are you working in earplugs? :)) even with 128 gigs of ram it must be super noisy.

I've put Opus to it and it says it will take 3-4h to do the process to the new weights. hoping it works!

you don’t need new weight. try vllm-moet from github. it will autogenerate 2-bit plane.

What kind of tps are you getting?

same. been running it in a dgx spark and it slaps

Google TPUs are built around a 128×128 systolic array of multiply-accumulate (MAC) units. Trainium 1, Trainium 2, and Inferentia 2 also feature a 128x128 systolic array.

You learn something every day. Today, it was the term "systolic array": A systolic array is a specialized grid of simple, interconnected processing units designed to execute parallel data operations—like matrix multiplication—by rhythmically passing data directly from cell to neighboring cell without writing intermediate results back to main memory.

The term comes from the biological word systole (the contraction of the heart pumping blood through the body). In a systolic array, data "pulses" through a network of processing elements on every clock cycle, driven by a global clock beat.

Just because scales are grouped by 128x128 tiles, does not mean you need a single compute tile that large. It works completely fine to process it with multiple smaller tiles that get given the same scales, like how this works on Hopper and Blackwell today

Huawei Ascend NPU

Blackwell. You could do it with wgmma on hooper but then you'll only be able to run 1 CTA per SM. There are cases where this is OK, but most commonly 128x128 mmas are primitive in tcgen05.mma i.e blackwell. The fundamental reasons is that the systolic array accelerator (TMA) until blackwell wrote to the cuda core registers themselves so you were limited by the register file size of the SM[SP]. In blackwell there's a separate TMEM where the mma unit stores it's output. You can even do higher 256x256 and such using a hardware feature called 2-CTA MMA, this is essentially them letting neighbouring pairs of SMs co-operate and access each other's memory.

As for sibling comments, huawei ascend is more of an NPU-style architecture where you can easily have much bigger MMAs as primitive. But you usually don't anyways for many reasons.

Huawei Ascend presumably?

Same. And I have come to use OMP (oh my pi) agent /advisor mode to put a 2nd model on the case (also mid-size one), reading everything. It can not block anything or change anything - just inserts comments in the text stream with 1 turn delay. Good portion of the time it's quiet. I'd say 1/2 of the time it's got something to say. About 2/3-rd of that the 'advice' is insubstantial or about something not-quite wrong. The good thing is the main model is confident - checks and then it stands its ground. Have not noticed it turning a right into a wrong b/c of advisor false alarm. And in 1/3-rd of the advice, it's a genuine defect teh advisor noticed, the main model works out a fix. This is my approximate feeling just observing the process, have not got collected the data. Afaik only OMP has advisor mode. Agent pi has plugin pi-omplike-advisor. For agent Hermes I had them code me an /advisor plugin (for now -0.1 old v0.18.x; yet to upgrade it to latest).

The problem is picking between models. I do not want to spend my time switching models and trying to decipher which model should be used for what. Maybe that's just a me problem that I need to figure out.

what plan are your 'fellow software engineers' using? I have a hard time even using up the Fable part of my allowance in a week of coding.

usually I am - Codex, make pi with deepseek to do something.

The issue for me is the data privacy if you're using their hosted prices, because you cannot opt out of data collection (and I'm not sure how you'd legally follow up on it if they did offer it, but didn't actually follow through). That means all of my code is being retained for training.

I've used it in some open source code though, and loved how fast it was.

My mind is changing on how valuable my code actually is though... it's the complete picture, how it's put together, the design, the UI, the attention to detail that's the real value.

I'm not a heavy user of agentic coding, but still use them quite a bit for some automation here and there. I've been going around shopping all the ~$10 subscriptions and I finally settled on openrouter + ds4 pro. The more intensive days cost me $1 and I set a $2 weekly limit which I've never hit the past 3 weeks, to me it's way cheaper than most subscriptions and I don't have to worry about maximizing my weekly quota/resets.

I wonder if OpenAI knew and deliberately price dropped to front-run the news cycle in their favor.

Chinese are pretty pragmatic. Even if they produce more expensive chips and memory, they are still going to focus on value. But who knows, maybe India will step up again like they did for Y2K 3 decades ago.

I agree. A meeting transcript posted recently from DeepSeek's founder suggests they're going to reach for larger models as soon as they can, and not look back. The model is already bigger than my machine can handle, though (I cannot put up with 25 t/s after getting used to over 100), so I'm not impacted.

We'll make sure there will be enough regulations to have neither the electricity to power it or the companies to build it.

Sorry we a focusing on leather handbags there.

Seems incredibly unlikely, doesn't it? (I hope for the same thing but it feels almost like wishful thinking!)

The EU is not a state or some federation. It's like a NAFTA+ It unfortunately doesn't have the risk averse VC system (or pragmatic state driven like China) ecosystem to support the scaling required.

But if models will be commodities, maybe that won't be such a terrible thing longterm?

Personally I don't. Wasting enormous amounts of resources on a technology that doesn't make money, can easily be copied, and eats up everyone's energy while at the same time can be downloaded over the internet is kind of stupid.

The fact that China could catch up in what, two years, shows that this is a commodity.

EU is out of game and can't be back anytime soon.

Incursions into other nations "just because"? Economic wars, and support for foreign governments or armed groups that caused harm or instability to the entire planet? You know? Planet Earth?

Sanctions and trade policies that hurt ordinary people? Cultural dominance through Hollywood (Holy Wood, again with the sick ruling elite and their witchcraft crap), social media, brands, the poisoning of food and water on global scale?

Etc? Etc?

Most people don't hate "Americans" or their companies, they are just against your ruling elite. Live one day outside the united states and understand how the planet works for the rest of mankind and you might start to grasp on reality.

And yeah, China will catch up with hardware pretty damn soon. Jesus, you guys bought your radios from Japan not 50 years ago. Sorry, can't dominate the world's economy forever. Never happened in history.

No matter how much black magic, dark rituals, sick crap and sacrifices people on the top throws at the wall.

We're humans, after all.

So yeah, it is not against America itself or American companies, but against the ruling class. So naturally , people get excited about China entering into the match!

3...2...1... Fight!

We'll never have fewer open-weight models than exist now. They won't suddenly disappear when labs stop publishing new ones. In fact, people will keep improving them and will keep distilling new frontier models into existing open-weight models.

Then non-Chinese labs can reopen, or they could offer lower prices right now not waiting for this event.

But imagine if Chinese labs would lose. Then there will be no open models, and the prices for closed models would be raised to the maximum.

My conspiracy theory is that this is the new space race, and the CCP encourages this to show the world what Chinese engineers are capable of, and tank the Anthropic/OpenAI valuation bubble as a desirable side effect.

As opposed to closed source American models where the federal government threatens your livelihood unless you build data centers to extend pax silica?

This will never happen btw.

I'm neither pro China, nor pro US. I'm pro open weights models, and I'm pro cheaper hardware. At this point I don't see any american frontier labs releasing SOTA open weights model, and I don't see ASML/Nvidia/Samsung monopoly getting any competition from anywhere apart from China in the near future.

And what is their nefarious plan after we get "addicted to cheap shit"?