Daily Tech News 24 July 2024

• SpaceX is cheaper and more capable than any other rocket company, and it's not even close. Here's why that's a bad thing. (Ars Technica)
  
  NASA estimated that de-orbiting the ISS in 2030 would cost $1.7 billion.
  
  SpaceX gave them a fixed-price quote of $680 million.
  
  The closest competitor, Northrop Grumman, came in around NASA's estimate.
  
  Sure, it would be great if we had a couple of other companies capable of competing with SpaceX. But the key here is scale, and SpaceX is creating its own scale with Starlink. It's not at all clear how another company is going to compete.
  
  
  
• Local Hyte distributor expects the Calliope Mori and Amelia Watson Hyte / Hololive limited edition cases in stock in the next few days. I've been chasing the Calli case for an entire year at this point, and was looking at having to spend hundreds of dollars to ship one by air from the US.
  
  They still won't be cheap, but they'll be a lot cheaper this way.
  
  There's a Dokibird model coming out in November but I think I'll have enough cases after these two. If it was Sana or perhaps Maid Mint I'd consider it, but while I like Doki I'm not sure I $300 like Doki.
  
  Or Pippa, but I don't need ants.

• Visual effects studio ModelFarm says fuck this we're going AMD. (Tom's Hardware)
  
  They say 50% of their 13900K and 14900K CPUs have failed, and their new systems will all be based on the Ryzen 9950X.
  
  
  
• The Intel problem - as finally confirmed by Intel, is twofold:
  
  First, the CPUs ask the motherboard for voltage levels high enough to fry their circuits.
  Second, the chips rust from the inside.
  
  
  
  Not a great combination.
  
  Intel is pushing out a microcode update next month to fix the first problem, but if your chip is already affected, this comes much too late.
  
  Intel has also been rejecting warranty returns despite knowing of these problems internally for some time.
  
  My most recent Intel system is a 12th gen laptop, so I escaped this one.
  
  
  
• AMD's new 9900X is slower than the 7800X3D for gaming. (Tom's Hardware)
  
  Not sure exactly how relevant this is, because AMD's 12 core chips are not ideal for gaming. Current generation consoles have 8 cores on a single chip, and AMD's 12 core CPUs have 6 cores on each of two chips, so they have cross-chip latency for games that need 8 cores.
  
  If you're focused on gaming, get the cheaper 9700X, or the 9800X3D when it arrives, or if you run heavy productivity workloads as well as games, go all out and get the 9950X.
  
  
  
• Facebook's new Llama 3.1 405b LLM is billed as the world's largest open-source AI model. (The Register)
  
  As a 16-bit model it requires 810GB of video RAM to run.
  
  There's also an 8-bit version that brings that down to 405GB.
  
  Which used to be a lot, and still is.
  
  
  
• 1 bit LLMs can be nearly as good though. (IEEE)
  
  These are typically 1 trit models though - they are trinary, so each element can have a negative, positive, or zero weighting.
  
  This would reduce Llama 3.1 405b down to around 80GB of video RAM.
  
  Which, yes, is still a lot.
  
  
  
• GitHub is starting to feel like legacy software. (Misty's Internet)
  
  Not entirely accurate. Legacy software often works very well, because nobody dares touch it in case it blows up.
  
  GitHub feels like legacy software that someone is pasting an ill-considered flashing interface over.
  
  Because it is.
  
  
  
• The Minisforum V3 tablet - the 32GB / 1TB model - is currently $949 on Amazon. (Liliputing)
  
  That's a pretty good price, though I don't know if I'd spend that much on a single device from a company in that tier.
  
  Though my Beelink PCs were about $250 each and they work perfectly, so maybe I'm overly cautious.

posted by Pixy Misa at 04:27 AM

| Access Comments