How AMD won

avatar
(Edited)

It's been two years since I last wrote about AMD and Intel on Steem. Coming into that year, 2017, AMD was nowhere, Intel was absolutely dominant. Indeed, they held a 99+% market share in the lucrative enterprise market. Despite all of that, in that post I wrote, "Without any doubt, AMD is going to command a significant lead throughout much of 2019". That's not some random fortune-telling, but based on what was strong evidence at the time. Let's see how it turned out.

Though that post and previous ones about AMD were absolutely not financial advice, some people commented or asked me in DM whether it would be a good investment. So how has AMD's stock done since then? It's gone from $10 to $46. Meanwhile, STEEM has gone from... oh, never mind.

Zen 2 is an engineering marvel

We all knew Zen 2 was going to be good, but what surprised even the more optimistic of us that they have revolutionized how CPUs are made.

Traditionally, CPUs are made using only one piece of silicon. That's how Intel's best CPUs are, even today. This means, to offer greater performance, you need larger and larger dies. As die size increases, cost of production increases exponentially. AMD's Zen 1 did something clever to mitigate this - by combining 4 CPU dies to make one package. This lead to AMD's Zen 1 32 core CPU costing far less to produce than Intel's behemoth 28 core CPU.

But Zen 2 takes things to a whole new level. It further deconstructs CPUs into smaller modules. In a CPU, you have the compute cores, cache, and then there's everything else to keep the CPU cores fed and running. Zen 2 breaks apart the "everything else" into its own dedicated IO (input/output) die, while building the CPU cores out of tiny "chiplets".


Source

The image above shows Zen 1 on the left - already a significant innovation over the status quo of having only one die. But Zen 2 on the right features 8 tiny CPU-only compute dies, with one central IO die for the common functions. Remember once again that for silicon, costs increases exponentially with an increase in die size. Similarly, costs are fractional the smaller your die size is. As you can see, AMD has been able to shrink each of the 8 CPU chiplets into a tiny die size, each smaller than a mobile SoC. As a result, by using 8 of these tiny dies, their costs are a tiny fraction of what it would be had they made a die that's 8 times larger.

The innovations don't just end there. The central IO die at first seems massive and expensive, but it's actually using an old 14nm process that makes it really cheap to manufacture, while the power-sensitive compute dies are on the cutting-edge 7nm process.

The end result is truly remarkable - AMD has been able to pack a gargantuan 64 cores into a single CPU. Remember, Intel is stuck on 28 cores!

Meanwhile, on the desktop, AMD has pulled the same trick, offering 2 of the very same chiplets - further economies of scale - taking mainstream desktop up to 16 cores. Once again, Intel's best desktop CPU offers 8 cores and draws considerably more power. Indeed, AMD's 16 core 3950X is so good, it has obsoleted Intel's entire workstation line-up. Just a few months ago, they used to sell the 18-core CPU for $2,000. Today, AMD's $750 CPU beats it while drawing much less power. One caveat - Intel's 9900K is still the best gaming CPU, but by a small margin, only if you have a 2080 Ti or such, and it's well behind AMD at everything else.

Intel falls further behind

Intel did manage to release it's 10nm Ice Lake parts, finally, but only for ultraportable laptops. Several years late, they were actually slower than some 14nm parts for CPU performance. Indeed, Intel's 10th gen offers both 14nm and 10nm parts in what is a complete mess of a product line-up. I can't decipher it, and there's no way the average consumer can tell a Intel Core i7 1075G7 apart from a Core i7 10810U. It's ridiculous.

But there's a good reason they had to do this. While Ice Lake offers a significant increase in GPU performance, it's limited to only 4 cores. So for the rest of the line-up, Intel has had to release their old 14nm parts, refreshed with some new features.

While Intel leads new products with laptops, AMD comes to laptops last. Due to this quirk, laptop is one area where Intel beats AMD handily... for now. This is because AMD still hasn't made the transition to Zen 2 and 7nm. But it is coming, very very soon. In fact, by January, we'll have 8 core CPUs in ultraportables. You best believe the 4 core Ice Lake will stand no chance. Not to mention, AMD's IGP will pull far ahead.

I wrote "throughout much of 2019 till Intel's 10nm Ice Lake arrives". While Ice Lake did arrive for small laptops, that's about it. Ice Lake on server is not releasing till late 2020, and all indications point to there being no Ice Lake on desktops at all. All Intel can muster now are more refreshes of their 14nm products which simply is a generation behind AMD's Zen 2 parts. By the time Ice Lake does arrive for server, AMD will be on Zen 3 which promises to be another step forward - though not the giant leaps made with Zen 1 and Zen 2.

Looking forward

I may have painted a rather dour picture for Intel, but the reality is that they have an incredible marketing budget, and continue to rake in the cash. They have long term partnerships with enterprise clients, and laptops are still dominated by Intel. AMD cannot beat Intel - they just need to keep up with the technological leadership and they will grow significantly coming from a state of near-bankruptcy earlier in the decade. For the consumer, this is an incredible win. Just 3 short years ago, Intel was selling their 10 core CPU for $2,000. Today, you can get a 12-core CPU that runs circles around that for $500. For $2,000, you can get AMD's incredible 32 core Threadripper that's at least 4x as fast as Intel's best from 3 short years ago. It just shows how stagnant Intel has been, and how important competition in. Without competition, everyone loses except the monopolist.

It's going to be a painful 2020 and possibly 2021 for Intel. They don't have anything on the horizon that can beat AMD's offerings. Meanwhile, AMD continue to make rapid progress - possibly a result of their smaller, more agile engineering-focused nature.

But don't count Intel out. I expect them to rebound strongly in 2022 or so - they simply must. They have a massive engineering team with some of the most talented people out there. Like they came roaring back with the Core series in 2006, it's likely they'll do the same in 2022. However, this time, AMD are much stronger and it's just as unlikely they'll falter like they did with Bulldozer. In short, competition is here to stay, and that's the best news for everyone.



0
0
0.000
8 comments
avatar

To listen to the audio version of this article click on the play image.

Brought to you by @tts. If you find it useful please consider upvoting this reply.

0
0
0.000
avatar

Questo post è stato condiviso e votato all'interno del discord del team curatori di discovery-it.
This post was shared and voted inside the discord by the curators team of discovery-it


0
0
0.000
avatar

Last time Intel did something great was when they cloned AMD64 architecture. Itanium was interesting, though.

0
0
0.000
avatar

And now they are going to clone AMD's chiplet system. Intel has already showed off packages with multiple dies and die stacking. To be fair, I'm sure they had the same ideas, it's just that AMD got their first.

0
0
0.000