AMD's 5nm CPUs could be monsters | Hacker News Reader

AMD's 5nm CPUs could be monsters

pcgamer.com

22 pointsstambros18 comments

As someone who hasn’t followed this space too closely, is there a good tl;dr explanation of how intel lost their lead? It doesn’t seem like an “innovator’s dilemma” situation since they seem to be getting beat at their core competency of laptop, desktop and server cpus.

There are a few different factors that all added up, at least from my understanding/point of view:

- AMD tried something new (FX series), that didn't go as well as they had hoped (they tried doing something similar to the P4 arc and thought they could do it better than intel). Which resulted in sub-par processors. So people continued to prefer Intel processors.

- Intel having no real need to continue to advance its processors with no serious threat, continued to do smaller iterations on its line, moving from a tick-tock cycle (big improvements, followed by a yearly refresh for additional small improvements) to more of a tick tick tock cycle (more yearly refreshes with marginal improvements).

- Intel also had issues getting their fabs down to smaller processors, which may have contributed to the change in how their cycles were going.

- Intel has also had more issues lately with speculate execution attacks, which due to their current design in how they implemented OoO-execution, has lead them to disabling some of this functionality. This disabling decreases some of their previous performance advantages. AMD has also been hit with some, but not to the same extent with intel, and usually not to the same level of performance degradation.

basically intel took too long getting to the next die shrink. they also failed to appreciate the advantage of amd's chiplet design (or couldn't pivot fast enough to do something similar). both of these allow amd to make more viable parts out of a single wafer, which means they can profitably sell more powerful CPUs at the same price.

it's worth noting that Intel still has faster individual cores (but only because they can clock higher). amd can sell you many more (slightly slower) cores for the money.

What does it mean "can sell you morenfor the money"?

Would Intel go bankrupt if they lowered prices?

it costs amd less money to make an n-core processor, especially when n > 8. so they can sell it for less while still making a profit. I don't know exactly how much intel could lower prices while still making a profit, but most of their parts aren't very competitive at the current prices.

They blew an incredible amount of effort and money on new processes that didn't pan out. The failure rate was too high for production. AMD doesnt fab their own chips so they can piggyback on the progress unlocked by the fabs that build for cellphones (TSMC in this case).

Also, transistor density is a better measure than 5nm/7nm/10nm. That feature detail in nanometers is more or less pure marketing term and doesn't compare accurately. Intels transistor density is still better than the latest Ryzen but it looks like they will be matched or surpassed by the next gen AMD.

Intel is at 14nm and amd is at 7nm, yet the transistor densities are higher on the Intel side?

Do you have any source to back this claim? I know that node sizes are mostly meaningless at this point, but still this is invalidating 3 generations of node jumps on the TSMC/AMD side vs none on the intel side, surely it can't be that meaningless of a number?

Wikipedia claims MTr/mm^2s of:

  Intel 14: 37.5
  Intel 10: 100.8
  TSMC 16/14/12: 28.88
  TSMC 10: 52.51
  TSMC 7: 96.5 to 113.9, depending on iteration.
  TSMC 5: 173

It really shows you how good Intel's IP is that they can even make competitive CPUs on a generation old process.

No, it's just that they use a different scale for vanity metric number. Intel's chips aren't competitive for the price; they just have customer inertia.

Even on a transistors per mm measure the 14nm++++ they are using is vastly inferior to TSMC’s 7nm.

Its only Intel 10nm that is equivalent and we don’t have desktop CPUs on that process yet (if we will at all).

this isn't an entirely fair representation of intel's 14nm. over the years, it has become extremely well optimized for high clocks. the 14nm parts actually have a slight disadvantage in IPC compared to amd, but an individual core is still faster due to the large difference in clock speed.

this has actually created an odd marketing dilemma for intel. despite having less IPC and worse power efficiency, intel's 14nm parts are actually faster than the new 10nm parts because the new process can't achieve such high clock speeds.

You can't really divorce IPC from the process used, the amount of nested logic in each pipeline stage is a direct function of gate delay and therefore the manufacturing process. On one process I may be able to fit 10 stages while on another I could fit 20. It's quite likely Intel's IPC would be much better on TSMC's 7nm vs their current 14nm.

This is also a problem when moving designs to a new generation FPGA. You may find your current level of pipelining is no longer optimal and you should do more each cycle.

People seem to have this weird idea that IPC is unrelated to how things are manufactured.

Intel's 10nm does have a higher density than TSMC 7nm[1]. However, Intel's desktop chips are still on 14nm and that is closer in terms of density to TSMC's short lived 10nm node.

[1] https://fuse.wikichip.org/news/2408/tsmc-7nm-hd-and-hp-cells...

I always thought AMD was attaching itself to inferior fabs by going fabless, but that was before the rise of smartphones pushing CPU and process development.

Intel missing the boat on smartphone CPUs was a massive mistake, and ARMs are coming for their servers and laptops.

> Intels transistor density is still better than the latest Ryzen

In the lab, yes, but aren't these the chips the Intel is failing to ship at scale, leading to desperate measures like rebadging old QA rejects as F-series CPUs?

>is there a good tl;dr explanation of how intel lost their lead?

Poor Management, Culture and Denial. Nothing technical really. ( They lost their Only Paranoid Survives Culture )

10nm was initially expected to be in 2016, but since 14nm was delayed it was naturally assumed it would be 2017 instead. But it was delayed again and in 2018 they shipped 10nm CPU with incredibly low yield.

All while their management and CEO promising 10nm were fine.

If you are counting, that is 4 years of delay. Meanwhile, TSMC, the leading edge Foundry for non-Intel node / products ( That is everything from Qualcomm, Apple, Nvidia, AMD, Broadcom etc... ) has been keeping their head down and grinding. And so they went from roughly 1 node behind ( ~2 years ) to 1 node advance ( ~2 years ).

One thing to point out is while they might have lost their technical lead, their most recent financial results is still recording breaking.

I have heard the Intel's 10nm node was overly ambitious and this caused their delays.

Reply on news.ycombinator.com