Intel’s Xeon 600 Brings Granite Rapids Muscle To High-End Desktops

Some folks act as if Intel and AMD have killed off the "high-end desktop" or HEDT market, but that's never really been the case. Instead, what happened is that mainstream desktops went from a maximum of four cores to 16  - 24 cores basically overnight, propelling HEDT systems out of reach for most enthusiasts. If you want a processor with more cores than you can shake a stick at, though—exclusively P-cores, too—as well as a humongous pile of PCI Express connectivity and memory bandwidth, Intel and AMD will gladly sell you one.

What Make Intel's Xeon 600 Processors Unique?

Indeed, that's exactly what Intel is announcing today: the Xeon 600 Processors for Client Workstations, which replace both the extant Xeon W-2500 and Xeon W-3500 series CPUs. As Intel says, it's doing a bit of product stack compression with this release, but we'll get into that in a moment. The point is, these are new Xeon processors meant for workstations, based on the Granite Rapids architecture.


Intel Xeon 600 series chips are structured in quite a different way from the previous generation; if you want to read more about that, check out our previous coverage of Granite Rapids. Intel says that these parts are fabricated on Intel 3, just like the Granite Rapids server processors, and that they are based on the Redwood Cove core architecture which debuted in Meteor Lake in the client segment back in December of 2023. Note that a portion of the 61% multi-threading performance jump here comes from comparing a 60-core processor to an 86-core CPU, but that's still a considerable per-core advantage for the new CPUs.


Big changes this time around include a humongous jump in L3 cache, but we'll get to that in a second. Indeed, many of the other changes are directly focused at the AI market, of course. As before, all 86 cores support AVX-512 and Intel's AMX extensions, although these chips are the first from Intel to support the FP16 data type. Intel says these chips are "the AI developer's dream," as they are based on the same architecture as Xeon server processors, and thus provide a consistent development platform.

Intel Xeon 600: Server Processor With A Client Chipset

As Intel says, HEDT is really "server CPUs with a client chipset," and that's what this slide is illustrating. The platform combines the Xeon 600 CPUs, sporting up to 86 CPU cores, 128 PCI Express 5.0 lanes, new CXL 2.0 for CXL Memory support, with the W890 chipset that newly includes Wi-Fi 7—the kind of thing that's important for a client workstation. Aside from that, the chipset platform doesn't change for this generation.


Just like the Granite Rapids server chips, the workstation version supports Multiplexed Rank DIMMs for improved memory bandwidth. We've reported on this before and have a decent explainer, but the short version is that it's basically like doing dual-channel accesses to each individual memory module. This improves bandwidth—up to 8000 MT/s transfer rate on these parts—but it does have a downside, which is reduced max capacity. Intel didn't say what the max capacity is with MRDIMMs, but it is going to be lower due to controller limits.


Here's the slide you've all been waiting for, which is the table of processor SKUs. As you can see, the higher-end configurations do indeed get a massive jump in L3 cache from previous-generation parts. Curiously, Intel's presentation credited the move to Redwood Cove for the huge bump in L3 cache even though that's really more of a platform design decision than anything to do with the processor core itself.

Only some of the SKUs will be getting boxed versions, and Intel restricts both overclocking features and MRDIMM support to the higher-end models in the family. The latter makes perfect sense; MRDIMMs' higher bandwidth isn't really necessary on the chips with fewer CPU cores. It doesn't offer reduced latency, only higher bandwidth, and that's helpful when you're doing a lot of crunchy AVX math, but the 24-core and below models simply don't have the core count to need the higher bandwidth.


On this slide Intel explains its decision to lock off overclocking only to the higher-end models. The company says that it's really intended as an option for system integrators and VARs who want to offer tuned configurations to their customers, and are willing to take on the increased support costs. That's totally valid; those tuned configurations are often less about what we normally think of as "overclocking" and more about power and platform tuning. Even still, this market is small compared to the market of adult enthusiasts who would be willing to pay perhaps $999 for a big fat CPU that they can then tune for higher clock rates to improve performance in client workloads. Unfortunately, the way Intel has segmented it, the cheapest chip for this use case is the $1699 Xeon 658X.

While that model retains the full 128 PCI Express 5.0 lanes and 8-channel memory support, it "only" has 24 CPU cores, which we must point out is the same number as the Core Ultra 9 285K, a chip which runs nearly a full 1 GHz faster than this Xeon part. Sure, those are a mix of P-cores and E-cores where this chip is all P-cores, but as an enthusiast myself, it's hard to justify, unless the increased bandwidth and PCIe connectivity that come with an HEDT system are an absolute necessity.

How Do Intel's New Xeon 600 Series Processor Perform?

We won't spend a lot of time on this, but Intel shared some benchmarks showing off the performance of the new chips. The first slide is showing SPEC Workstation 4.0 speed, offering a range of performance results from +/-3% on the low end all the way up to 61% faster in Finance Services when comparing against the previous generation Xeon w9-3595X. On the second slide, Intel claims performance gains 27% and higher in three professional benchmarks.


In these two slides, Intel emphasizes its rendering performance, where it claims to be 74% faster than the previous generation in Blender CPU, and 29% faster in Topaz Labs AI video upscaling. The company celebrates the release of Intel Open Image Denoise 2.4, which is a bit like NVIDIA's DLSS Ray Reconstruction for offline rendering. It's accelerated on the AMX units of the new CPUs utilizing the added FP16 functionality. Pretty cool, although Intel didn't give any performance numbers or other comparisons.


Finally, Intel says the new chips are faster at AI versus the previous generation. Xeon 698X is apparently 24% faster at linear algebra, 18% faster at large-set data analysis, and 16% faster at CPU-based AI inference versus the previous generation. We'd love to see benchmarks versus AMD's latest EPYC CPUs, which are also very good at AI.


Now, this part is a very interesting move from Intel. These new chips absolutely do have their place, but as Intel itself points out, it's not as clear-cut as "desktop versus workstation" anymore, because modern desktop CPUs are so fast. Intel recommends a "desktop workstation" with a Core Ultra 200 processor and a W880 chipset for "latency-sensitive interactive & moderately-threaded workloads," while suggesting that the Intel Xeon 600 CPUs offer better performance in "highly-threaded, high-memory bandwidth, large memory-capacity, & high PCIe expandability workloads," with examples of each.


Intel presents these figures, comparing a Core Ultra 5 245K against a Xeon 636. The comparison isn't particularly flattering considering that the Core Ultra part is all of one-third the cost of the Xeon machine, before factoring in the cost of motherboards for the monster socket and memory to fill the four channels, but that's exactly the point. The Xeon 636, despite losing in some workloads, can simply do things that the Core Ultra can't. It's a candid and clear declaration from Intel's marketing and we approve.


That pretty much wraps it up. The new Xeon 600 processors come with up to 86 modified Redwood Cove cores, fat L3 caches, MRDIMM support, up to 128 PCIe 5.0 lanes, CXL 2.0 support, all the vPRO management features IT departments know and love, and radically simplified naming versus the previous generation. Hopefully, that instinct for simplified naming carries over to the company's client CPU division soon too. Intel says the Xeon 600 Processors for Client Workstations will be available starting in late March of this year.