AMD and Intel differ so much in their approach to CPU design that trying to compare Ryzen Gen 3 to Intel Coffee Lake Gen 10 is a chalk and cheese endeavour.
While both want to dominate the x86 world, AMD and Intel are significantly different in concept and design – but there is room for both.
The critical difference that separates
AMD and Intel processors today is that AMD uses a chiplet design, while Intel
uses a monolithic approach.
But what do these terms even mean? And what do the differences between AMD and Intel really mean for you as a customer?
Gamer, author and part-time Gadgeteer Arjun Lal writes:
AMD and Intel do the same but very differently – Let’s take a deep dive
Intel’s Monolithic Approach: Big, single and increasing thinner (nm) chips
In Intel’s monolithic approach, everything about a processor – all its cores, cache, graphics, and I/O – is on a single die. With smaller processor designs–between 2-4 cores this is easy to do. Each die is relatively tiny, and you fit lots onto silicon wafers – leading to a relatively lower cost to manufacture a complete system on a chip.
Even if defect rates (yield) were to increase, wastage levels are low enough that Intel’s marginal cost of production isn’t hit. Every chip Intel builds–from dual-core Celerons right up to 28-core Xeon behemoths has its own die and design.
To be fair to Intel, it will move to a hybrid chiplet design if only to offer more options. In the example below it could use the same monolithic CPU core die and add custom I/O, Comms (e.g. 5G) and more making it a very flexible chip.
AMD’s Approach: Multiple, discrete chiplets build the chip
AMD’s chiplet design is the opposite.
The fundamental unit of every Ryzen processor is the CCX, a four-core/eight-thread
set of CPU cores. The CCX is a quad-core processor. Two CCXs pack together with
L3 cache and I/O on a separate die to form a CCD.
And up to seven CCDs can stack
together in a Ryzen Multi-Chip-Module (MCM). CCXs in a CCD communicate with
each other, and other CCXs via the Infinity Fabric interconnect. Infinity
Fabric binds the CCXs to the cache, I/O and interconnects different CCDs.
Monolithic vs Chiplet: What are the implications?
First, Intel’s per-unit production cost is lower with smaller processor designs. A dual-core Celeron or Pentium sits on a small low cost, high-yield die.
Secondly, Intel’s performance across
cores is better due to a physics: a monolithic die has less of latency than connected
While Infinity Fabric is very fast, there’s still a latency
penalty as the CCDs are physically separated. Electrons can only travel so fast,
according to Galileo.
With Ryzen Gen 1 and 2, the latency
penalty had a noticeable performance impact, especially in games. However, AMD
has significantly increased cache size up to 72 MB for the Ryzen 3900X. This,
together with intelligent scheduling buffers tasks so that latency isn’t a big
Why Ryzen’s hard to beat on price
Linear and Exponential Price scaling does it every time!
With Ryzen, the smallest CPU is a single
CCD, with 2 CCXs, effectively an eight-core processor. How on earth do they
sell quad-cores and Hexa-cores, then? This is where the real advantage of the
chiplet design shows itself.
No matter how high silicon yields are,
there’s always a defect rate. And on a newer process like the 7nm process, that
defect rate can be 20% or even higher. Every
core on a processor has a chance of being defective.
So, the more cores on a single die,
the more likely it is that at least one core is defective.
In Intel’s monolithic die this means throwing
out the whole CPU if even one core is defective. This causes Intel’s
manufacturing cost scale to scale exponentially up with core count. An 8-core
Intel part costs them far more than
twice as much as a quad-core.
The anatomy of a Ryzen Gen 3: Salvage parts and core scaling
With Ryzen core count scales
near-linearly with price. Since each Ryzen CPU is made up of CCXs (quad-core
units), AMD just has to watch yields on that quad-core building block.
Wastage is far, far lower than for a
monolithic eight-core or sixteen-core die. And even when individual CCXs have
defects, AMD doesn’t throw them away. Instead, they disable one or more cores
to help create cheaper SKUs.
For instance, the hexacore Ryzen 5
3600, a price-performance champion, has the same
core layout as the pricier, octa-core Ryzen 7 3700. It has the faulty CCX
disabled on each CCD, for a total of 6 cores. By utilising salvage parts and
multi-chip scaling, AMD keeps prices low, even when core counts rise.
So, Intel uses large, single dies,
while AMD makes use of multiple, smaller dies put together. It’s hard to
understate how significant an impact this has had on the CPU market over the
past two years.
Exactly how did Ryzen shake up the market?
Since Intel’s Sandy Bridge (2011) it has
been content to sell mainstream quad-core CPUs with relatively high
AMD emulated Intel’s going multi-core with Bulldozer, but that experiment failed miserably.
Its failure lead to AMD’s success. Its
new Ryzen chiplet architecture was able to offer significantly improved instructions
per cycle (IPC).
With Ryzen Gen 1, AMD could offer more
cores and more threads at every single
price-point than Intel.
The Ryzen 3 1200, with four-cores
and four-threads, went up against dual-core i3 and Pentium.
The Ryzen 5 1600 had six cores
and 12 threads chasing Intel i5 buyers.
This was unprecedented: Intel’s i7s
were quad-core eight-threads, and only its ultra-premium HEDT parts offered
more than four cores.
Intel still managed to deliver better
single-threaded performance, and for some time, its processors were still
meaningfully better for gaming.
Ryzen Gen 3 changed that. IPC
improvements meant that clock-for-clock, it outperformed Coffee Lake. Ryzen’s multiple ‘parts’ clock lower than Intel
counterparts offsetting this advantage slightly.
With no noticeable gaming advantage to
Intel, buying Ryzen simply became a no-brainer. In DIY markets like South Korea AMD has over 50% Ryzen
market share – numbers AMD hasn’t hit in over a decade.
Intel takes drastic steps.
Starting with Coffee Lake, core counts
increased. The i3 became four-cores while i5s offered six-cores, and so on.
Intel also cemented its overall
performance dominance by adding another performance tier at the top-end, the
premium i9 models – these were clocked high enough and offered enough cores and
threads to keep even the fastest Ryzens at bay.
This gave Intel nominal performance
superiority, even though you could probably count the number of Australian i9
buyers on one hand.
Intel has also been making noises about multi-chip designs: Foveros 3D stacking, which debuted with Ice Lake, allows them to deploy multi-chip modules with higher core counts, onboard memory and more. It’s Project Athena laptops are all about this.
AMD and Intel – the Chip war is good for customers
Competition is a great thing for
customers. Intel no longer has a monopoly. It has to deploy its massive
resources – an order of magnitude more than AMD – to deliver competitive
The years to come might not be the
best for its shareholders, but customers will have no shortage of great CPU
GadgetGuy’s take – AMD and Intel – may the Core Wars continue
AMD and Intel are a little like the Qantas
and Virgin air war – they fight over the 1% difference between 49 and 51% market
share while mostly doing the same as each other.
Arjun has explained that while AMD and Intel do the same, e.g. x86 it is not the same under the surface. But does it matter?
Perhaps to gamers and high-performance computing but to Joe and Jane Average – no.
With the advent of Ryzen, notably Gen 3 (3, 5, and 7 series) and its low power U versions for laptops there is no stigma or downside in buying AMD for Windows-based PCs. But as my dad used to say, “It is easy to sell a cheaper product. It is much harder to sell a better one”. AMD cannot claim victory unless it is a better chip.
But Intel is serious in wanting to maintain its lead. Its 10th generation i3, i5 and i9 processors are spectacular and more affordable. You can read more here (skip to the end of this ITy Bytes article)