Intel returns to boxed workstation CPUs with Xeon 600 — Granite Rapids WS delivers up to 86 cores, 4TB of memory, and 128 PCIe 5 lanes

Intel Xeon 600
(Image credit: Intel)

After nearly three years, Intel is returning to the desktop workstation with the hotly-anticipated (and long-rumored) Granite Rapid-WS series. Now known as Xeon 600, the new series covers all of Intel’s bases for desktop workstations. With previous-gen Sapphire Rapids-WS, Intel split its offerings between two different lineups — Xeon W-2500 and Xeon W-3500 — but all Granite Rapids workstation chips will live under the same Xeon 600 branding.

Granite Rapids has been deployed in the data center for about a year and a half, and in that time, AMD introduced its Zen 5-based Threadripper 9000 chips, leaving an open spot in the market for Intel to release its next-gen workstation CPUs. In total, Intel has 11 SKUs for Xeon 600, five of which will show up as boxed models for individual retail sale. Intel hasn’t provided a firm release date for the CPUs yet, but it says new W890 motherboards and systems from brands like Dell, Lenovo, Supermicro, and Puget will be available starting in late March.

Compared to refreshed Sapphire Rapids-WS parts, Intel broadly claims Xeon 600 delivers up to 9% better single-threaded performance and up to 61% higher multi-threaded performance. The latter metric is explained by much higher core counts on Xeon 600 chips. The previous-gen flagship, the Xeon w9-3595X, topped out at 60 cores. Now, Intel has two SKUs that go above that mark, with the Xeon 698X sporting 86 cores and the Xeon 696X coming in at 64 cores.

Like its data center counterpart, Xeon 600 chips use the Redwood Cove microarchitecture that first debuted in Intel’s Meteor Lake mobile chips. Here, the core counts are massively scaled up, however, and keeping with Intel’s new split of Xeons with a heterogeneous architecture, Xeon 600 CPUs exclusively use P-cores with Hyper-Threading.

Intel Xeon 600 ‘Granite Rapids-WS’ specs

Swipe to scroll horizontally
Row 0 - Cell 0

698X

696X

678X

676X

674X

658X

656

654

638

636

634

Cores / Threads

86 / 172

64 / 128

48 /96

32 / 64

28 / 56

24 / 48

20 / 40

18 / 36

16 / 32

12 / 24

12 / 24

Frequency (Base / Boost)

2 GHz / 4.8 GHz

2.4 GHz / 4.8 GHz

2.4 GHz / 4.9 GHz

2.8 GHz / 4.9 GHz

3 GHz / 4.9 GHz

3 GHz / 4.9 GHz

2.9 GHz / 4.8 GHz

3.1 GHz / 4.8 GHz

3.2 GHz / 4.8 GHz

3.5 GHz / 4.7 GHz

2.7 GHz / 4.6 GHz

All-core Turbo

3 GHz

3.5 GHz

3.8 GHz

4.3 GHz

4.3 GHz

4.3 GHz

4.5 GHz

4.5 GHz

4.5 GHz

4.5 GHz

3.9 GHz

L3 Cache

336MB

336MB

192MB

144MB

144MB

144MB

72MB

72MB

72MB

48MB

48MB

Base TDP

350W

350W

300W

275W

270W

250W

210W

200W

180W

170W

150W

Memory channels

8

8

8

8

8

8

8

8

4

4

4

MRDIMM Support

8000 MT/s

8000 MT/s

8000 MT/s

8000 MT/s

8000 MT/s

PCIe 5.0 Lanes

128

128

128

128

128

128

128

128

80

80

80

Boxed

Yes

Yes

Yes

Yes

Yes

Suggested Price

$7699

$5599

$3749

$2499

$2199

$1699

$1399

$1199

$899

$639

$499

Above, you can see the full list of Xeon 600 SKUs, which range from $499 for the 12-core Xeon 634 up to $7,699 for the flagship Xeon 698X. Note the five SKUs that will be available individually in boxed retail units. Intel says all SKUs will be available in tray, but the boxed lineup tops out at the 64-core Xeon 696X. The split largely mirrors AMD and how it’s treated Threadripper 9000 chips, with the main range topping out at 64 cores, but the Threadripper Pro 9000 WX range climbing up to 96 cores.

Short of the bottom three SKUs, there are some platform features that are consistent across the Xeon 600 range. X-series SKUs are unlocked for overclocking, and they all support octa-channel memory with officially sanctioned speeds up to 6,400 MT/s. You also get 128 PCIe 5.0 lanes, CXL 2.0 support, and Intel’s AMX accelerator in each CPU core, which now supports FP16 instructions. The chips also support up to 4TB of memory, doubling what’s available on AMD’s Threadripper 9000 WX range and quadrupling support compared to the base Threadripper 9000 range. Granted, even a 1TB kit of DDR5-6400 RDIMMs will run you about $28,000 right now.

Intel Xeon 600

(Image credit: Intel)

For the first time in workstation, Intel is bringing support for MRDIMMs, or Multiplexed Rank DIMMs. MRDIMMs are supported in the data center with Xeon 6 CPUs, but this is the first time we’re seeing them in workstation chips. MRDIMMS include two ranks of memory chips along with multiplexer chips. The multiplexer essentially allows the bandwidth of both ranks to combine and double transfer rates — in the case of Xeon 600, with support for up to 8000 MT/s.

MRDIMMs use the same physical connector as RDIMMs, so they drop into existing connectors (assuming the CPU supports MRDIMMs). The applications of MRDIMMs are really only relevant in HPC settings where memory bandwidth is critical, and as such, Intel isn’t supporting MRDIMMs across the full stack. They’re only supported with the top five SKUs, starting at the Xeon 674X.

Intel’s Xeon 600 ‘Granite Rapids-WS’ performance claims

Intel broadly claims a 9% improvement in single-threaded performance and 61% improvement in multithreaded performance with Xeon 600 compared to Xeon W-2500 and Xeon W-3500. According to Intel, those numbers come from scores in Cinebench 2026, as you can see in the disclaimers and configurations we included at the end of the gallery below.

Looking at SPEC Workstation 4, Intel says the flagship Xeon 698X offers a 17% improvement in AI, 22% in the energy subcategory, 61% in financial services, 19% in life sciences, and 10% in media and entertainment compared to the Xeon w9-3595X. In the productivity category, the Xeon 698X posted identical performance, while in product design, it showed an undisclosed regression.

In specific apps, Intel says the Xeon 698X finished the Blender Junkshop render 74% faster than the Xeon w9-3595X, as well as sped up AI-powered upscaling with Topaz Labs Video Upscaler by 29%. Intel attributes the latter speed up to the AMX accelerators inside Xeon 600 cores. To that end, Intel is introducing Open Image Denoise 2.4, which it says is accelerated by FP16 instructions available in the Xeon 600 AMX.

In development, data analysis, and AI, Intel claims 24% better linear algebra performance (as measured in algorithms in Intel’s fork of NumPy/SciPy), 18% faster large dataset analysis with SPEC Workstation 4’s data science workload, and 16% faster AI inference with SPEC Workstation 4’s ONNX inference test.

Critically, Intel didn’t share any competitive benchmarks with Threadripper 9000. In a Q&A session with the press, Intel’s Jonathan Patton said, “We’re looking to be very competitive within the market, offering better performance per dollar for more value for the workstation spend… This is a very highly expandable platform. We have up to 4TB of memory capacity, supporting two DIMMs per channel, our competitors do not. We have advanced instruction sets — AMX, as we mentioned a little bit — and our vPro technology, so we continue to offer a very competitive platform.”

Hopefully, we’ll have those comparisons soon. Intel says Xeon 600 motherboards with the W890 chipset will launch in late March, as will workstations from OEMs like Dell, Lenovo, and Supermicro. We still haven’t learned about a firm release date, nor a release window for boxed Xeon 600 chips.

Google Preferred Source

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

TOPICS
Jake Roach
Senior Analyst, CPUs

Jake Roach is the Senior CPU Analyst at Tom’s Hardware, writing reviews, news, and features about the latest consumer and workstation processors.

  • thestryker
    I'm surprised to see MRDIMM support making it to the workstation parts as I just assumed that would be a point of differentiation. I've been looking forward to seeing what overclocking on GNR looked like in general (first overclockable part on an Intel EUV node), but adding MRDIMMs to the mix is a whole new frontier here. It's a shame these are so far out of my justifiable range of purchase price.
    Reply
  • Stomx
    What overclocking on GNR? What overclocking is at all for the high end workstations be it Intel or AMD or any others? Workstations chips all already overclocked server chips, and to keep with the TDP of additional reduced core count. From 128 to 86 like in this case. Or from 128 to 96 like in AMD Threadripper.

    And with current and near future chips you will think about underclocking them when the total power will hit the limit of the home 110/120V electric power per line because two such chips with 650-750W (when you stress all their resources power can overhit the TDP) plus good GPU will easily kick your circuit breaker. And even if you switch on 220/240V the total 2-2.5kW of heat such system generate will sometimes push you hard to find reasonable cooling solutions for the entire room as well as for each individual processor

    As to the price I'd say it is actually pretty reasonable. If you think differently, buy used and be happy, you will lose nothing, only win. In modern conditions CPU prices are negligible compared to the totally absurd RDRAM prices
    Reply
  • bit_user
    Like with some other recent examples we've seen from Intel, a promising product looks to be undermined by bad launch timing.

    I look forward to benchmarks, but DRAM prices are going to seriously rain on its parade.
    Reply
  • bit_user
    Stomx said:
    Workstations chips all already overclocked server chips,
    Scatterbencher did some interesting stuff with overclocking Sapphire Rapids Xeon w7-3465X. In the end, they managed to achieve over a 50% performance increase on some benchmarks! That's quite amazing, IMO!
    https://skatterbencher.com/2023/09/02/skatterbencher-63-intel-xeon-w7-3465x-overclocked-to-5100-mhz/
    Reply
  • thestryker
    Stomx said:
    What overclocking on GNR? What overclocking is at all for the high end workstations be it Intel or AMD or any others?
    The same overclocking there always is? This has literally been a thing since Intel first started releasing unlocked Xeon SKUs.
    Stomx said:
    Workstations chips all already overclocked server chips, and to keep with the TDP of additional reduced core count.
    This is astoundingly false. If you look at the lineup you'd see that there are equal (and some even higher) TDP server parts at every core count where there is an equivalent to the workstation parts. What the workstation parts have that the server parts don't is turbo boost profiles.
    Stomx said:
    And with current and near future chips you will think about underclocking them when the total power will hit the limit of the home 110/120V electric power per line because two such chips with 650-750W (when you stress all their resources power can overhit the TDP) plus good GPU will easily kick your circuit breaker.
    So now you're making up some random hypothetical about running two high end workstation machines on a home circuit? This is very nonsensical to say the least.
    Stomx said:
    As to the price I'd say it is actually pretty reasonable.
    Who said it wasn't? The price being reasonable is absolutely not the same thing as it being a justifiable purchase.
    Reply
  • Stomx
    bit_user said:
    Scatterbencher did some interesting stuff with overclocking Sapphire Rapids Xeon w7-3465X. In the end, they managed to achieve over a 50% performance increase on some benchmarks! That's quite amazing, IMO!
    https://skatterbencher.com/2023/09/02/skatterbencher-63-intel-xeon-w7-3465x-overclocked-to-5100-mhz/
    That was 7nm 28core chip of 2023. No way you easily overclock current 128-core processor. The further in passed times the larger was overclock. I was probably the first in history overclocker when on the start of PC era by mistake reached 100% overclock of 8086 because could not find its stock 4.77Mhz crystal and used variable frequency LC generator reaching 10MHz !

    We will soon forget about overclocking 2nm and below chips due to potential rapid degradation of their small features specifically sensitive to the top temperatures.
    Reply