In 2019 people have been talking nonstop
about AMD’s chiplets and the pros and cons of this modular approach to
building chips. Of course we shouldn’t expect Intel to
stand still and let AMD take all the chiplet glory. I’ve gotten some
information about an Intel CPU codenamed Clarkdale that is also chiplet-based
let’s take a look at it. Here’s Zen 2 on the left and Intel Clarkdale on the right. In
Zen 2, AMD has separated the CPU and cache onto a die or chiplet made
with the latest 7 nm process while all the I/O is on another die made
with a cheaper process. I apologize that the Clarkdale is a little dirty but it’s
the best photo my source could come up with. We can immediately see some
similarities: both have two dies on an organic package and neither one is using
an interpose or EMIB. These photos aren’t exactly to scale so don’t try to
infer too much about the die sizes. The larger die on top is the I/O die with
the memory controller and PCIe and the smaller die on the bottom holds the
cores and the cache. In both processors the CPU die is made on a newer more
efficient fab process while the I/O die is made using an older process that is
cheaper. This is because I/O doesn’t benefit as much from the higher
transistor performance of newer fab processes. The two dies are connected by
an on-package coherent interconnect AMD calls this Infinity Fabric and Intel’s
is QPI. One thing chiplet design really makes obvious here is that the cores
take up less than half of the space of a modern CPU. There are some differences in
approach between Intel and AMD. AMD left room for a second CPU chiplet so they
can scale from 4 up to 16 cores in the same package. Intel decided to have just
one CPU die here although they clearly have room for two but they integrate
a GPU into their I/O die. These processors aren’t targeting exactly the
same market because Matisse is for the upper mainstream market using discrete
graphics while Clarkdale is for a lower end market that might want integrated
graphics. Now let’s look at some more detailed
specs of these chips. I want to caution you that specs can change before
something is released. Clarkdale is clearly a low-end product using the
tried-and-true 32 and 45 nm processes and we can see Intel has only
put 2 cores with 4 MB of cache while AMD has 8 cores per chiplet with pretty
massive cache. Now this low-end theme continues with I/O where Intel is using
pretty old DDR3 and PCIe 2.0. Not surprisingly, AMD’s 8 core has a higher
TDP than a dual core although 87 watts is pretty high for only two cores. You’ve
probably seen that AMD released Ryzen 3000 a few months ago and Intel is
planning to release Clarkdale in… wait 2010? This can’t be right. Did Intel copy
chiplet technology and send it back in time almost 10 years? At this point you
may realized I’m just trolling you… or am I? What is the difference between chiplets and multi chip modules that have existed for decades? Is an I/O die just
another name for a northbridge? Does the history of PC hardware repeat in cycles
of integration and decomposition? Is time a flat circle? You decide. And by the way
my source for this video it’s not a leak from somebody in Taiwan; it’s memory. In
the PC industry does anyone remember or care
about anything that happened more than a few years ago? Do we have so little to learn
from 2010 that it’s pointless to even bring it up? I don’t know, maybe not. Next
time I swear I’ll have some more serious analysis. Sadly I don’t have any
interesting information about Intel’s real chiplet CPU Rocket Lake, at least
not yet. Come back later and we can talk about Ice Lake, the Zen 2 die, some
thoughts about Zen 3, we might even take detours to the cloud and low Earth orbit
until then I guess you can troll me, I mean continue the conversation on Twitter @wmf.

Intel’s First Chiplet CPU Revealed
Tagged on:                     

8 thoughts on “Intel’s First Chiplet CPU Revealed

  • October 11, 2019 at 2:22 am
    Permalink

    What a meme lord.

    Reply
  • October 21, 2019 at 2:05 pm
    Permalink

    Look at this absolute champion.

    Reply
  • October 21, 2019 at 4:21 pm
    Permalink

    Nice patent infringement

    Reply
  • October 21, 2019 at 5:24 pm
    Permalink

    Scale and modularity. that is the difference. It's been boring thinking with "new" chiplets but that will change very very soon. High end GPUs and CPUs are comming in the very near future in a chiplet style package. Server first of course.

    Reply
  • October 21, 2019 at 6:12 pm
    Permalink

    Hmmm…

    Reply
  • October 21, 2019 at 8:44 pm
    Permalink

    oi IBM Power4 cpu's want their chiplet design back (2001)

    Reply
  • October 28, 2019 at 11:50 am
    Permalink

    Opened my old low end laptop to repaste cpu and found a chiplel CPU.
    Got bamboozled for sure.

    Reply
  • October 29, 2019 at 9:35 pm
    Permalink

    Chiplets actually go back even further than this, the Pentium D for example.

    Or the QX9300 which are two X9100 cores dies on a single substrate. Or the Core 2 Quad!

    Reply

Leave a Reply

Your email address will not be published. Required fields are marked *