Intel challenges AMD's Epycs with a 144 e-core Xeon

Computex With the launch of its many-cored Xeon 6 processors at Computex on Tuesday, Intel is closer to reclaiming the core-count lead over competitors AMD and Ampere.

With up to 288 efficiency cores, Intel's Sierra Forest CPUs can boast more hardware threads than the latest chips from either of its main rivals. Unfortunately, that 288-core monster is not the part we're getting now.

While Intel has technically kept its promise to deliver Sierra Forest in the first half of 2024, for now we're only getting the smaller 144-core version of the chip. Chipzilla's highest core-count parts won't arrive until early 2025.

Sierra Forest nonetheless represents a fresh chapter for Intel. This Xeon is the first to have been made with Intel 3 process tech – a heavily refined and rebranded version of what used to be called 7nm. It's also Intel's first datacenter chip to employ a disaggregated chiplet architecture that splits off I/O from the compute dies – similar to that used in AMD's Epyc – and Intel's first datacenter platform to use only its efficient cores (e-cores).

Stripped down cores

The last point can't be understated – these are not the same cores as we've seen on past Xeons. Sierra Forest relies on the Crestmont cores used in Intel's Meteor Lake parts, introduced late last year.

And like other e-core equipped chips in Intel's lineup, Sierra Forest's have been stripped down to the basics. The first place you're likely to notice this is the lack of support for simultaneous multi-threading, aka hyperthreading.

Considering that Sierra Forest is aimed squarely at cloud, web scale, digital services, and 5G workloads, that may not be as big a deal as you might think. Hyperthreading is often disabled in these kinds of environments for tenant security, or to improve performance predictability.

The more pressing difference is that these chips lack support for instruction sets we've come to expect from the past several generations of Xeon. Most notably, Sierra Forest lacks support for AVX-512 and Intel's newer advanced matrix extensions used in AI acceleration.

Here's a rundown of how Intel's Xeon 6 e-cores chips compare to its upcoming p-core ones:

While it makes sense to strip out little-used features and pack in more cores, it's a very different approach to the one employed by AMD with its Bergamo and Siena Epycs. Those chips use a shrunken down version of its Zen 4 core that, while 32 percent smaller, is otherwise feature complete. The only concession is that those cores clock lower than the full fat Zen 4 siblings.

No slouch … in the right workloads

Even with its slimmed down cores, Intel boasts that its e-core Xeons are as much as 4.2x faster at workloads like media transcoding compared to its now five-year-old 2nd-gen Xeon Scalable parts. That improvement blunts criticism voiced yesterday at Computex by AMD CEO Lisa Su, who noted that many datacenter CPUs are five years old and looking tired.

Thanks to its substantially higher core density and more efficient process tech, Intel argues that a single 144-core Sierra Forest server can replace as many as three 2nd-gen systems.

Things do get a bit trickier when looking at gen-on-gen performance. Intel hasn't shipped e-core Xeons before, so comparing against its older parts isn't exactly an apples-to-apples comparison.

But assuming your workload doesn't rely on specialized accelerators such as the AVX or AMX support reserved for its performance core (p-core) Xeons – more on those in a bit – Intel's benchmarks show the tiny little cores baked into its new processors are actually quite potent.

According to Intel, its 144-core Xeon can not only match but exceed the performance of its fifth-gen Xeons in a variety of general compute, database, media, and networking workloads, and nearly matches its performance in web workloads.

Obviously, we recommend taking all of these claims with a grain of salt. But the takeaway is clear: Intel is positioning Sierra Forest not only as a cloud and web scale processor, but as a consolidation play for enterprises due to refresh their ageing hardware.

This isn't really that surprising. AMD and Ampere have made similar consolidation claims about their respective many-core chips. If you've got a couple of racks full of systems running Nginx flat out, consolidating that to a smaller server fleet makes sense once the hardware reaches end of life.

A lot of more to come

As we alluded to earlier, the Sierra Forrest SKUs that Intel revealed at Computex Tuesday are the first of several Xeon 6 products due to appear in coming quarters.

You can find a full SKU list for the parts announced today below, but notice that all of these parts fall under Intel's new 6700-series platform.

Intel's 6700 series platform will eventually include both Sierra Forest e-core (6700E) and the upcoming Granite Rapids p-core (6700P). The platform is actually one of two that will fall under the Xeon 6 line up. The second is a much larger 6900-series platform offering substantially higher core-counts, I/O capacity, and memory channels.

Looking at the two, it's hard not to feel like Intel is trying to compress its usual "tick, tock" cadence into a single generation. Despite being based on the same underlying architecture, the two platforms are quite different.

The 6700-series parts are the smaller of the two with support for up to eight memory channels at 6,400MT/sec or 8,000MT/sec when using MCR DIMMS on a p-core Xeon – no MCR DIMM support on Sierra Forest it seems. The chip also boasts 88 lanes of PCIe 5.0 per socket and has a rated thermal design power (TDP) of 350W.

Intel's 6700E parts top out at 144 cores, while its 6700P platform will ship with up to 86 – a full 22 more than last gen.

Curiously, while Intel's dual socket compatible 6700E parts will debut soon, its 6700P parts – which support four and eight socket configurations – won't arrive until Q1 2025.

Instead, the version of Granite Rapids coming in Q3 of this year will employ Intel's larger 6900-series platform.

Bigger and badder chips on the way

Intel's 6900-series is physically much larger. We're told that up to two Sierra Forest or three Granite Rapids compute dies will fit on the package, bringing the max core count to 288 and 128 respectively.

The compute and I/O will also be upgraded to make them more competitive with AMD's latest generation of parts. Intel has long trailed AMD on the number of memory channels, PCIe speed and bandwidth, and Intel's 6900 will narrow that gap considerably.

The chips will boast up to 96 lanes of PCIe 5.0 and 12 channels of DDR5 capable of running at 6,400MT/sec or 8,800 MT/sec when using MCR DIMMS on Intel's p-core Xeons.

For reference, AMD's Genoa parts released in late 2022, boast 128 lanes of PCIe 5.0 and 12 channels of DDR5 memory running at 4800MT/sec.

The larger silicon surface area does come at the cost of higher power consumption, with the platform maxing out at 500W per socket.

Intel's 128-core Granite Rapids 6900P processors are due out in Q3, while its 288-core Sierra Forest 6900E chips will launch alongside its lower core count 6700P parts in Q1 of 2025.

Generational blurring

The "tik-tok" cadence of generational improvements and refinements is generally associated with Intel, but it's one that's become quite blurred in recent generations – particularly in the datacenter.

As a reminder, after years of delays, Intel's 4th-gen Xeon Scalable processors – code named Sapphire Rapids – made their debut in early 2023 only to be supplanted roughly 12 months later by its 5th-gen Emerald Rapids products.

Those two products are notable in that they are reminiscent of Intel's Xeon 6 roadmap. Sapphire Rapids supported mammoth machines, up to eight sockets, while Emerald Rapids was limited to two. The same is true for the 6700P and 6900P-series platforms.

In a sense, Intel's roadmap has become so compressed as to be a rolling release cycle as opposed to distinct generations of products. To this end, Intel's next-gen e-core Xeons are already in the works and planned to debut next year. If, say, a 7700E arrives in the northern spring it'll overlap with the availability of the 6900E due out a few months earlier.

Whether this will simplify or complicate purchase decisions remains to be seen. What we do know is that Clearwater Forest will be the chip to watch: it'll be the first Xeon to ship with Intel's next-gen 18A process node next year.

• We're the Ryzen force in CPUs for AI PCs: AMD
• AMD reveals the MI325X, a 288GB AI accelerator built to battle Nvidia's H200
• Nvidia said to be prepping AI PC chip with Arm and Blackwell cores
• Microsoft invites punters to test drive custom Arm-based Cobalt 100 CPU VMs in Azure

The core wars intensify

Much like the gigahertz race of the late '90s and early 2000s, the past decade or so has been marked by a trend toward ever higher core counts.

AMD arguably fired the first shot with the introduction of Epyc in 2017, pulling ahead of Intel with a 32-core part. Two years later AMD extended its core count advantage with the launch of a 64-core Epyc and in late 2022 and mid 2023, the House of Zen released 96 and 128 core versions of its popular datacenter processors.

Intel, for its part, has trailed AMD’s core count at each turn, only catching up with AMD's Epyc Rome parts with the launch of Emerald Rapids in December.

It has been a similar story among the variety of many-cored Arm processors from Microsoft, Amazon, and Ampere. The latter was among the first to announce a 128-core processor called Altra Max in 2021, and its latest Ampere One chips feature up to 196 cores.

With the arrival of the 128-core and 144-core variants – Granite Rapids and Sierra Forest – arriving this year and a 288-core e-core chip early next, Intel is no longer trailing its competitors in this arena.

That's not to say the fight is over. During its keynote at Computex, AMD revealed its upcoming Turin platform would boast up to 192 cores. We suspect this will be the successor to Bergamo and is using AMD's compact Zen 5c cores, but we'll have to wait until it (probably) arrives in the second half of 2024.

Meanwhile, Ampere recently revealed it's moving to a 3nm process node for the 256-core version of its AmpereOne processor – but that won't arrive until next year. We can also expect core counts among the major cloud providers to inch upwards over the next few years as well.

In any case, it seems the core wars are only heating up. ®

Now read: Intel brings a big fork to a server CPU knife fight.