AMD Still Believes in Moore's Law, Unlike NVIDIA

[TheLostSwede]

Back in September, NVIDIA's Jensen Huang said that Moore's Law is dead, but it seems like AMD disagrees with NVIDIA, at least for now. According to an interview with AMD's CTO, Mark Papermaster, AMD still believes that Moore's Law will be alive for another six to eight years. However, AMD no longer believes that transistor density can be doubled every 18 to 24 months, while remaining in the same cost envelope. "I can see exciting new transistor technology for the next - as far as you can really plot these things out - about six to eight years, and it's very, very clear to me the advances that we're going to make to keep improving the transistor technology, but they're more expensive," Papermaster said.

AMD believes we'll see a change in how chips are being designed and put together, with chiplets being the future of semiconductors. Papermaster calls this "a Moore's Law equivalent, meaning that you continue to really double that capability every 18 to 24 months" although it's not exactly Moore's Law in the traditional sense. AMD also appears to be betting heavily on FPGA technology in some of its market segments, for something the company calls adaptive computing. As to how things will play out, time will tell, but with both AMD and Intel going down the chiplet route, albeit in slightly different ways, we should continue to see new innovations from both companies, with or without Moore's Law.



View at TechPowerUp Main Site | Source

 

[thegnome]

Exactly, it would be naive to think Moore's Law is dead. It has just changed.

 

[Solaris17]

One thing I would be scared of with chiplets is physical size. I mean if you look at some of the chips now and in the past, like thread ripper/ x299 and the server chips now the physical size is getting pretty crazy, though I guess there was never really an expectation. I guess I find it a little weird that eventually I could be cooling something the size of a cell phone on my motherboard.

Just seems weird to me.

I think the future will be a break from silicon or the traditional processes anyway. Like using a metric besides nm. Quantum is too far off. We are more likely to see slotted CPUs or a break through in the current process/elements before that kind of thing is common; even if its the logical evolution.

 

[TheLostSwede]

One thing I would be scared of with chiplets is physical size. I mean if you look at some of the chips now and in the past, like thread ripper/ x299 and the server chips now the physical size is getting pretty crazy, though I guess there was never really an expectation. I guess I find it a little weird that eventually I could be cooling something the size of a cell phone on my motherboard.

Just seems weird to me.

I think the future will be a break from silicon or the traditional processes anyway. Like using a metric besides nm. Quantum is too far off. We are more likely to see slotted CPUs or a break through in the current process/elements before that kind of thing is common; even if its the logical evolution.

Well, there's already this.

Cerebras

Cerebras is the go-to platform for fast and effortless AI training. Learn more at cerebras.ai.

www.cerebras.net

 

[ARF]

Maybe AMD is afraid that its CPU business is dying.
The chiplets is the last resort but definitely only a temporary solution.

Going further there are two solutions:
1. Faster software
2. The cloud but it will require faster internet connection speeds

 

[Ravenlord]

Moore's Law is biggest joke in a whole IT. I never liked this fake "law" which isn't law at all. 100% dependent on how tech companies are doing.

 

[AnotherReader]

Going further there are two solutions:
1. Faster software
2. The cloud but it will require faster internet connection speeds

I won't comment on the rest, but faster software would be a miracle. I suppose you haven't heard of Wirth's law.

 

[TheLostSwede]

Moore's Law is biggest joke in a whole IT. I never liked this fake "law" which isn't law at all. 100% dependent on how tech companies are doing.

Well, it's not a law in any sense of the word, but it's something of a guiding principle when it comes to silicon design and performance improvement over time.
However, it's no longer what it once was.

 

[ARF]

I won't comment on the rest, but faster software would be a miracle. I suppose you haven't heard of Wirth's law.

That's worse than Moore's.
Imagine running a 64-bit Windows 95 on Ryzen 9 5950X. How fast would it be?
Many of the slow-downs are artificially added planned obsolescence shenanigans by M$ which asks that you pay more and sells the new hardware.
It's all a conspiracy.

 

[Auer]

Is Moore’s Law Really Dead?

A postmortem on “Cramming More Components Onto Integrated Circuits”—the most influential article of the 20th century.

www.wired.com

 

[AnotherReader]

That's worse than Moore's.
Imagine running a 64-bit Windows 95 on Ryzen 9 5950X. How fast would it be?
Many of the slow-downs are artificially added planned obsolescence shenanigans by M$ which asks that you pay more and sells the new hardware.
It's all a conspiracy.

It isn't just Microsoft. It's an industry wide trend; the latest example is microservices.

 

[TheinsanegamerN]

Maybe AMD is afraid that its CPU business is dying.
The chiplets is the last resort but definitely only a temporary solution.

Going further there are two solutions:
1. Faster software
2. The cloud but it will require faster internet connection speeds

The “cloud” is thinly veiled planned obsolescence DRM that needed to DIAF yesterday

 

[Assimilator]

"AMD believes Moore's Law will still be relevant for another 6 to 8 years" isn't the same thing as "AMD still believes in Moore's Law" in any way shape or form.

Silicon has come to the end of its useful life as a semiconductor, and we don't yet have something to replace it (at least, not something as cheap or simple).

 

[Vayra86]

One thing I would be scared of with chiplets is physical size. I mean if you look at some of the chips now and in the past, like thread ripper/ x299 and the server chips now the physical size is getting pretty crazy, though I guess there was never really an expectation. I guess I find it a little weird that eventually I could be cooling something the size of a cell phone on my motherboard.

Just seems weird to me.

I think the future will be a break from silicon or the traditional processes anyway. Like using a metric besides nm. Quantum is too far off. We are more likely to see slotted CPUs or a break through in the current process/elements before that kind of thing is common; even if its the logical evolution.

Why scared though

We have been stuck to a socket size for how long now? In the meantime, parts of the whole system have moved into the CPU, and motherboards clearly have the capacity to shrink to lots of sizes for lots of use cases. There's a lot of free space there.

Also, larger chips have more wiggle room in terms of temperature. Temperature-regulated activation or load regulation of core complexes? Sure... AMD is already moving into a fixed peak temp with the current gen, so they know how to extract performance adhering to a strict temp limit, its like a perfection of throttling.

Larger chips do require changes in the economy of the whole fabrication/production process and its cost. The first thing that comes to mind is: longer lifetimes of products. Pay more, use longer, because honestly, the E-waste we create now is retarded if you look at what the chip is capable of.

 

[Daven]

This is so frustrating. Can someone start a charity for the semiconductor victims of Moore’s Law? Its apparently had a profound negative psychological effect on these poor bastards.

 

[Solaris17]

Why scared though

I explained it lol gotta read the whole thing.

As for size I kind of wonder how big they can get away with I almost wonder if they will run into odd latency problems with super big dies if they go that route. Maybe at that point it would just be creative positioning of the chiplets they already do similar in the arc itself so it would really just be doing it in a bigger scale.

Thinking about it more given the already existing chiller design they might even be doing that kind of thought already.

Interesting times ahead for sure.

 

[hurakura]

Go AMD !!!

 

[TheoneandonlyMrK]

I explained it lol gotta read the whole thing.

As for size I kind of wonder how big they can get away with I almost wonder if they will run into odd latency problems with super big dies if they go that route. Maybe at that point it would just be creative positioning of the chiplets they already do similar in the arc itself so it would really just be doing it in a bigger scale.

Thinking about it more given the already existing chiller design they might even be doing that kind of thought already.

Interesting times ahead for sure.

I think they have enough experience on 2.5/3D to Safely say they can do it in reasonable package sizes, but can they make it affordable at scale.
I think wafer cost and obviously max reticle size put limits on the chiplets size individually And the substrate especially if silicon becomes more important and the limiter(of package sizes)

I agree interesting.

 

[R-T-B]

Imagine running a 64-bit Windows 95 on Ryzen 9 5950X. How fast would it be?

It'd be pretty fast at booting. It'd also fail to really use any ram over 128MBs just like the original 9x kernel.

It's not really a conspiracy. It's human nature to be more lazy with more resources. That said modern OSes do way way more than 9x could even dream of (Heck, OS/2 Warp 4 kills 9x there).

The “cloud” is thinly veiled planned obsolescence DRM that needed to DIAF yesterday

It's a buzzword for "putting something on a server somewhere else."

Killing the whole concept would kill the internet as we know it. But the buzzword status it has gained isn't helping anyones understanding of it, that's for sure.

 

[RedBear]

If you basically redefine Moore's Law of course you can keep believing in it. Intel has even stated that Moore's Law is "alive and well", for that matter, despite what happened with Intel 14nm and Intel 10nm/7 in the last several years... We'll see in due time whether it's Huang who's making things up to justify the extravagant prices of his GPUs or it's the competition that is making up claims just to get some visibility in the news.

 

[Darmok N Jalad]

One thing I would be scared of with chiplets is physical size. I mean if you look at some of the chips now and in the past, like thread ripper/ x299 and the server chips now the physical size is getting pretty crazy, though I guess there was never really an expectation. I guess I find it a little weird that eventually I could be cooling something the size of a cell phone on my motherboard.

Just seems weird to me.

I think the future will be a break from silicon or the traditional processes anyway. Like using a metric besides nm. Quantum is too far off. We are more likely to see slotted CPUs or a break through in the current process/elements before that kind of thing is common; even if its the logical evolution.

As transistor density increases and dies decrease in size, spreading them out across a substrate is one way to more effectively cool and divide up the thermal density. Even the 7900xtx has more surface area that it would be if it were one giant GPU. The last remaining solution is what AMD and others are doing. Design the chip to operate at a max temp, cover it in thermal sensors, and let it run at a threshold. It’s dumping the same heat out of less surface area, so engineering changes are made and stands temperature conventions in their ear, but the end result getting dumped into the cooler is about the same.

 

[sam_86314]

"AMD believes Moore's Law will still be relevant for another 6 to 8 years" isn't the same thing as "AMD still believes in Moore's Law" in any way shape or form.

Silicon has come to the end of its useful life as a semiconductor, and we don't yet have something to replace it (at least, not something as cheap or simple).

Makes me wonder what will be used in the future. GaN?

 

[RogueSix]

Um, Papermaster pretty much said exactly the same thing as Huang because Huang as well as Papermaster both said that the cost aspect of Moore's Law is dead. Let us see what Huang really said in full context without simply ripping the isolated "Moore's law is dead" quote...

“Moore’s Law’s dead,” Huang said, referring to the standard that the number of transistors on a chip doubles every two years. “And the ability for Moore’s Law to deliver twice the performance at the same cost, or at the same performance, half the cost, every year and a half, is over. It’s completely over, and so the idea that a chip is going to go down in cost over time, unfortunately, is a story of the past.”

... and Papermaster said according to the TPU news article...

However, AMD no longer believes that transistor density can be doubled every 18 to 24 months, while remaining in the same cost envelope. "I can see exciting new transistor technology for the next - as far as you can really plot these things out - about six to eight years, and it's very, very clear to me the advances that we're going to make to keep improving the transistor technology, but they're more expensive," Papermaster said.

^ They are basically saying the exact same thing.

 

[grammar_phreak]

Jensen claimed Moore's law is dead while the AD102 has nearly 3x as many transistors as the GA102.
I'm thinking of one of those popular memes but I'm drawing a blank.
As far as costs go.... TSMC did raise prices and Nvidia no longer gets the repeat customer discount from TSMC.

 

[Bwaze]

Um, Papermaster pretty much said exactly the same thing as Huang because Huang as well as Papermaster both said that the cost aspect of Moore's Law is dead. Let us see what Huang really said in full context without simply ripping the isolated "Moore's law is dead" quote...



... and Papermaster said according to the TPU news article...



^ They are basically saying the exact same thing.

Well, Huang said what he said.

Everybody focuses on his first statement: “Moore's Law's dead.”

He continued with another statement: “And the ability for Moore's Law to deliver twice the performance at the same cost, or at the same performance, half the cost, every year and a half, is over."

But that second statement doesn't make Moore's law automatically invalid - because the original Moore's law doesn't make any observations on cost:

"Moore's law is the observation that the number of transistors in a dense integrated circuit (IC) doubles about every two years."

Of course you can say that lowering cost is what enables this doubling every two years, but you can't just change what the "law" actually states and then moan that it doesn't hold any more.

And we all know what he actually said with this. That price uplift from RTX 3080 at $699 to RTX 4080 at $1199 is somehow the new law, deal with it.