rerx

Many shows have hit their strides in later seasons. IIRC I found Breaking Bad a bit boring in season one. (Granted, probably the writing was already better than season one of Foundation.)

GPU cluster scaling has come a long way. Just check out the scaling plot here: https://github.com/NVIDIA/Megatron-LM

External NVSwitch (for more than 8 GPUs) isn't available for purchase yet. But even now you typically buy your A100 or H100 training GPUs in servers where each GPU has an individual 400 Gbps Mellanox networking adapter, then connect them in a sophisticated switched network that provides full throughput between all GPUs of the cluster. The deployment is not that custom, typically you would let Nvidia and your vendors implement their reference design for you: https://www.nvidia.com/en-us/data-center/dgx-superpod/ Then run an open-source software stack on top of CUDA etc.

I don't know how well the TPU hyper-torus interconnect performs, but the networking topology seems to be less general than switched NVLink or InfiniBand.

It's strange that Netflix would do that. In Germany, they pioneered consistently offering original audio. (For instance, Amazon Prime Video was available earlier than Netflix here, and originally most of their catalogue was dubbed to German exclusively.) Plenty of services where you can rent this movie also offer it in English (OV): https://www.werstreamt.es/film/details/2097679/the-green-kni... So it may very well be some weird licensing deal where they only get to offer this specific film for streaming without a surcharge if they stick to dubbed audio.

Meta question: Why has this submission moved off the front page so quickly?

> At that point you need to call customer support to clear your authenticated devices. Hassle.

Yikes, which services require you to call support in such an instance? In my experience you can often kick out old devices right from the new device that you are trying to use (annoying, but manageable).

It's so weird that they still haven't learned to provide multiple audio tracks and subtitles. I don't know how sizable the portion of the market is that they lose this way, but people like you and me won't ever consider one of Google's services to rent films. Presumably, US-centric decision-making is to blame.

The first season could have been better, but it has great production values and as far as I can tell they see it as an investment and they are not going to cancel it soon. So for now I'm optimistic about a second season.

At least in Germany I can find a streaming source via justwatch.com for most movies that I want to watch. Often they are included in one of the flatrate offerings, but the vast majority of classics are available to rent from Amazon and Apple for ~4 euros.

Compared to competitors, Apple TV+ currently appear to put much more emphasis on quality over quantity in their original content. I really enjoyed some shows like Severance, For All Mankind, Trying, Acapulco, Foundation. It reminds me of the time when Netflix Originals stood for content to be excited about. What they don't have is a back catalog of licensed films and shows included with the flat rate offering (they all cost an extra fee).

Please help out us outsiders. My guesses are: MPK = Menlo Park? FRE = Fremont?

WhatsApp users who sync their contacts provide Meta with a social graph annotated with phone numbers.

It's utterly amazing to me that a single person, Volker Wertich, designed the game and programmed all of it in raw Assembly (more or less in a year). https://en.wikipedia.org/wiki/The_Settlers_(1993_video_game)...

Very interesting! Do you know about any longer public write ups about the matrix management style and organization of engineering at Nvidia?

Ah, that's true, thanks. It's the same type of NVLink as on the A40 GPU. https://images.nvidia.com/content/Solutions/data-center/a40/...

Ampere NVLink (NV3) was 600 GByte/sec, with Hopper (NV4) it's 900 GByte/sec. https://www.nvidia.com/en-us/data-center/nvlink/

Market segmentation. Back when the Pascal architecture was the latest thing, it didn't make much sense to buy expensive Tesla P100 GPUs for many professional applications when consumer GeForce 1080 Ti cards gave you much more bang for the buck with few drawbacks. From the corporation's perspective it makes so much sense to differentiate the product lines more, now that their customers are deeply entrenched.

You can, for instance, connect two RTX 3090 with an NVLink bridge. That gives you 48 GB in total. The 4090 doesn't support NVLink anymore.

The two cards show as two distinct GPUs to the host, connected via NVLink. Unification / load balancing happens via software.

You can also join a pair of regular PCIe H100 GPUs with an NVLink bridge. So that topology is not so new either.

That's the SXM version, used for instance in servers like the DGX. It's also faster than the PCIe variation.

78x is roughly two orders of magnitude in typical physics parlance. If you take a more CSy stance and count powers of two, it would be six to seven orders of magnitude. Sounds entirely fair to me.

SFINAE

The Jeopardy winning Watson wasn't a large language model (LLM) as it is understood today. That technology didn't exist at the time. It was more of an expert system.

It's surprising really that they get away with showing absurdly low quality information like that next to their ads.

Where did you pull that number from? That's off by a factor of 1000 or so.

https://www.boerse-frankfurt.de/equity/mercedes-benz-group says the market capitalization is 70.29B Euros.

It would be perfect if Rosetta supported AVX instructions. For instance, I could just install pre-built tensorflow binary packages in a x64 container on Lima then.

Link to the new tool: https://www.deepl.com/write

The tooling is there to convert existing CUDA code into a form that runs on AMD GPUs as well: https://docs.amd.com/en-US/bundle/HIPify-Reference-Guide-v5.... They have drop-in replacements for most Nvidia-specific libraries like cuDNN, too.

So, while the documentation still doesn't make it trivial to get started, the software side definitely seems to have improved in recent years.

The car is absolutely not preferable if you value using your time for anything else than staring at the road and concentrating on driving.

Google Maps also suggests 6.5 hours by car compared to 4.5 hours by train between Munich and Berlin.