• Zen 6 is here but not for you.  (WCCFTech)
  
  Only the server versions are shipping now, but that tells us pretty much how the desktop versions will perform, because the CPU chiplets used are identical; only the I/O die changes.
  
  And what we see from the CPU chiplets is 27% better performance at the same clock frequency, 10% higher clock frequencies (and desktop chips may go higher than that), and 50% more cores - twice the total performance.
  
  Also interesting is the Zen 6c low-power chiplet.  Unlike Intel's low-power cores it is functionally identical to the full Zen 6, just implemented in a denser, power-optimised layout.  But where Zen 5c had 16 cores with 2MB of cache per core - twice as many cores but half the cache per core of Zen 5 - Zen 6c has 32 cores per chiplet and 4MB of L3 cache per core.  128MB in total, which is more than a desktop 9800X3D chip.
  
  And the 32 cores all being on a single die means there is no added latency between the cores and between the cores and the cache.  Or not much, anyway; L3 cache accesses look to be about 10 cycles slower than Zen 6.
  
  As to when we'll see Zen 6 on the desktop - that's unanswered.  Engineering samples for the laptop versions are already circulating, so it's up to AMD's marketing, so somewhere between tomorrow and never.
  
  
  
• Tom's Hardware found a 93% performance gain comparing a 256-core Zen 6c with a 192 core Zen 5c.  (Tom's Hardware)
  
  So right in the ballpark of 2x.
  
  

Tech News

• AMD also announced MEXT, which appears to be a predictive swap file.  (WCCFTech)
  
  All modern operating systems already have swap files, where unused pages in memory are written out to disk to free them up for other tasks.  What MEXT does is try to read the pages back in before they are actually needed so this doesn't kill performance - and reportedly works pretty well.  It's still slower than having enough DRAM, but it's faster than traditional swap mechanisms.
  
  
  
• Who is Alice Huang and what has she been up to?  (Tom's Hardware)
  
  And is she why you can't buy an RTX 5090?
  
  
  
• ]Terafab, the SpaceX / Tesla chipmaking join venture - is targeting production of a million wafers per month by 2032.  (Tom's Hardware)
  
  That's great, but if those aren't memory wafers it doesn't do us much good.
  
  They are memory wafers?  Or at least, 70% of them?
  
  Well...  Good.
  
  
  
• How does Apple's top-of-the-line M5 Ultra compare with Nvidia's RTX Spark on local AI?  (Tom's Hardware)
  
  Like a bug, you say?
  
  On the one hand, a 256GB M5 Ultra Mac Studio costs twice as much as the 128GB RTX Spark for twice the performance.  On the other hand, there is a 96GB model that is only a little more expensive than the RTX Spark, and if your LLM fits it still delivers twice the performmance.