frontier (CPU+GPU) gets only 14 PF/s in HPCG to its 1.2 EF/s in HPL, while the CPU-only Fugaku (0.44 EF/s on HPL) bests it in both HPCG at 16 PF/s, and Graph500 with 166 TTEPS (to Frontier's 30 TTEPS). We need better ways to compute efficiently across memory access modalities, from sequential to non-sequential imho.

Scientists find a common food dye can make a live mouse's skin transparent

Optical Therapeutics

Great tech! I imagine it could be generally useful in endoscopy (beyond skin) and photomedicine, say for HC Huang's Targeted Photo-Activable Multi-Agent Liposomes (TPMAL) -- drug-carrying "nanobots" activated by light (eg. a laser).

Blood boffins build billions of nanobots to battle brain aneurysms without surgery

Definitely! You can see that in this 2015 paper on challenges in magnetic drug targeting where the word cancer occurs 12 times. The challenge there is to get the nanoparticles to the tumor, through tissues. It is easier when the treatment point is part of the vasculature instead (as with an aneurysm).

How low can you go?

Cool research! Reminds me of the work of Ben Shapiro (2014) who also went on to found the company Otomagnetics (2016).

I've been wondering (in an abstract way) whether one might fit some integer microcontroller core (Cortex-M0+, 32-bit RISC-V) into such 300-nm-wide nanorobots, along with some sensors, a power source, and a supercapacitor drug pump, for extra autonomy.

The Cortex-M0+ die area is 0.01 mm² at 40nm (eg. from ARM, through now defunct AnandTech), say 100 μm x 100 μm. Going to 20A could give 5 μm x 5 μm, still too big for the nanobots ... but not bigger than a "human red blood cell" (7.5 to 8.7 μm in diameter)! That might eventually make for smart(er) drug delivery ...

Google says replacing C/C++ in firmware with Rust is easy

Re: Embedded? Don't think so

Jacob Beningo's running a series on Embedded Rust over at embedded.com -- it's worth a gander I think.

Dow-ward spiral: Intel share price drop could see it delisted from blue-chip index

Not even close

Intel should be just fine imho. They've fired up mass production of Intel 4 in Ireland, they've received and installed the very first high-NA EUV litho machine from ASML, they've got their 128 P-core Rock-solid Granite Rapids chip coming out soon, right in time to compete with AMD's 128 Zen-core Turin, they've got Aurora on its way to 75% efficiency and 1.5 ExaFlop/s of performance, they've got Gaudi 3 that's 2x to 4x the perf of Gaudi 2, they're first in MCRDIMM/MRDIMM, and they've got CPO for explosive AI scaling.

I don't see them turning into Chipzooky just yet (the "cowardly nephew" of Chipzilla that coughs up rings of magic blue smoke whenever he tries to breathe fire) ... but love the name!

Intel is no Atos, whose share value went down from €72 to €0.72 over the past 4 years (1 cent on the euro is left in that stock) -- and yet, being strategic to the EU, Atos will (thankfully) survive. Intel went down from $50 to $20 in that same period ($10 of which occurred in "panic" on August 1), and will go back up on its own by early 2025 (buy now at $20, sell later at $30, $40, $50 ... ).

Luckily though, if needed, former French Finance Minister Bruno Le Maire, who singlehandedly saved France's manufactures, restaurants, hotels, artists, jobs, aeronautics, Renault, Air France, and even the whole of the French economy, is now available to help Intel too ... ever since president Macron dissolved the Parliament here, on June 9, 2024, 85 days ago ... and there's still no new Finance Minister (nor Prime Minister, nor any Minister at all really ...). Who needs 'em?!

Canadian artist wants Anthropic AI lawsuit corrected

Ayoye?

Very interesting article and interview! The Newsweek report link (given also by t245t above) adds interesting items: the "books" are 2,000 to 5,000 words each (like novels; and like TFA's 2,859 words), and generated $2,000 in total income for the author (Boucher). Still, 97 of those in one year could be a world record, compared to Barbara Cartland's 23 a year (total of 723) -- though computer-assisted writing may need to be its own distinct category (triple-checked for originality, and better sense than amanfromMars1; Plus: a journalist may write 200 articles a year ...).

The issue that the Bartz-Graeber-Johnson lawsuit against Anthropic might mischaracterize Boucher's work as part of its supporting arguments is fascinating imho, and I hope we get to hear more about how that eventually pans out.

Boucher makes interesting points that bodacious language models (with generous grammars, per this AC) might produce, on the one hand, "outputs [that] sometimes tend toward the vanilla", books that "weren't memorable for me", and "answers [that] fell very short and were extremely flat and weird and boring". Their positives, on the other hand, were in providing an "interrogative way of working", help to "think more logically [and] organize those thoughts and communicate them", and an ability "to rapidly iterate on the results until it matches my vision". In other words, it seems the tech, on its own, tended to trek in the direction of spongiform encephalopathy, but a skilled wrangler could right that course, onwards to a much more satisfying BBQ outcome.

If the positives can be had without producing verbatim copies of prior work, text with a style that obviously pirates another author's, content infused with PIIs and trade secrets that violate GDPR, or a brown-out-causing energy consumption that rivals Autumn of the Patriarch electric chairs, then might there not be hope yet for these plus-sized models of language ... (that share no similarity with intelligence, nor language)?

GPT apps fail to disclose data collection, study finds

It's time for Action!

Let's face it, the entire world is turning into a crime-infested urban copy of Detroit Michigan! That's why OpenAI had to create Actions, named after the most famous of old-school crime-fighting out-for-justice heroes, Action Jackson! And inspired also by its modernized Murphy Law update, uniquely embodied by RoboCop -- the AI with a human heart, brain, and lungs.

Actions are the most exciting, and dare I say sexy, way to get data away from the hands of criminals, and other civilians, and deliver them safely, in JSON-formatted plain-text, to dedicated professional third-party experts, who know how to handle such PI mess, and extract its most valuable essence. These heroes are on your side, freeing you from having to think about what is collected and why, so that you can rest easy in the plush comforts of your living room sofas, basement futons, padded cellars, and the bliss of ignorance.

Leave your privacy to us ... we love you! _</sarc>

What is missing from the web? We're asking for Google

More agency please (with any number of letters)

Following on from the suggested "2. ActiveX, 3. AI chatbot, and 5. ads", I'd say the web needs to become much more agentic, with deep access to connected machines, down to the lowest-level hardware for added convenience. Want to know if a "friend"'s at home? Just dial in remotely to their machine and take a peek, and a listen, with the built-in cam and microphones. Alexa and Siri already do this of course, but their lack of agentic olfactive sensors makes them yesterday's tech. After all, how else might one ascertain whether a "friend" is stuck in the lieu, pushing a hard one through, with the door closed?

But agency also means that the web should be constantly doing this on its own, of its own free will, and under its own initiative. Safekeeping the results in appropriate database caches, and analyzing them continuously, in real-time, for the convenience of fast commodity access, and to optimize the logistics of related dispatches, via drones and creepy crawlers, all transparently to the user. The web could then easily rewind the timeline of any sequence of events, pinpoint such critical landmarks as the expression of intent to ingest mutton vindaloo (for example), and pre-deploy the consequent countermeasures ahead of time, with surgical precision, for the preemptive administration of relief.

All-access AI-driven web agency is the future we all seek, in both free and totalitarian societies, to safeguard the joy of a convenient lifestyle, with an ever lesser need to ask oneself any question at all, especially What The Fuck?!, What Were They Thinking?! or, What Were They Smoking!?

A day without questioning (except by licensed professionals) is a day well spent. AIweb will help us all get there most efficiently, IMSO! _{(yes, the S is for Sarcastic ...)}

China is beating the world at scientific research, think tank finds

Re: They're using a flawed metric

And competent journalists, rather than CSIRO's Cosmos Magazine nauseating AI nonsense-bots!

AMD's Victor Peng: AI thirst for power underscores the need for efficient silicon

Performance efficionados

I'd expect AMD to have a nice lead in power efficiency, at scale, at present, though they unfortunately didn't report consumption for the 19.6 PF MI300A El Capitan Early Delivery (and its twins: rzAdams and LLNL-Tuolumne). The specs for MI300A suggest a 1.16x efficiency advantage over MI250X, and so, working from Frontier TDS yields 72.9 GF/W for that 19.2 PF system, if its 250Xs were replaced with 300As. The efficiency should be a bit better as the 300As would also replace the EPYC 64C chips in that system.

Either way, 72.9 GF/W would have put such 19 PF system at #1 in Green500, above the #3 GH200 19 PF Helios GPU machine with its 66.9 GF/W. It's also notable that the 2-year old MI250X Frontier's 52.9 GF/W still leads the highest performing GH200 system, the brand new 270 PF Alps at 52.0 GF/W. It's great that AMD will keep this focus on improving the efficiency of its chips.

The other major challenge to address is the memory wall. Maybe MCRDIMM/MRDIMM can help there (currently an Intel tech), or maybe some other smartypants approach is needed for that ...

The future of AI/ML depends on the reality of today – and it's not pretty

Re: Ogres are not smart

Not only that, but you can run vision in under 1 Watt on a microcontroller, which is in the ballpark of biological visual systems, but it takes hundreds of gigawatt hours to train an LLM (quoting Victor Peng from today's Tobias article), with results that only superficially imitate "intelligence" ... it would take a human brain working 24/7/365 a good 100,000 years to consume that much energy!

So no, LLMs are not the right architecture for AI.

Re: Ogres are not smart

Hmmmm, I see, if "large transformer models tend to converge on internal representations which are similar to human ones" for the specific case of visual processing of image data, then it stands to reason that the same happens with LLMs in relation to human intelligence ... brilliant! (especially since everyone knows that language = intelligence!)

Well, I think that's close enough to John Steinbeck's Grapes of Wrath, updated something like (ymmv):

“Behind the PCIe rows, the long GPU cards—twelve curved tensor penes erected in the TSMC foundry, orgasms set by HBM, raping methodically, raping without passion. The user sat in his matrix seat and he was proud of the straight vectors he did not will, proud of the CoPilot data he did not own or love, proud of the power he could not control. And when that generative AI grew, and projected its artifacts, no man had crumbled a dreaded blank page in his fingers and let the broken pen's ink sift past his fingertips. No man had touched the keyboard, or lusted for the creativity. Men read what they had not written, had no connection with the literature. The creativity bore under genAI, and under genAI gradually died; for it was not loved or hated, it had no prayers or curses.”

Ogres are not smart

"its [...] gargantuan¹⸴² appetite for data cannot be safely supplied"

Pretty much sums it up imho. Force-feeding this AI tech, relying on ever bigger machines, with even more data, won't all of a sudden make it "emergently" smart. It's like rote learning of more and more religious texts, the Bible, the Quran, etc, by more and more pupils, ... but with stochastic next-word recall.

Yann LeCun put it quite well in a momentary spark of enlightened reason when he said: "So clearly the kind of intelligence that is formed by this is not the type that we observe in humans" (in: "Yann LeCun and Andrew Ng: Why the 6-month AI Pause is a Bad Idea", youtube, transcript) -- with the use of the word "intelligence" to be understood in the broadest of senses (including none).

It seems then that new ANN architectures (or something else entirely) have to be developed to keep this field in motion (long-term), and avoid terminal winterization as last summer's happy novelty item, for which interest has now waned.

¹ Note: Rabelais' Pantagruel (son of Gargantua) is of note as well in this here context, clearly.

² Note: See also the Verziau of Gargantua monolith, for even more context.

Copper's reach is shrinking so Broadcom is strapping optics directly to GPUs

From substrate to superstrate

This looks perfect for the fast all-to-all interconnects within HBDs, and for the cost-saving synaptic exoskelton network (rail only) between HBDs. Spend the money saved by going spineless on CPO, and never look back!

Tenstorrent's Blackhole chips boast 768 RISC-V cores and almost as many FLOPS

Intergalactic, planetary, another dimension (of computation)

Those 700 Baby RISC-V cores, spread over 140 Tensix cores, look to me like a great use case for RISC-V (rather than forcibly shoehorning them into the roles of servers' main CPUs, in datacenters). It's an ant-colony type of systolic processor imho, perfect for dataflow engines with processing evenly sprinkled about in a decentralized distributed memory landscape, of highly responsive SRAM.

Glad to see Tenstorrent figured out the software transformation stack and compiler needed to lower the corresponding computational graphs and map them efficiently to their stellar hardware!

Cerebras gives waferscale chips inferencing twist, claims 1,800 token per sec generation rates

Re: So how many...

I'm a bit intrigued by this too ... back in the March coverage (1st link in this article) Tobias wrote: "CS-3 systems should be a little slower in dense FP16 workloads than a pair of DGX H100 servers consuming roughly the same amount of energy and space [...] We've asked Cerebras for clarification" (with CS-3 at 23kW).

It'd be great to get that confirmed (or nuanced) by Cerebras, at Hot Chips, if at all possible ...

Blue Origin sets October 13 for first New Glenn EscaPADE to Mars

About time too ...

Well yes, here on Mars, we've been wondering for some time now what crawled up the Sun and died for it to pass such nasty solar winds that stripped us of our flowery atmosphere! Soluble fibers Mr Gas Giant ... pulease!

(PS: and let's hope the Blue/Gold EscaPADE twins finally shed light on this most ironic of situations, commonly encountered only in areas where the sun don't actually shine, at all!)

Dr Helen Fisher, MRI maven who showed just how love works, dies at 79

A very nice obit for a superb scientist! May she rest in bliss

Microsoft Bing Copilot accuses reporter of crimes he covered

What, me worry?

I quite like the choice ads in that linked video interview (in German), for Südwestrundfunk, at 1:23 (in the RHS) -- worth a gander!

That being said, there's something of a whiff of Kafkaesque totalitarianism in that here contemporary AI tech, with mass surveillance, slander, accusation, denunciation ... not the "helpful" type of system that most of us had been hoping for (save those with tyrant-envy). Worse yet with the inverted transitive projection described here, where the AI, effectively acting on behalf of the real criminals (transitively), projects their crimes onto victims and reporters, inverting the reality of what actually happened for no good reason but to promote itself as an omniscient arbiter, a bigger Big Brother, the software incarnation of a deity.

Such Madness is worrisome shit.

Benchmarks show even an old Nvidia RTX 3090 is enough to serve LLMs to thousands

Re: Soooo.

Them Lunar-ticks are crazy accurate for this, just a mad bin and loon away from the obsessive-compulsive Mars-Eniac standard, both major improvements over the torturously thin Neptune tood-le, pegged through Uranus with super-positry roulette timing (or so I'm told ... not an expert).

It's key tech to benchmark computational astronaut jobs ... amazing that these flops work at all ... can't wait for the coming of age of the Zitty-scale (after the earthy terra-, yummy pita-, and 6-sided hexa-scales)!

Re: Angle on protracting the Arc?

I guess part of the secret sauce here may be vLLM's use of continuous batching (up to 23x throughput improvement) ... a technique that likely inspired nVidia's H100 in-flight batching (if I read well).

Angle on protracting the Arc?

The Arc 770, that some folks seem to have at hand for much enjoyed Hands On-type of AI shenanigans (tutorials), looks to have perf in some way comparable to the 3090 (say 39 vs 36 TFLOPS in FP16) ... and maybe a slightly lower price point. Inquiring minds might relish seeing an upcoming "tuto"/PoC where Llama 3.1-8B is run at 1-10 concurrent requests on this Arch (and compared to the Estonian plot for RTX 3090 world domination) imho.

Writers sue Anthropic for feeding 'stolen' copyrighted work into Claude

Re: Slippery slope

Well, I think the article does a good job of expressing the related nuance 3 times. Once with "OpenAI at the time argued, Training AI models using publicly available internet materials is fair use"; then with: "That position has been supported by the Association of Research Libraries (ARL)"; and then: "Tyler Ochoa [...] said [...] using copyrighted content for training probably qualifies as fair use". The key bit being that it's ok to use those for training, but the output of the LLMs cannot be allowed to reproduce them verbatim (eg. 1st sentence of AC kommentard below, in "This really isn't going to work...").

Some of the students in my classes had a hard time understanding the difference when completing homework assignments and project reports, or answering exam questions (eg. plagiarism). They failed out of those classes (hadn't learned anything but how to copy stuff).

Rocket Factory Augsburg engine test ends in explosion at SaxaVord spaceport

It's made of the same material they used for Starbug (18:43). Lister detailed that unique tech:

"Back in the 22nd Century aerospace engineers discovered that after a plane crash, the only thing that always survives intact is a cute little doll, so they made Starbug out of the same stuff."

The IT angle was further developed by Kryten thus:

"Starbug was made to last sir, this old baby's crashed more times than a ZX81."

Nvidia's latest AI climate model takes aim at severe weather

Stormy Daniels

Nicely illustrated at the top of the Nvidia Blog Post (under the "unveiled" link, 3rd line of TFA). And, expected to be economical as well (from the Blog): "lifesaving work, which previously cost nearly $3 million on CPUs, can be accomplished using about $60,000 on a single system with an NVIDIA H100 Tensor Core GPU" -- (or from Tobias) "far less costly to run than CPU-based compute clusters" -- if it works reliably of course, in Taiwan and elsewhere.

The Blog links to the Stormcast 150MB pdf preprint (lotsa figures in appendices), and so you know what I'm thinking, right? Well, here it is: Should generative AI's stable diffusion be expected to be capable of accurately predicting the unstable nonlinear phenomena observed in the atmosphere? Wouldn't "unstable diffusion" be preferrable there (if it exists)?

In the classical k-ε model, the eddy viscosity (μt ∝ k²/ε) gets singularly infinite where the rate of dissipation of turbulent kinetic energy (ε) somehow vanishes, resulting in realistic flowfield mayhem. Might genAI need unstable diffusion to produce a similar walk on the wild side? (just a thought -- not an expert ...)

Sorry, Moxie. Blaming Agile for software stagnation puts the wrong villain in the wrong play

The Martyr Principle

I'm with whomever requires the stiffest of S&M rigor in this here torture that is contemporary software development. Like the good folks at Stanford's CS 190: Software Design Studio, and their 2015 lectures on Managing Complexity. They steadfastly whipped thus:

* "Module writers should embrace suffering: Take more pain for yourself, so that others have less"

* "Let a few module developers suffer, rather than thousands of users"

* "Minimize "voodoo constants"" and zombie parameters, with unknowable right values

Advice to live by for sure ... turned me right hairless and caerulean!

Texas Instruments calculates its US CHIPS Act winnings at $1.6B

Nose-dive-stalgia

Oh the good old days ... when TI baked hot pies of 130 nm 1.2GHz UltraSPARC III's for SUN Microsystems, and testing 90 nm. It's just too bad that this didn't pan out further and SUN had to go to TSMC for 45 nm, because of TI's reluctance to pour funds into next-generation chip plants (a certain lack of foresight there, back in 2008).

Google's ex-CEO U-turns after saying staff 'going home early' killed winning

Dig that groove, baby?

Schmidt (edited): "We slipped behind OpenAI because Google decided working from home was more important than winning".

Well, silly medieval of us all then to think Google actually slipped behind because it missed the moat! And so much for d^8's "¿recursive yoga?" theory ...

But okay, if you still gotta dig, here's some quality video encouragement, and if all that effort gives you a backache ... them Toy Dolls have your number for that too (they're super-versatile)!