Intel expands its Atoms’ radius with C3000 SoCs

Intel's Core architecture gets the lion and the tiger's share of attention around here, but the company still makes Atom chips. To wit, Intel just rolled out the Atom C3000 SoCs, its third series of Atom chips. In all, Intel announced 20 new models packing anywhere from two to 16 Goldmont cores, support for 128 GB or 256 GB of DDR4 ECC memory across one or two channels, support for up to four 10 Gigabit Ethernet connections, and Intel's Quick Assist Technology encryption acceleration.

Each pair of cores in each SoC shares 2 MB of L2 cache. Buyers can choose models with anywhere from one to eight pairs of cores. Device manufacturers can add a pair of PCIe x8 slots, up to 16 SATA ports, and USB 3.0 using the six-to-20 lanes of High-Speed I/O (HSIO) connectivity, depending on the exact Atom model. Memory can come in the form of SODIMMs, UDIMMS, or RDIMMs.

Intel divides the Atom C3000 chips into three groups: one intended for use in servers and cloud storage machines, one for network and enterprise storage, and one for IoT applications and extended ambient temperature conditions.

  CPU

Cores

Speed Power HSIO

Lanes

Memory Support Ethernet Quick

Assist

Atom C3955 16 2.1 GHz 32 W 20 2400 MT/s, 2 ch 4 x 10 GbE No
Atom C3950 16 1.7 GHz 24 W 20 2400 MT/s, 2 ch 2 x 10 GbE No
Atom C3850 12 2.1 GHz 25 W 20 2400 MT/s, 2 ch 2 x 10 GbE No
Atom C3830 12 1.9 GHz 21.5 W 12 2133 MT/s, 2 ch 2 x 10 GbE No
Atom C3750 8 2.2 GHz 21 W 12 2400 MT/s, 2 ch 2 x 10 GbE No

Server and cloud storage SoCs

We're talking about an Intel product launch, so product segmentation is the name of the game. The server and cloud storage models shown above are available with eight to 16 CPU cores, but the Quick Assist encryption acceleration is not available. All models in the group have a pair of integrated 10 GbE controllers, except the Atom C3955 that packs four.

  CPU

Cores

Speed Power

HSIO

Lanes

Memory Support Ethernet Quick

Assist

Atom C3958 16

2.0 GHz 31 W 20 2400 MT/s, 2 ch 4 x 10 GbE up to

20 Gbps

Atom C3858 12 2.0 GHz 25 W 20 2400 MT/s, 2 ch 4 x 10 GbE up to

20 Gbps

Atom C3758 8 2.2 GHz 25 W 20 2400 MT/s, 2 ch 4 x 10 GbE up to

10 Gbps

Atom C3558 4 2.2 GHz 16 W 12 2133 MT/s, 2 ch 2 x 2.5 GbE +

2 x 10 GbE

up to 5 Gbps
Atom C3538 4 2.1 GHz 15 W 12 2133 MT/s, 2 ch 2 x 2.5 GbE +

2 x 10 GbE

up to 5 Gbps,

compression only

Atom C3338 2 1.5 GHz 8.5 W up to 10 1866 MT/s, 1 ch 2 x 2.25 GbE No

Network and enterprise storage SoCs

All members of the network and enterprise storage family (save the entry-level Atom C3338) sport at least some form of Quick Assist acceleration. Another interesting note is that some models also support the burgeoning 2.5GBASE-T and 5GBASE-T networking standards. Thus far, the only products we've seen embracing this standard pack Aquantia controllers. Gigabyte has already announced one motherboard based around the Atom C3958 SoC.

  CPU Cores Speed Power HSIO

Lanes

Memory Support Ethernet Quick

Assist

Atom C3808 12 2.0 GHz 24 W up to 20 2133 MT/s, 2 Ch 4 x 10 GbE up to 20 Gbps
Atom C3708 8 1.7 GHz 17 W 20 2133 MT/s, 2 Ch 4 x 10 GbE up to 10 Gbps
Atom C3508 4 1.6 GHz 11.5 W 8 1866 MT/s, 2 Ch 4 x 2.5 GbE up to 5 Gbps
Atom C3308 2 1.6 GHz 9.5 W 6 1866 MT/s, 1 Ch 4 x 2.5 GbE up to 5 Gbps

IoT and extended ambient operation temperature SoCs

The last group of Atoms is the smallest. The number of CPU cores in this lot tops out at 12, but all models support Quick Assist with at least 5 Gbps throughput. Before "IoT" achieved buzzword status, these chips would probably have been marketed for embedded applications. In any case, they are rated to operate at temperatures from a bone-chilling -40° F all the way up to a blistering 185° F (-40° C to 85° C).

According to Intel, Quick Assist technology offers up to 20 Gbps of cryptography throughput and as much as 20 Gbps of compression output, freeing up CPU cycles for other tasks. The new SoC's enhanced networking capabilities look suited to Intel's vision of positioning these systems as network appliances and 5G base stations with server capabilities.

According to Intel's Ark, the top-of-the-line Atom C3958 rings in at $449, though the SoC pricing is probably purely academic because consumers will only find them soldered onto motherboards or as parts of complete systems. The least-capable chip, the Atom C3338, lands at a meager $27. Somewhat more detailed information is available in Intel's Atom C3000 Processor Product Brief.

Comments closed
    • AMDisDEC
    • 2 years ago

    Those extended temp -40C-85C parts are going to do every well in embedded.

    • DragonDaddyBear
    • 2 years ago

    Wow, I hope this takes 10 GbE closer to mainstream. I wouldn’t be shocked to see boards with these cheaper than an excess system + card.

    • jarder
    • 2 years ago

    I see Gigabyte are calling their new board a “Server/Workstation Motherboard” . “server” I get, but “Workstation” is a bit of a stretch, especially considering there are many desktop CPU’s that would make for better workstations.

    • the
    • 2 years ago

    Every time I see some of these neat industrial/embedded parts from Intel that have high core count and low power, I wonder what it’d take to sneak one of these into a laptop. The higher memory support and virtualization extensions could also move this chip if it where in a workstation laptop.

    The obvious is the lack of a GPU but there are enough PCIe lanes to tackle this. Say a C3950 with 16 cores even at 1.7 Ghz, how would that compare to say a quad core 3.5 Ghz i7-7567U. Single threaded performance of the i7-7567 would undoubtedly win but multithreaded results be an interesting comparison. Integer workloads would be determined by how well applications could take advantage of the extra IPC of Kaby Lake over a Goldmont core. I strongly suspect that the C3950 would win here. Who would win floating point workloads would be dependent on applications utilizing AVX/AVX: Twice the clock with twice the vector width vs. a chip with four times the core count. With vanilla SSE code, the results could very well swing in favor of the C3950 where as the i7-7567U would win where it could use AVX. Things like video encoding using Quicksync vs. software codec on the C3950 wouldn’t have a clear winner on paper.

      • NoOne ButMe
      • 2 years ago

      enough money to fund the R&D of a suitable Mobo and everything else into a laptop.
      So I imagine in the millions of dollars range.

      If I ever win the lottery (and I don’t play, so that would be hard) I would love to make a company (after setting aside a suitable chunk for myself) to make stuff like this.

        • the
        • 2 years ago

        There is another potential barrier to this: Intel doesn’t want it to happen. I don’t think they look fondly upon designs that that circumvent their artificial product segmentation by doing something clever.

        I totally understand the desire to start a company and build things like that. It would be expensive but I do see an opportunity in the market. with these chips. Ditto for Xeon D in laptops, especially some future Xeon D’s which will gain an on-die GPU.

      • DavidC1
      • 2 years ago

      The question is:

      How many people are willing to buy a laptop with shoddy battery life(idle of 20W or more) and per thread performance of less than half of most Ultrabooks, and doesn’t support the latest vector extensions?

        • the
        • 2 years ago

        Is the idea really that bad, especially if certain features (i.e. three of the four ethernet controllers) turned off? I would suspect that real battery drain would be the requirement of an external GPU.

        As for demand, I would see this product popular in the workstation laptop niche. Much like most workstation laptops, they wouldn’t be incredibly high volume parts but they would carry a nice price premium. Poor battery life here isn’t as critical if other features can be met. The potential for 128 GB of ECC RAM in a laptop is higher than mobile Xeon E3 v5 chips for example . Due to the previously mentioned battery and higher price (gotta include a GPU), these would not be useful in any sort of main stream application.

        Per thread performance would only be half running AVX workloads where recent Core i7 chips can double their potential throughput using it. As for not supporting the latest vector extensions, this is only a deal breaker if the software strictly requires it. Generally speaking, most workstation software doesn’t require AVX but certainly can take advantage of it. That’s where these Atoms chips would have trouble competing against mobile i7/Xeon E5 chips. However, for multithreaded workloads, even with AVX on modern mobile i7/Xeon E5, the raw 16 cores could overcome the AVX and clock speed advantages in multithreaded workloads.

          • DavidC1
          • 2 years ago

          The server oriented microprocessors do not have the latest power management features of the consumer chips.

          It seems C2000 series have up to C6 power states. Laptop ones go down to C10. Also there’s the practical power usage. Consumer workloads are very bursty, meaning power management is worth doing. Server workloads run constantly. Also reliability issues would prevent extensive power management features from being implemented. Reliability in this case means for performance.

          To reduce platform power(that’s including the display) by half on Haswell Noteboooks they needed to do work with nearly every vendor making the platform.

          On Desktops you have people running on Always On power states despite the fact vast majority of workloads lose minimal amount of performance running on Balanced, which has the upside of saving lots of power when doing nothing. I assume Server folks would be even more cringy bringing Laptop state power states to a platform when it could create issues with performance and compatibility.

          Intel would have to see a market in order to see laptop specific Atom C series available. Market” for Intel means something that could potentially impact their bottom line in a meaningful way. HEDT platforms are probably much larger than the Atom C on Notebook market yet they largely base it off their server platform.

          “Per thread performance would only be half running AVX workloads where recent Core i7 chips can double their potential throughput using it. ”

          It’d be worse. Remember Skylake chips are even faster, probably by 15%. I don’t think Denverton is at Broadwell level per core, they’d need another iteration or two to do that. The Skylake you mentioned also has an advantage of being clocked 2x.

          2x clock
          ~50% perf/clock
          Vector extensions

          Though the 3.5GHz mobile chip you mentioned is a dual core version. There’s that. Still with the massive perf/clock/core advantage it would only be useful in extremely parallel applications. Halfway parallel applications might be merely equal or even worse than the dual core Core i7.

          This doesn’t prevent you from creating a Atom C3000 based Laptop of course. Niche markets exist and vendors are profitable selling for them.

    • ronch
    • 2 years ago

    I wonder just how efficient or inefficient the Goldmont core is if you clock it as high as other proper desktop processors containing the same number of cores.

      • Klimax
      • 2 years ago

      Highly unlikely you’d ever reach those frequencies.
      These designs are highly tuned for far smaller range of operation to achieve much better efficiency. Number of stages, types of transistors and other things. Try to clock too high and signals won’t have time to reach next stage or transistors to switch (or stabilize)

        • tipoo
        • 2 years ago

        It’s a 12-14 stage pipeline, Kaby Lake is 14-16, most of their big cores have started at 14 for years (31 for Netburst, wiki made me guffaw again at that). So it would probably clock a touch lower than them, but probably also has a lot of headroom for clocks over what it’s currently at, just that its efficiency would most likely fall off a cliff a bit further from where it is.

        Now my question is how high 31 stage Netburst could clock fabbed on 10nm with esoteric cooling!

          • tay
          • 2 years ago

          Is it still in-order?

            • srg86
            • 2 years ago

            Silvermont is 2 issue out-of-order.
            Goldmont is 3 issue out-of-order.

            • tipoo
            • 2 years ago

            Atoms went OoO with Silvermont

          • DavidC1
          • 2 years ago

          Jaguar is 14 stages. Pipeline stages are determined by branch mispredict stages. [url<]http://diit.cz/sites/default/files/styles/media_gallery_thumbnail/public/amd_jaguar_slide_11.jpg?itok=R9l8R93n[/url<]

    • NovusBogus
    • 2 years ago

    I’ve heard good things about Goldmont; supposedly the lower power ones can almost challenge a mobile series Sandy Bridge CPU so I’d expect the juiced-up ones to do quite well.

      • Ummagumma
      • 2 years ago

      I have an ASRock motheroard (from Newegg, UEFI BIOS only) with an embedded J4205 Pentium processor/SoC soldered on it. That board is impressive when compiling Linux kernel code; it simply feels “just darn fast & responsive”. The CLI is very responsive, never seems to bog down. The onboard graphics chip of the J4205/J3455 have impressive capabilities; DirectX 12, triple monitor support, H.264 decoder, HEVC/H.265 (10-bit) decoder, VP8 & VP9 decoders. The heat sink on the SoC barely gets warm to the touch even when all CPU cores are busy compiling code.

      I have not checked out the video other than VGA since I find Intel HDMI to have horrible behavior (fails to resync or display output even with “consoleblank=0”) when I power cycle the flat panel display connected to that output. Compare Intel’s poor performance here to a simple, low cost RaspberryPi 3 that will automatically power up that flat panel display when the RPi is powered up and properly resync when that flat panel is power cycled. How an RPi gets this behavior right and Intel can’t pull their thumb out, and Intel hasn’t pulled their thumb out on this issue for years now, is beyond comprehension.

      As for comparing performance to a mobile Sandy Bridge, yeah, I would say they are quite comparable.

      I would love to find a reasonably priced (less than 300 USD new, or “Grade A refurb”) quad core J4205 or J3455-based laptop/netbook style device for portable leisure/coding use. Everything I find seems to be limited to the J3355 dual core SoC.

        • Lianna
        • 2 years ago

        [url=http://www.gearbest.com/tablet-pcs/pp_557447.html?<]CHUWI Hi13 2 in 1 Tablet PC[/url<]

        • srg86
        • 2 years ago

        I’m soon going to be buying the Celeron version of this board with the J3455. I plan on running my normal benchmark of single and multi-threaded GCC compiles, I’m very interested to see how this runs now.

        As for the HDMI issue, I’ve used that on Braswell, Ironlake and haswell and never seen an issue. Could it be your panels?

      • tipoo
      • 2 years ago

      With 12-14 pipeline stages, I wonder how high they could clock. They won’t be as clock limited as say 6 stage Jaguar, so they should go pretty high.

        • DavidC1
        • 2 years ago

        Bobcat – 13 stages
        Jaguar – 14 stages

          • tipoo
          • 2 years ago

          Oh, I guess six was fetch alone, with only three being critical

          [url<]http://meseec.ce.rit.edu/551-projects/fall2015/3-3.pdf[/url<] [url<]https://nathanlamont91.wordpress.com/2015/03/22/my-report-on-the-amd-jaguar-quad-core-cpu/[/url<]

    • xeridea
    • 2 years ago

    The Atom C3750 in your table has power of 2.2 GHz. Is there a translation from GHz to Watts?

      • morphine
      • 2 years ago

      Well, our flux capacitor got stuck at warp 3 speed and our watts-o-meter was measuring in the wrong units.

        • xeridea
        • 2 years ago

        Power is surely more than 2.2W
        [url<]https://ark.intel.com/products/family/29035/Intel-Atom-Processor[/url<]

          • morphine
          • 2 years ago

          Ack, thanks. It’s 21 GHz, er, I mean, 21 watts.

        • curtisb
        • 2 years ago

        But did it go to plaid?

          • K-L-Waster
          • 2 years ago

          That’s ludicrous…

    • DPete27
    • 2 years ago

    [quote<]Gigabyte has already announced one motherboard based around the Atom C3958 SoC.[/quote<] See that!!! You CAN do 4 full sized DIMM slots on a mITX board.

      • EndlessWaves
      • 2 years ago

      Sure, with a smaller CPU socket.

        • DPete27
        • 2 years ago

        There you go. Raining on my parade.

    • DavidC1
    • 2 years ago

    [url<]https://www.servethehome.com/intel-atom-c3955-16-core-top-end-linux-benchmarks-review/[/url<] Performance is really goood. Intel claims up to 3.4x over previous generation. Next generation's Goldmont Plus core is said to be 30% faster. If you consider the lack of Hyperthreading on these parts, with Goldmont Plus it won't be that far behind Broadwell ones.

      • NoOne ButMe
      • 2 years ago

      I recall seeing Goldmont plus being theorized at Ivy Bridge level somewhere recent. wish remembered 🙁

      • tipoo
      • 2 years ago

      Impressive, considering how much further back Silvermont was than their big cores.

      • srg86
      • 2 years ago

      Goldmont is also a 3-issue core, where Silvermont was only dual issue. I’m sure that also helps.

        • tipoo
        • 2 years ago

        Impressive that they’re getting up to Broadwell levels with a 3 issue core. Clearly the larger cores aren’t hitting their full issue widths very often.

          • DavidC1
          • 2 years ago

          Well, programs inherently have limits on instruction level parallelism. The peak issue rate is there just for few cases where burst throughput is needed. I think wide issue cores are more for the vector units.

          The cars are capable of going at 200km/h, but the speed limits are at 120km/h. The higher performing cars merge quicker and accelerate quicker of course.

            • tipoo
            • 2 years ago

            Understandable. I do wonder how far these smaller cores can go, and if they’ll start to get to 80-90% of the most modern large cores performance at a fraction the cost. So ironically Intels true fight from the bottom up may come from…Themselves?

    • hungarianhc
    • 2 years ago

    My FreeNAS server currently runs on an ASROCK board with a C2750 CPU on it. It has been rock solid and fast enough for my needs for a few years. I cannot wait to upgrade to a C3958 based motherboard in the not too distant future. The Supermicro board is fanless, supports 10gigE, IPMI, etc. I’ll upgrade from 32GB to 64GB of RAM as well. I know it will be expensive, but I don’t upgrade this system that often…. For anyone who wants a great, low power, high performance FreeNAS CPU, I recommend this one. Yes yes you can get a Xeon D for cheaper, but the integrated, itx, fanless elegance… and having 16 cores… make this a buy for me :-).

      • Ummagumma
      • 2 years ago

      I would consider “fanless” very very carefully. So consider reading this review of a C3955-based motherboard:

      [url<]https://www.servethehome.com/supermicro-a2sdi-h-tp4f-review/[/url<] I have a C2758-powered video server processing 720p/10fps max. feeds from 7 cameras (with a few more planned) and it's CPU gets quite "toasty" (50+ C.) without added "air cooling". With added "air cooling", albeit "noisy", it gets down to 40-45 C.

        • Anton Kochubey
        • 2 years ago

        50 C is nothing – modern Intel CPUs feel just fine even at 80-90 C.

        • Nomgle
        • 2 years ago

        50c isn’t toasty – these new Atom chips are rated to 85c as per the article.

    • Mr Bill
    • 2 years ago

    That first sentence inplies an exclusivity for Intel Core architecture to that of ANY other vendor. I know you don’t intend that meaning. For example: ‘Gets the lions share of Intel generated excitment around here’.

      • chuckula
      • 2 years ago

      [quote<]For example: 'Gets the lions share of Intel generated excitment around here'.[/quote<] Let's get real. They should have said: [quote<]Intel's Core architecture gets the lion and the tiger's share of [i<]unimpressed, no reason to upgrade from my 2500k[/i<] around here[/quote<]

    • chuckula
    • 2 years ago

    On the spectrum of destructiveness to clothing, you have ThreadRipper way over on the right, and 16-Core Atoms Codename: LintRoller on the left.

      • blahsaysblah
      • 2 years ago

      Did you read the specs on the Giga-byte board? two 10Gb/s SFP+ connections. Four mini-SAS ports for 16 drives. ECC/RDIMM support with 4 sticks for 128GB…

      Drool FreeNAS box.

        • chuckula
        • 2 years ago

        Actually these parts are quite interesting if you are into a low-power NAS box or even a smaller web or database server. The top-end 16 core parts aren’t cheap at about $450 but for this market segment they are quite attractive and actually compare well with the Xeon D in some workloads.

          • blahsaysblah
          • 2 years ago

          It seems like a pretty good board to tuck away as your “storage controller” in your TR build.

Pin It on Pinterest

Share This