back to article SmartNICs haven't soared so VMware will allow retrofits in old servers

Buyers have not rushed to adopt SmartNICs in their VMware environments, but the virtualization giant will soon support their use in more servers – a move it hopes will spur a new wave of adoption. SmartNICs – aka data processing units (DPUs) – are network cards that include some compute and storage capacity. Hyperscalers use …

  1. Tom Womack

    Is CXL pricing really as keen as claimed here?

    I needed a lot of memory in a server last year, so bought 12x16GB DDR4 sticks from bargainhardware for £42 each. The going rate is now £42 per stick for 32GB sticks, if I wanted 1536GB it would be a thousand pounds even if I threw away all the current DDR4.

    It seems unlikely that a PCIe card with a controller significantly more complicated than the normal kind and 256GB of brand new memory chips can compete with that on pricing.

    Is this a weird artificial market brought into existence by Dell's insane pricing on DDR5?

    1. Snake Silver badge

      Re: Is CXL pricing really as keen as claimed here?

      "It seems unlikely that a PCIe card with a controller significantly more complicated..."

      Which brings up another good question for those who pay attention: exactly what new security flaws will these independent processing units bring to the server environment? If, pretty much, almost every CPU introduced in the past 15 years has had security flaws, exactly what new flaws are going to appear on a device that a device driver-driven, independent processing, *handling every network packet passed through the system* bring to the table? A register overflow flaw can theoretically make every single packet, and therefore every single transaction, open to attack.

      Did anyone test these devices to failure before they rolled this plan out??

      1. talk_is_cheap

        Re: Is CXL pricing really as keen as claimed here?

        Attack from what? One of the points of such a card is that it is running dedicated software, rather than all the general software deployed to the server.

        1. Korev Silver badge

          Re: Is CXL pricing really as keen as claimed here?

          And if the dedicated software and/or hardware has holes then it'll be hackable

          1. talk_is_cheap

            Re: Is CXL pricing really as keen as claimed here?

            You have to hack the system that the device has been installed into, find the device, and then hack it. One of the key aspects of security is to layer the security and having dedicated devices with dedicated APIs/Software layers is a long-established solution.

            It is somewhat easier to assign a security B2 (structured protection) and/or C2 (controlled access protection) rating to a card installed into a general server running general software such as VMWARE Sphere that in turn is running basically random applications.

            1. phuzz Silver badge

              Re: Is CXL pricing really as keen as claimed here?

              If all the network communication is flowing through the device, then it might be possible to hack it directly via malformed packets or something, and then use that as a way to attack the rest of the server.

              For example

  2. ChoHag Silver badge

    Let's see if I have this straight

    New device is made, can't be put in anything or used in the ways that anybody wants, vendor wonders why it won't sell.

    Is that about the gist of it?

  3. Nate Amsden

    can you run mix/matched systems?

    Can you run some servers with SmartNICs and others without in the same cluster/network? I would assume so, as anything the SmartNIC offloading in one server is simply processed in software on the others(that don't have SmartNICs)? But if not then that could be a big roadblock for deployments too.

  4. Simon 49

    Majority of ESXi servers are disk/RAM bound

    With the third place bottleneck CPU, and network I/O below that too. Mostly been selling servers with 2 or 4 25G ports, or 2 X 100G, I've yet to see servers push > about 55 except during bandwidth testing. All of which adds up to SmartNICs not really giving ROI in general purpose use cases. Yes if you're building 5G telecom nodes or a very large private cloud fine, but otherwise VMware's move to limit each socket license to 32 cores runs right into this - where its moved the scale out v scale up value point, plus failure domain becomes huge if you're at only small numbers of huge hosts.

    1. BlueInfra

      Re: Majority of ESXi servers are disk/RAM bound

      There is efficiency to be gained by delivering packets directly to the VM's instead of bouncing the packets through a host networking core. Presume that networking latency would drop a bit and you might get a core back to use for compute.

      Since the products exist there is no need to guess. Data should be available.

      For AWS, their networking core was more complicated, with lots of tunnels, so they got a lot out of the transition.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like