@SebAaltonen I don't think this is correct, at least for the 5600X. It averages at 4.6GHz full load, so compared to PS5's 3.5GHz it has almost equal "cores*frequency". Then it has 4x the L3 coherent among all cores and +19% IPC.
The 5600X is comfortably above the PS5's CPU.
@SebAaltonen Oh no, I hope you get better.
Just be sure you get some real rest doing the least possible physical activity. Like sit in a comfortable couch all day and try not to move anything but your fingers. That'll do the trick.
@opinali@Kepler_L2 IIRC, there's a number of games that get a 1440p and 1800p output post-temporal upscaling in the PS5 and SeriesX, and then use a cheap spatial upscaler.
@SebAaltonen I bet the following weeks will have a significant proportion of Fable and Opus token generation dedicated to hacking the latest PS5 and PS5 Pro firmwares and porting the game to PC.
Many hackers have been preparing this for months, so it'll be a matter of days or weeks.
@SebAaltonen X3D probably comes later as usual.
3D cache is an incremental manufacturing, packaging and logistics effort.
OTOH with Zen6 you can buy super fast CUDIMM modules in the future and you can also populate all 4 slots with no clock throttling.
@SebAaltonen Not if he's declared guilty of being complicit to a massive GPU smuggling operation to China, putting into question a lot of past revenue and forecasts.
IMO there's way too much attention on the mistresses and not enough on what's actually really dangerous to him and to Nvidia.
@keaukraine@SebAaltonen@LuftkoppTim Graphics don't have to be limited if the LLM decoding is properly throttled.
RTX5080 decodes Gemma4 12B FP4 at ~100tk/s so Helix AT2 should do >120tk/s.
Limit decode to 8tk/s and that's 7% of GPU/RAM bandwidth on Helix and 10% on PS6. They just need capacity to fit the models..
@SebAaltonen@LuftkoppTim BTW I don't think AAAs will target 4K60, ever. We're prob looking at quarter base resolution + ML upscaler as standard.
Also: 1) if RT and PT are so fast on RDNA5, no more shadowmaps; 2) if ML inference is so fast, neural textures become norm. RAM savings could be huge.
@SebAaltonen@LuftkoppTim RDNA5 should support MXFP4, meaning 4-6GB would be enough for a custom 8-12B model.
Limit decode to ~8tk/s and you have a very capable reasoning and dialogue engine sipping little power, processing and bandwidth resources.
@SebAaltonen@Spybef0rey0ubuy Really? Buy while a GPU trafficking scandal with multiple possible arrests is still unfolding around the company?
Sure, PE ratio is down, but that what we don't know what will happen to revenues from trafficking.
@SebAaltonen To be fair, the price Sony set for the EU was indeed preposterous considering the relatively low increased manufacturing cost it had over the base PS5.
Even if the APU was twice as expensive and it had 2x larger storage, it came out in a time when 1TB of PCIe gen4 was ~$60.
@TripperDay@SebAaltonen@liquidtrading Yes, I can see how that could become a problem in this economy.
Good thing Jensen learned how to flip burgers, in case his retirement fund dries up.