Jerry 🦙💝🦙Mar 17

I’m catching up on the news today and see that Nvidia forecasts $1T in GPU sales over the next 18 months. That is crazy. I don’t know how that works at any level.

Also on the nvidia news front, they are making a version of the gpu for orbital datacenters. Now I’ve been around for a long, long time and I’ve heard some pretty dumb ideas in that time, and I feel like orbital datacenters is at least in the top 10.

Show threadRootWyrm 🇺🇦Mar 17

@jerry to be clear and unquivocal about this:

None of that works. At all. In any form. It's all straight up fabrication. They can't even design silicon that works right on Earth.

And $1T of "GPU sales" would be more than the GDP of Taiwan. It's more than the GDP of Switzerland.

Show threadda_667Mar 17
@rootwyrm @jerry Fuck sake, heat dissipation in vacuum would make things hard enough. Yes, it can be done, but doing so at the scale of an entire datacenter in orbit would be fucking insane.
Show threadA wizard did it  🇨🇦Mar 17
@da_667 @rootwyrm @jerry Nevermind heat dissipation, hardening devices to prevent bit flipping from cosmic rays becomes exponentially more difficult the denser the chip is. I read an article that the most advanced CPU that's used on the ISS is the equivalent of an Android phone from the mid 2010's
Show threadRootWyrm 🇺🇦Mar 17
@Mustardfacial @da_667 @jerry there are firms that claim to be flying Ryzen embeddeds which are "radiation tolerant," which is of course, bullshit.
AFAIK the fastest actually radhard CPU remains the BAE RAD750 which is a PowerPC 750 that can run at a blazing 118MHz.
Show threadA wizard did it  🇨🇦Mar 17

@rootwyrm @da_667 @jerry "Radiation tolerant" just means they slapped some shielding over the chip and called it a day.

I found the article I was initially thinking about, and of course it was from 15 years ago
https://spacenews.com/samsungs-nexus-s-aboard-nasas-final-shuttle/

Samsung’s Nexus S aboard NASA’s Final Shuttle

Samsung's Nexus S aboard NASA's Final Shuttle

SpaceNews
Show threadRootWyrm 🇺🇦Mar 17

@Mustardfacial @da_667 @jerry yeah, those are 'rad tolerant.' I can't find any actual radhard silicon faster than the RAD750 (which has admittedly been improved.)

For those unfamiliar, "radiation tolerant" also means that your CODE must accept and compensate for a VERY high error rate and the potential of entire sections of silicon to just fail. So of 8 cores, a maximum 3 are usable. (3 to verify operation, plus 2 spare, which doesn't even work because it's a monolithic CCD.)

Show threadA wizard did it  🇨🇦Mar 17

@rootwyrm @da_667 @jerry Looking around the only radhard microprocessor I can find that beats the RAD750 is the Gaisler GR765. It runs at 800MHz, but even that is a SPARC/RISC-V processor and is still in development.

Most of the actual radhard silicon I see is in the 150nm range, which is a laughable size compared to nVidia's current GPU's built at what...5nm?

Show threadRootWyrm 🇺🇦

@Mustardfacial @da_667 @jerry yeah, the GR765 hasn't flown, the RAD750 has. Also, the GR765 is using a LEON SPARCv8 IP core. People truly do not understand this part.

SPARCv8 was first introduced in 1989. It's that old. But wait, it gets worse. The LEON3FP is a V8E (cut down) 32-bit version. This is not sun4u. Not sun4d. Not even sun4m. It's sun4e. All the clock in the world will not make up for all of that.

Mar 17 at 4:49pmWeb
Show threadMeatzieMar 17
@rootwyrm @Mustardfacial @da_667 @jerry this is fascinating, anything else good to look up if I want to learn more about radhard stuff? Who is the standards body for this?