You’re jumping to a lot of conclusions there based on no hard data. Nvidia didn’t pay out million dollar salaries. The shares that employees already own appreciated significantly. The whole lazy rich employee thing is probably real but we have no idea if there will be any tangible impact to their roadmap. Lots of companies have made employees rich without crashing and burning after.
Speculation and Rumors: Nvidia Blackwell ...
- Thread starter pharma
- Start date
You’re jumping to a lot of conclusions there based on no hard data. Nvidia didn’t pay out million dollar salaries. The shares that employees already own appreciated significantly. The whole lazy rich employee thing is probably real but we have no idea if there will be any tangible impact to their roadmap. Lots of companies have made employees rich without crashing and burning after.
1. Don't believe the published R&D cost. It's CEO math marketing BS. He put stuff inside that doesn't belong there
2. I can say from first hand experience that Jensen is the most demanding CEO I have ever met. And from the feedback I have these days, he is pushing the company like never before. He is accelerating all product developments and gives higher targets to every department. In consequence, Nvidia is firing on all cylinders that most employees do overtime. To the point that Rubin DC will tape out in August when Blackwell won't be shipping yet. Honestly, Jensen is very much in an Andrew Grove "Only the Paranoid Survive" moment because he knows that everybody and his dog wants a pie of his AI lunch
And that's what AMD is missing.
CoWoS capacity is set to increase significantly this year though. As per reports from Trendforce, TSMC is set to expand capacity by 150% by Q4 -https://www.trendforce.com/news/2024/03/13/news-tsmc-reportedly-sensing-increased-orders-again-cowos-production-capacity-surges/
Probably planned as per customer requirements so might well be possible.
By 50% unless anything's changed.
And that capacity could easily be eaten up entirely by B100/200 if Nvidia could have their way, given the stupid levels of demand for AI horsepower right now.
People often mean the same thing when they say this. '150% of x' is the same as saying '50% higher than x'. You're right technically, just trying to avoid confusion.
Expand by 150% is incorrect though. It’s either expand to 150% or expand by 50%.
Yes I know, I'm just saying people make this mistake even though they mean the same thing. Hence why I said Degustator is technically correct. I'm trying to avoid further confusion or arguing over this, though that's obviously not working. lol
If we're getting technical, as we must in a technical forum lest anarchy take hold, CoWoS capacity is increasing quicker than shit through a goose and, as a result, Nvidia (and TSMC) are going to be richer than Croesus before the bubble bursts, to put it plainly.
Basis the figures quoted in the article it is actually 150% (Though you have to reference an earlier article as well - https://www.trendforce.com/news/202...kor-ase-also-enter-advanced-packaging-for-ai/ ). It says capacity was 14-15k wafers per month in Dec 2023 and was to increase to ~33k-35k per month in Q4'24 which the newer article says is revised to ~40,000 wafers per month by Q4'24. Which is ~150%.
CoWoS is the biggest capacity limitation at the moment so it would certainly help Nvidia and other companies as well. Though 150% is large enough that it should satisfy AI demand and still have some leftover for consumer products as well.
Maybe for next year, likely this year the capacity will only gradually become available as the year progresses
Hot Chips 2024, Monday 26th August 2024
NVIDIA Blackwell GPU: Advancing Generative AI and Accelerated Computing - Ajay Tirumala and Raymond Wong, NVIDIA
NVIDIA Blackwell GPU: Advancing Generative AI and Accelerated Computing - Ajay Tirumala and Raymond Wong, NVIDIA
NVIDIA Might Shift To Improved FOPLP Packaging For A.I. Chips In 2025 - Report
NVIDIA is reportedly shifting GPU packaging to FOPLP ahead of schedule as tight capacity creates risks for GB200 Blackwell supply.
wccftech.com
輝達封裝進化台鏈沾光 面板級扇出型封裝商機一觸即發 | 科技產業 | 產業 | 經濟日報
為緩解CoWoS先進封裝產能吃緊問題，供應鏈透露，輝達（NVIDIA）正規劃將其「地表最強AI晶片」GB200提早導入面...
The only 512 bit card I can even remember is the Radeon HD 2900XT, which had 8 on each side of the PCB. I don't think there's been anything beyond 384 bit since. Given the need for active cooling for the modules these days, would likely be on the same side.
Ah you just reminded me that Hawaii(290/390 series) were 512-bit and I just checked and they did indeed have sixteen 512MB memory chips on one side for the 8GB versions.
I'm a bit skeptical it's 512-bit, but if it is, then with GDDR7 and the maximum number of DRAM chips, they could go up to 96GiB on professional cards which is more than the original H100... that'd be quite interesting! As long as they don't include NVLink, there's still strong differentiation for B100/GB200, and they'd probably limit the consumer/GeForce version to 32GiB as well (half as many chips, and 16Gbit rather than 24Gbit).
GT200 was 512-bit as well, but the photos of those I can find from back in the day were also 8 chips on each side of the PCB:
Zotac GeForce GTX 280 Amp! Edition Review
Today NVIDIA launches their new GTX 280 graphics cards. Zotac as one of NVIDIA's premier partners has engineered a factory overclocked version of the GTX 280 which offers additional performance. Overall the Zotac GTX 280 Amp Edition is the fastest graphics card on the planet today, but also the...