Besides Xeon processors that are officially mentioned on its website and price list, Intel has tens of ‘off roadmap’ server CPUs only available to select customers that have special requests. Recently journalists from ComputerBase discovered that Intel has Xeon Platinum 8284, the company’s fastest 28-core chip for multi-socket servers. The CPU runs 300 MHz faster than the ‘official’ Xeon Platinum 8280, but costs considerably more.

Intel’s Xeon Platinum 8284 packs 28 cores with Hyper-Threading that run at 3.0-4.0 GHz, feature a 38.5 MB cache, a six-channel memory controller supporting up to 1 TB of DDR4-2933 with ECC, 48 PCIe 3.0 lanes, and other capabilities found in codenamed Cascade Lake CPUs. Since the chip runs at 300 MHz higher base frequency when compared to the Xeon Platinum 8280, it has a 240 W TDP, up from 205 W. Meanwhile, Tcase of the CPU (the maximum allowed temperature on the IHS of the processor) was reduced to 65°C (down from 84°C), so the CPU requires a very sophisticated cooling system that can take away 240 W at the aforementioned temperature.

Being Intel’s fastest 28-core CPU for multi-socket servers, the Xeon Platinum 8284 processor costs $15,460 (recommended customer price for 1k unit order, RCP), whereas the Xeon Platinum 8280 that runs at a 300 MHz lower frequency, costs $10,009 for 1ku.

Intel Second Generation Xeon Scalable Family
(Cascade Lake)
  Cores Base
Freq
Turbo
Freq
L3
Cache
TDP
(W)
Optane Price
(1ku)
Xeon Platinum 8200
8284   28 3.0 4.0 38.50 240 Yes $15460
8280 L 28 2.7 4.0 38.50 205 Yes $17906
8280 M 28 2.7 4.0 38.50 205 Yes $13012
8280   28 2.7 4.0 38.50 205 Yes $10009
8276 L 28 2.2 4.0 38.50 165 Yes $16616
8276 M 28 2.2 4.0 28.50 165 Yes $11722
8276   28 2.2 4.0 38.50 165 Yes $8719
8270   26 2.7 4.0 25.75 205 Yes $7405
8268   24 2.9 3.9 35.75 205 Yes $6302
8260 L 24 2.4 3.9 25.75 165 Yes $12599
8260 M 24 2.4 3.9 25.75 165 Yes $7705
8260   24 2.4 3.9 25.75 165 Yes $4702
8260 Y 24 2.4 3.9 35.75 165 Yes $5320
8256   4 3.8 3.9 16.50 105 Yes $7007
8253 L 16 2.2 3.0 35.75 165 Yes ?
8253 M 16 2.2 3.0 35.75 165 Yes ?
8253   16 2.2 3.0 35.75 165 Yes $3115

The Xeon Platinum 8284 is not mentioned in Intel’s pricelist, and not under Cascade Lake on Intel's ARK database, but it is searchable if you know the exact number. This typically means that the CPU is only available to select customers or even a customer. That said, it is possible that apart from higher clocks, this 'semi-custom' off-roadmap processor may come with features that go beyond that and this might explain the huge price difference when compared to the model 8280.

Related Reading

Source: Intel’s ARK (via ComputerBase)

Comments Locked

81 Comments

View All Comments

  • Qasar - Friday, July 19, 2019 - link

    " AVX 512 is not in mainstream computers until Sunny Cove " um... i think you might be wrong on that, even just a little, as skylake X and cannon lake both have avx512
  • ravib123 - Friday, July 19, 2019 - link

    The other thing the scalable has going for it I believe is pci-e lanes. I use scalable to build out full lanes of pci-e based flash storage arrays. No AMD options that fit the bill, even the consumer side they still seem very focused on direct cpu performance.

    I don’t see AMD losing to the scalable any time soon.

    Intel always release these special order tray only processors. In the last series of xeon there was a tray only I used all the time because lower power and full speed and the cost difference was about 300$. I can’t tell you how many times customer cooking equipment failed and I was happy with the lower heat cpu resulting in no hysteria.
  • Spunjji - Monday, July 22, 2019 - link

    This doesn't make much sense. EPYC has a lot more PCIe lanes available than Xeon Scalable - 128 PCIe 3.0 on the first generation. Surely that would be superior for building PCIe-based flash storage arrays?
  • phoenix_rizzen - Wednesday, July 24, 2019 - link

    EPYC motherboards have 128 PCIe lanes direct from the CPU (whether single socket or dual socket).

    Xeon SP has 36 PCIe lanes direct from the CPU socket. In order to match an EPYC system, you NEED a 4 socket motherboard, 4 separate CPUs, and 4x the number of RAM sticks to fill it. How is that "better"? You're paying extra for the multi-socket motherboard, extra for the multi-socket CPUs, extra for the DIMMs, extra, extra, extra.

    If you want PCIe lanes, especially for a storage system where CPU speed isn't really relevant to total I/O throughput, you're better off going with an EPYC system.
  • DanNeely - Thursday, July 18, 2019 - link

    To me this sounds like a customer willing to throw money at Intel for even higher binned than normal chips. On the one hand it seems a bit silly, on the other - provided Intel can find enough chips to meet the desired performance target - charging that customer 50% more is a big pile of free money.
  • twotwotwo - Friday, July 19, 2019 - link

    Yep. Don't know that this is the use case, but it's worth a lot to high-frequency traders to be just a *little* lower-latency than competitors, which could drive stuff like this. Wonder if that price would hold up in a world where trades settle on the hour. :)
  • Kevin G - Thursday, July 18, 2019 - link

    I'm curious what other off roadmap chips Intel has.

    No Cascade Lake has on package Omnipath though Intel has hinted that they are holding them back until the 200 Gbit controllers are ready. Though at this pace, that'll be alongside Cooper Lake or Ice Lake-SP now.

    Intel also had on package FPGA options with only one model publicly announced previously: the Gold 6138P. However there were supposed to be far more options off road map. Intel still is talking about this but with the 10 nm delays and lack of public parts, I would presume that all this is now off roadmap.
  • AshlayW - Saturday, July 20, 2019 - link

    This CPU is DOA. Rome 64C will offer almost twice the throughput in MT workloads at the same or lower power consumption, and likely half the price or less.

    Intel Xeon basically is only relevant heavy AVX code, up to avx512, that requires lots of branches or dependency, otherwise you should probably be writing it for a GPU or vector processor. (A GTX 1660 Ti has higher FP32 throughput than this Xeon 28C at an order of magnitude less money).

    Intel needs its 10nm node right now, to have any chance of even competing against Rome based epyc. Even Intel internal memo acknowledged this.
  • chada - Monday, July 22, 2019 - link

    The title reminds me a little bit of sketch by Daniel Tosh. I'll have to clean it up obviously, but the gist is that he's explaining to his girlfriend that it is not her versus a '6' he meets on the road. It is the girlfriend versus EVERY '6' he meets on the road.

    So to say that $5500 for 300 MHz is a bit disingenuous. Instead, it is more like ALL of the 300 MHz. I mean, sheesh 28 cores! Insane! So is $5500, but the comparison holds.
  • speculatrix - Wednesday, July 24, 2019 - link

    I've seen Intel CPU variants in AWS which are unlisted on ark.intel.com. Some are a bit quicker with a few more cores.
    I've always assumed that when Intel do the final die test, they fuse out bad cache and cores. Ones which happen to be perfect, not needing the spare silicon, can then be sold at a premium to select customers.

Log in

Don't have an account? Sign up now