How good is their game-streaming at cutting out the various 'filler' time(loading screens, startup animations telling you that Nvidia is the way it's meant to be played) and does it have some sort of 'suspend' state to handle cases where you would pause a game if it either doesn't have save on demand anywhere or takes a little while to load from save?

I know I have a bunch of games that I don't actually play all that much, in wall time terms; but Steam's 'executable run time' number is 2x-10x as high because I'm not necessarily going to save, quit, open, load if a game can be paused and alt-tabbed without crashing or continuing to eat all the GPU time to render a static image I'm not looking at; and I'm not necessarily going to do anything other than pause if I'm called away for 15 minutes by something else.

Unless the streaming service is impeccable at cutting out the various little waits it seems likely that people are going to start resenting the time limit even more than their level of seriousness and actual hand-on-controller-and-attention-locked time suggests.

Given that they already purchased Activision Blizzard King, including the "King" part of it, they clearly already have a considerable interest in casual players.

What is less obvious is how that is supposed to work for their streaming service. The stuff aimed at casuals, sensibly, is normally targeted at either potato-tier PC specs or the (pretty punchy in absolute terms) mobile devices that people worth targeting have.

It is much less clear whether there's a 'casual' market that demands current-gen console performance in very limited amounts rather than mobile or iGPU level titles. Given the extreme prevalence of various 'free to play' models that do sneaky monetization later it also looks like the 'casual' market's (significant) amount of money is largely locked up behind the sort of price sensitivity that requires clever psychological tricks to break through. I'm not sure that they'll be a good fit for something as obvious and psychologically burdensome as a ticking countdown timer hanging over every minute they play.

I have no particular reason to assume that Nvidia is lying about what this software does; but unless they've stumbled on a veritable miracle it's going to mostly be a mechanism for running largely separate models of whatever size your hardware allows; not anything close to what the datacenter comparison implies.

There's a reason why Nvidia sells most of their higher end models intended for parallel use with NVLink between the chips inside the chassis and a lot of either infiniband or RDMA-capable ethernet between multiple chassis; as well as shoving a considerable amount of RAM per GPU onto those parts. And it's because latency and bandwidth matter a lot.

As it is people seek out the unified memory parts, Apple's or strix halo, or nvidia's little GB10, because the GPU having to go to main memory over a mere x16 PCIe link is already a serious problem. If you think either 256 or 512 Gb/s directly to the PCIe root on the same die as the main memory controller doesn't cut it I have some...bad news...about how well NICs are probably going to work for you, especially with the Nvidia parts that don't support RDMA even if you have NICs that do.

As long as I know how fast I'm going and how much fuel I have left, what do I care about engine speed and oil pressure and coolant temperature, and so on? I don't need to monitor all of the minutiae of my engine's operating point while I am driving. That is handled by the FADEC. The FADEC will tell me when and if there is a problem. I do not need to monitor these things myself. This emotional attachment to the past is just that, an emotional attachment.

There are lots of businesses that are inexplicably still around.

Steak n Shake
Friendlys
Giordano's
Hooters
Many more

Physical ones are definitely rare beasts. Hypervisors make having support for it a lot more generally useful, though. You can still run into weird edge case limitations(certain PCIe devices, under some circumstances, really don't like being detached from the host and popped onto a VM); but the hypervisor will definitely be able to tell guests stories of dynamic reconfiguration that the hardware isn't electrically or mechanically capable of supporting; and being able to add or remove CPUs or RAM without rebooting is nice.

Why are you doing AI work with GPUs? For the cost of a pair of GPUs you can have a DGX Spark.

Tell me you've never used an AI coding tool, without telling me ...

I don't know about the rest of the peanut gallery, but I can say first hand that claude is excellent at commenting and documenting code and architecture, as well as keeping track of lessons learned.

It could certainly just be a numbers game; tons of ICE vehicles out there; or whether there's an issue with accumulated cultural assumptions and engineering priorities.

I can at least imagine that there's a difference in mindset between "this is a 12v lead acid starter battery; it'll put out a solid few tens of amps for the starter motor assuming it is working correctly; but its energy density is kind of trash and we mostly prefer not to dump additional systems onto it, especially when that exposes them to the slightly gross output of the alternator" and "This battery has been designed to fully provide for all energy requirements of a vehicle; if you displease it it will probably kill you and then burn vigorously for a day or two. A foundational assumption of all systems that need energy is that they'll be connected back to it in one way or another because there aren't other energy sources here."

Not like electric vehicle designers couldn't/don't make mistakes, and certainly some of their mistakes burn with the really exciting commitment of a large light metal fire; but the engineering culture around an ICE vehicle seems much more likely to produce a string of individually reasonable, collectively dangerous, "c'mon, it's just one little 12v system" requests that get bolted onto a fairly antique battery without too much broader thought; while analogous feature demands in a fully electric vehicle will either be judged high energy enough to potentially get brought into the high voltage wiring(which, barring gross negligence across the design is treated like it can and will kill you) or totaled in to some low voltage system budget that will be handled by a DC-DC converter designed with actual care because going from 600v to 12v isn't something you do as casually as just slapping another 12v requirement in parallel with the other stuff on the 12v battery and calling it good.

Except you are accepting their stated reasoning as gospel. They don't want to free up RAM because of a RAM shortage. They want to free it up to run local AI models on the phone, mostly for data harvesting, behavior recognition, and surveillance.

... Into a trash can that's owned by law enforcement.

IANAL, but don't lie and say you forgot your PIN/password. It is a crime to lie to a cop. Instead, just refuse to give it to them because you don't have to.

This sounds like it was written by someone who actually knows what they are doing, so I assume that they know better than I; but I am surprised to hear that it is impersonating a system dll; since (in all comparatively recent windows versions) system components are signed. Absolutely tons of fuckery you can do to them and around them; but if dpapi.dll doesn't have a signature that checks out the mechanisms that try to keep OSes up and running on marginal hardware subject to random power loss will likely eventually object and attempt a repair, even if there isn't any more paranoid EDR tooling in place.

Is the on-disk copy of dpapi.dll untouched and only the in-memory instance tampered with?

I just can't believe that It was the guys with their finger on GPU allocations, not just the expendable PC OEM they were trying to throw under the bus, the whole time. Whoever could have guessed?