I remember Nvidia being "afraid" of them all, yep.
GTC 2024
- Thread starter pharma
- Start date
DavidGraham
Veteran
Meanwhile A100 and H100 made NVIDIA 5 times as big as AMD and Intel combined, so comparatively speaking, Frontier is a drop in the ocean.
Again, tame reception compared to H100. Most are buying MI300 because they can't buy H100 (due to price or unavailability, but mostly the latter).
Bondrewd
Veteran
I love bubbles too.
y2k Cisco vibes. You know how that ended.
All hail Cisco.
Zoomers getting their own dotcom bubble is funny, but that's even less sustainable than it was back then.
not really?
What does even qualify for a "tame reception"?
You really need to stop with this meme, H100 supply leveled off, especially for hyperscale.
Channel checks say as much (4mo lead time worst case is very managable. a lot better than the 2017/2018 Great Xeon Crunch).
MS is buying gobs of 300X because it's good at running Cop(e)ilot.
Very, very good. That's it!
Not sure why Meta is vacuuming piles of 300X for. But Zucc is weird like that, and they love AMD now (more on that in 2026).
DavidGraham
Veteran
What piles? the man announced he is buying 350K of H100 GPUs in 2024 alone! And they are even buying Blackwell too on top of that. How many MI300 they are getting again?
Meme numbers compared to H100/H200/B200. As you would say it.
You can't seriously believe that, this has been debunked several times already, the dotcom bubble had no real demand underneath it, in contrast to right now. If it's a bubble then MI300 and MI400 would be it's first victims. AMD seemed quite hesitant to get on board at the beginning (like they usually do), but now they are committed and are praising AI as much as the next guy.
Bondrewd
Veteran
Chungus ones.
Tens of thousands of units!
I've already referred to a specific UBS note elsewhere, find it and you'll have your definitive unit count.
A lot.
No idea why. MS at least has a definitive usage model.
Kinda?
The bull thesis is AMD capping 15% GPGPU TAM with MI300.
That's a lot!
The bear one is 8-9%, so let's settle at 12% in the middle.
Yes dog walking chatbots and "rent inference farm" startups are real demand.
Truly.
Bondrewd
Veteran
Cisco had a healthy PE back then.
The demand just cratered off just like it always does with any bubble.
DavidGraham
Veteran
We shall see when the year ends. I remind you of your enthusiasm about MI250X, and how it turned it out in the end.
Not just that, Medical, hyperscalers, governments, film makers, robotics, car makers (Tesla), social platforms (TikTok, Twitter), ... etc.
Oh, and concerning the price, it's not 30K to 40K, NVIDIA recalled that number, they are not announcing anything yet, although that maybe the price of the lowest tier (B100), and not B200 or B200 water cooled version.
Bondrewd
Veteran
Not much to see here, those aren't my numbers.
They won a bunch of really-really key systems?
lol.
Really not the own you think it is.
goddamit it's still somehow relevant, won that tiny-ish italian machine recently and has no business doing so.
Kinda?
It's just an extension of known applications and not something the Street hypebeasts over.
Yeah toy chatbots so far.
Office copilot seems like something with a revenue model?
There's some potential with automation of low-level non-critical support jobs (rip India GDP) but that's it.
Only Saudis.
But Saudis buy anything.
Not really?
This has to be a joke.
They were buying those for recommendation shit already.
They're the same part, just different clock and power target.
Hyperscale pays like 25-27k.
Value! (unironically).
oh yeah lol, it's the first NV DC part since V100 to have all the HBMs enabled ootb on launch.
Competition!
Last edited:
DavidGraham
Veteran
UAE, China, India, Japan, UK .. etc.
NVIDIA Powers Japan’s ABCI-Q Supercomputer for Quantum Research
NVIDIA today announced that Japan’s new ABCI-Q supercomputer — designed to advance the nation’s quantum computing initiative — will be powered by NVIDIA platforms for accelerated and quantum computing.
nvidianews.nvidia.com
UK finds £225M for Nvidia-powered Isambard-AI supercomputer
5,448 GraceHopper superchips and 200PFLOPS gets you somewhere in the global public top ten
www.theregister.com
AMD's guidance for 2024 is 3.5 billion for data center GPUs, that's 1/32 of NVIDIA, so let's wait till the end of the year before you open up any premature victory champaign.
Like 3 key systems? compared to how many from NVIDIA? Did those 3 systems get them anywhere marketshare wise? Heck, Intel already won a system and it got them nowhere, you know none of these token government funded cheap systems are anything serious market wise.
Riiiight, these people are spending billions joking around and building toys to dump them in the ocean afterwards.
Bondrewd
Veteran
Those are standard HPC deals.
Those are not my numbers.
Again, consult the UBS note I mentioned.
more and those are Very Important Systems.
Like Titan was for NV.
YES. YESSSSSSSSSSSSSs. yes. absolutely.
Zucc managed to drown tens of billions of dollars into VR shit not even 3 years ago.
Valley big tech and VCs have a lot more money than they have common sense. Thought we had that figured out eons ago?
Cisco's PE ratio @ 2000 was nearly 200. NVIDIA is no where near as high.
People love to compare current stock rally with dotcom bubble, but the PE ratios of the hottest companies today are actually more reasonable. MSFT is < 40, AAPL < 30, GOOG ~25, AMZN ~ 60. Even "AI meme stock" such as SMCI is only ~70.
Bondrewd
Veteran
That's because the overall market and valuations in general were reasonable.
Their valuation is cuckoo bananas (like the rest of ML stocks but I digress).
Either way it's the fundamental problem of VCbros and big tech burning piles of money like coal seeking a problem for a new solution.
Money that's finite, at least for VCbros.
Obviously no one can say whether this is a bubble or not, but no matter how you measure it, today's stock is no more bubble than in 2000.
For example, people tend to think a PE ratio of 100 "does not seem too bad" but in the past that's crazily expensive, because interest rate was much higher back then. For example, in 2000 the Feds funding rates were ~6%. Today we are near that rate but still not as high and people are already asking the Fed to reduce it.
So we have lower interest rates and yet lower PE ratio than in 2000, so I'd say we are much less "bubbly" now than in 2000.
GN's recent takes on all new GPUs has been borderline cringe worthy. They seemingly say things which their viewers like to hear now instead of providing actually correct and relevant technical data.
Also - Intel ads. Lots of Intel ads. Remember how "GPU busy" was the best thing ever since fps?
Also - Intel ads. Lots of Intel ads. Remember how "GPU busy" was the best thing ever since fps?
Was referring to actual mainstream media. CNBC etc. Steve covers it in the video.
Similar threads
