DavidGraham
Veteran
Elon Musk is builing a new AI project: a gigafactory comprising of a 100K H100 GPUs, he couldn't wait for H200 or B100/B200.
Is gonna be a monumental waste of GPU's, time, effort, money and energy for everybody involved. 'xAI/Grok' is not gonna ever amount to anything that other competing AI's wont do much better. Elon is trying to just brute force this and Grok is gonna be irredeemably bad regardless.
Sorry I meant what more should they be doing on work graphs to meet your bar.
It’s really doubtful investment in tensor cores is to blame for lack of RGBE and pointer support on Nvidia hardware. They aren’t competing technologies and solve different problems.
What’s the source for the Spider-man descriptor issue? Is it being handled differently/better on other architectures?
DegustatoR
Legend
Nvidia h/w support RGBE just fine.
Including the new optional support for render targets and UAVs that Lurkmass pointed to?
A probable issue for work graphs /w mesh nodes on their implementation is that there's going to be potentially a lot of transitions between compute/mesh nodes which will cause WFI/subchannel switches (implict barriers) ...
If you look at his profiling data you'll notice that one hotspots involves the CopyDescriptorsSimple API so now that we games (Monster Hunter Rise exhibits similar issues) that are going against their one of API usage recommendations I think it's time that we stop kneecapping other architectures that are capable of loading descriptors from memory directly. D3D12's descriptor 'heap' abstraction model has not aged well in the bindless era ...
https://wccftech.com/nvidia-mediate...computex-2024-setting-a-new-era-of-computing/
Computex 2024 will be a huge deal for the AI PC segment, as MediaTek & NVIDIA are expected to present their AI PC SOC solution at the event.
Computex 2024 will be a huge deal for the AI PC segment, as MediaTek & NVIDIA are expected to present their AI PC SOC solution at the event.
MediaTek uses off the shelf ARM designs so any secret sauce would likely come from Nvidia. Will be interesting to see how well the arch scales down. The Windows on ARM AI PC market is non-existent but they must know something we don’t. If Intel can deliver with Lunar Lake this ARM renaissance will be over before it begins.
It's a shame about that SVE/2 implementation will still only be 128-bits. Just means that no ARM vendors are actually serious behind the idea of being an alternative to modern AAA PC gaming ...
Last edited:
nVidia can't stop growing and their numbers are just up there with those of Apple and Microsoft.
www.forbes.com
Apple might be beaten soon. I take morbid pleasure on that. Go nVidia!
How Big Can Nvidia Get As It Threatens Apple And A $3 Trillion Market Cap?
If you believe Nvidia’s CEO that we’re amidst a “new industrial revolution,” then an investment in Nvidia is akin to an investment in the steam engine in the 1700s.
www.forbes.com
Apple might be beaten soon. I take morbid pleasure on that. Go nVidia!
one thing is interesting, although I don´t know if it´s true, but is suppose to be manufactured at Intel fab using Intel 3 process
Mod note: paste twitter url into the media button. Change X.com into Twitter.com
And it will create the correct media link for it to show.
Last edited by a moderator:
cheapchips
Veteran
From the embedded Bloomberg interview it sounds like they are targeting next year for their 'PC AI' chip. That's when Qualcomm's exclusivity deal expires.
DavidGraham
Veteran
About CUDA, and TPUs ..
Semiconductor legends take a stroll in a Taiwanese night market — Nvidia, TSMC, MediaTek, and Quanta heads seen eating dinner
www.tomshardware.com
Semiconductor legends take a stroll in a Taiwanese night market — Nvidia, TSMC, MediaTek, and Quanta heads seen eating dinner
A sign of things to come?
it looks like chinese cooking festival meeting
Blackwell Ultra and X800Ultra is next "year".
Rubin, Vera, and X1600 is the "year" after, 2026(?) which had the 3.6TB/s NVLink 6.
Yeah, missed the year.
Still 12x within 10 years should still be one of the fastes speed up in networking.
