Jensen Huang Tells the Media That Moore's Law is Dead

[hat]

Well, it seems like it's hard to innovate on the chip design side, so all these companies rely more and more on node shrinks, which in turn makes for more expensive chips, as the foundries want to recuperate their investments within a set period of time.

I'm sure it is hard. You know what else is probably hard? Developing the machines that manufacture these chips. It's looking more and more like a one sided effort to me. Remember the "MAKE MOAR CORES" meme from the Bulldozer days? Someone needs to make a similar one that says "MAKE MOAR TRANSISTORS!".

 

[ARF]

There is no such thing "free market". The market is always dictated by the capitalists.
Even if they can't sell, they have two options:
- to hide the inventory;
- to do whatever it takes to convince the clients that these are "wrong" and should pay whatever is asked.

 

[openbox1980]

Everyone keeps saying "AMD has a chance to grab marketshare", but that's not going to happen....because even when AMD offers a better value, everyone still buys nvidia (say what you will about AdoredTV, but a year or so ago he did a lot of research into this phenomenon and used the time from the late 2000s and early 2010s as an example for a multpart video on the subject. AMD offered faster cards for better value, and Nvidia still outsold them by a magnitude)...what they really mean is they want AMD to drive prices down so they can buy an Nvidia card cheaper, but that'll never happen unless people actually buy the AMD cards!

People keep complaining about Nvidia and their prices (which unfortunately is going to force AMD to follow suit to please there shareholders), but as long as those same people keep buying Nvidia products, not only will prices not go down, they'll keep going up.

The issue for me on the Nvidia vs AMD cards, Nvidia cards are always more available.
I'm on a laptop instead of a desktop. Laptops, Nvidia+intel rules the market.
I was looking forward for those new AMD 6000 series APUs on the laptops, but after waiting for months, I went with a intel 12th gen cpu. To this day, 6000s APUs are still very rare.

 

[ARF]

I'm on a laptop instead of a desktop.

I have these stickers

 

[TheLostSwede]

I'm sure it is hard. You know what else is probably hard? Developing the machines that manufacture these chips. It's looking more and more like a one sided effort to me. Remember the "MAKE MOAR CORES" meme from the Bulldozer days? Someone needs to make a similar one that says "MAKE MOAR TRANSISTORS!".

Much like game developers relying on faster and faster hardware every other year, so they don't have to improve their coding skills or optimise the code...

 

[Vayra86]

Moore bullshit from the Leather Jacket, gosh

 

[ymdhis]

It's greed and inflation you jacket wearing dumb fuck, not Moore's Law.

 

[AnotherReader]

AD103 is 300 mm^2, and the similar N5 process is 70% more expensive than the N7 process used for Big Navi. The Navi 21 die is 520 mm^2. This means that TSMC would charge almost the same amount to manufacture AD103 and Navi 21. So the 4080 12 GB, even with good margins, could have been priced similarly to the 2020 MSRP of the 6800 XT.

 

[medi01]

and I bet RDNA3 matches if not beats rtx 4xxx series in raw fps and undercuts costs at same time. we will find out soon. Nvidia. lol.

Announced, early Nov, available 3 week-ish after is what reputable leaker says.

There is no such thing "free market". The market is always dictated by the capitalists.
Even if they can't sell, they have two options:
- to hide the inventory;
- to do whatever it takes to convince the clients that these are "wrong" and should pay whatever is asked.

What you describe is monopoly, not capitalism.

 

[mplayerMuPDF]

Sounds like an excuse to me. It seems like there's a lot of reliance on foundries to make better processes so they can just cram more transistors in, rather than improving the architecture itself. This is why we had about 57 *lake generations. Even with the increased density from better manufacturing processes, they just keep making bigger and bigger chips. Guess they're having issues obtaining the next set of design specs from the aliens or something...

Not really a great observation, to be honest. The Garbage (Sky) Lake derivatives were all on 14 nm(+*n) except for that one weird limited edition laptop series (which I don't know why the hell anyone would want to buy Garbage Lake on 10 nm/Intel 7 (R)). They were doing the opposite of just cramming more transistors in to improve performance during that period. They were tweaking their 14 nm node and the Garbage Lake architecture to (barely) increase performance. In fact, "just cramming more transistors in" would be a far more fitting description for what they did with Sunny Cove (and derivatives) as the Sunny Cove core is massively wider and larger. In fact, it has a disproportionally larger number of transistors compared to the actual IPC increase. (No, I am not some genius. I got all this from the Anandtech Sunny Cove deep dive). And that explains precisely why Rocket Lake runs so hot, has lower clocks and the i9-11900 has less cores than the Comet/Coffee Lake (aka Garbage Lake+*n) i9: they didn't have their Intel 7 (R) to compensate for those massively larger cores, so they could fit less cores on the die and struggled with the heat output of those physically huge cores (due to same transistor size). It also explains perfectly why Intel came up with the whole E core (i.e. cram Atom cores on "regular" chips) idea because Sunny Cove and derivatives are a pretty damn inefficient use of die space (which is bad for multithreaded performance/$), it's just that they desperately need that (singlethreaded) IPC to maintain the singlethreaded performance lead.

 

[Deleted member 177333]

Well, something else that's dead is my willingness to shell out crazy $ for brand new NVidia cards. That ended with the astronomical price increases that started with Turing. Happy to buy 2nd and even 3rd hand market.

 

[Godrilla]

Nvidia is past overdue for an ego correction!

 

[trsttte]

hard to innovate on the chip design side

Was it ever easy? These words from Jensen seem like nvidia is assuming it's own incompetence, like look at the competition: amd is moving to chiplets, intel barely entered the gpu market and is already moving in that direction and did some very cool architecture stuff right out of the gate (still to be seen but the xmx accelerators and systolic arrays seem really cool), even apple is putting out some very impressive designs.

And what is Nvidia doing? Being a crybaby saying moore law is dead because it can't produce a high performance gpu without almost doubling previously established power limits lol

Foundries have become more expensive, he's not wrong, but some other things didn't and it's still on nvidia to design better stuff, not just rely on fabs advancing their nodes

 

[CrAsHnBuRnXp]

That's why AMD is using chiplet design on RDNA3. GET SKILL LEATHER JACKET BOYS WOOOOO welcome to the thunderdome Nvidia, get skill yeeee bois

and I bet RDNA3 matches if not beats rtx 4xxx series in raw fps and undercuts costs at same time. we will find out soon. Nvidia. lol.

And when it doesnt because they never do, what will you say then?

 

[Space Lynx]

And when it doesnt because they never do, what will you say then?

6800 XT already did it. I was very impressed how much it went toe to toe with a 3080 and 3080 ti even in some games. hell, there are even a couple of games it beats a 3090.

and it was much cheaper during the entire life cycle. AMD has already been within striking distance performance of Nvidia, and undercut them in price. its one of the reasons I love them.

even in the Ati AGP days this was true.

 

[THU31]

Moore's Law might be dead, but greed is eternal.

 

[N3M3515]

Nvidia trying to clear excess 3000 series inventory. Regardless, it's down to the consumer if they want to pay these prices. If people don't want to or can't, Nvidia will have poor sales and be forced to adapt.

How is nvidia trying to clear excess 3000 series inventory, if the 3000 series are still expensive. And every amd offer is much better.

RX 6950 XT $1000 ------ RTX 3090 $1070
RX 6900 XT $680 ------ RTX 3080 Ti $870
RX 6800 XT $580 ------ RTX 3080 $704
RX 6800 $520 ------ RTX 3070 Ti $610
RX 6700 XT $370 ------ RTX 3060 Ti $450
RX 6650 XT $280 ------ RTX 3060 $400
RX 6600 $240 ----- RTX 3050 $300

 

[openbox1980]

I have these stickers

View attachment 262635

i have those but from the other team

 

[hat]

Not really a great observation, to be honest. The Garbage (Sky) Lake derivatives were all on 14 nm(+*n) except for that one weird limited edition laptop series (which I don't know why the hell anyone would want to buy Garbage Lake on 10 nm/Intel 7 (R)). They were doing the opposite of just cramming more transistors in to improve performance during that period. They were tweaking their 14 nm node and the Garbage Lake architecture to (barely) increase performance. In fact, "just cramming more transistors in" would be a far more fitting description for what they did with Sunny Cove (and derivatives) as the Sunny Cove core is massively wider and larger. In fact, it has a disproportionally larger number of transistors compared to the actual IPC increase. (No, I am not some genius. I got all this from the Anandtech Sunny Cove deep dive). And that explains precisely why Rocket Lake runs so hot, has lower clocks and the i9-11900 has less cores than the Comet/Coffee Lake (aka Garbage Lake+*n) i9: they didn't have their Intel 7 (R) to compensate for those massively larger cores, so they could fit less cores on the die and struggled with the heat output of those physically huge cores (due to same transistor size). It also explains perfectly why Intel came up with the whole E core (i.e. cram Atom cores on "regular" chips) idea because Sunny Cove and derivatives are a pretty damn inefficient use of die space (which is bad for multithreaded performance/$), it's just that they desperately need that (singlethreaded) IPC to maintain the singlethreaded performance lead.

Ah yes, fair point. How could I have forgotten all the 14nm jokes in every Intel thread? Still, I maintain my point that the manufacturers rely too much on foundry process improvements alone to improve their product.

 

[mplayerMuPDF]

Ah yes, fair point. How could I have forgotten all the 14nm jokes in every Intel thread? Still, I maintain my point that the manufacturers rely too much on foundry process improvements alone to improve their product.

Honestly, I think it is wishful thinking to think that we could get continuously get significant performance improvements just by rearranging the same transistors to get magically superior architectures. Every once in a while it may be possible, sure, but not most of the time. People should also just learn to curb their expectations and optimize (or demand more optimized and minimalistic) software.

 

[hat]

Hey, I'm all for optimized and minimalistic software. Imagine how well things would run without a bunch of unnecessary bloat? That would be so much better, and not just for performance reasons...

 

[enzolt]

The only thing dead here is 40 series sales.

 

[evernessince]

How is nvidia trying to clear excess 3000 series inventory, if the 3000 series are still expensive. And every amd offer is much better.

RX 6950 XT $1000 ------ RTX 3090 $1070
RX 6900 XT $680 ------ RTX 3080 Ti $870
RX 6800 XT $580 ------ RTX 3080 $704
RX 6800 $520 ------ RTX 3070 Ti $610
RX 6700 XT $370 ------ RTX 3060 Ti $450
RX 6650 XT $280 ------ RTX 3060 $400
RX 6600 $240 ----- RTX 3050 $300

Nvidia might assume that people will buy their cards even if they are quite a bit more expensive. Given Nvidia's pricing of the 4000 series, can't say they have their head screwed on right in regards to price.

 

[N/A]

1.5 improvement is still pretty good if you look at their GA100 N7 to H100 N4 big die, bump the clock speed by 20% and you are almost getting there. But yes it;s almost over, 3,2,1 and there is no sub 1 Node.

 

[mechtech]

""A 12-inch wafer is a lot more expensive today than it was yesterday, and it's not a little bit more expensive, it is a ton more expensive,"

ok yes,........ but I don't see no invoice saying how much more.

Also for an over simpllification and simple math, say 8nm process then move to 4nm. 8x8 = 64nm2 and 4x4 = 16nm2. So a factor of 4, so if the same chips were fabbed, (neglicting yields, etc) a 4nm wafer should yield 4x the chips as an 8nm wafer. So if the cost is 4x (400%) more for the 4nm wafer over the 8nm wafer, the end price should be roughly the same per chip. Now of course if the transistors double........