AMD Rolls Out EPYC Embedded 9004 CPUs: Up to 96 Cores

AMD
(Image credit: AMD)

AMD introduced its latest EPYC Embedded 9004-series processors featuring its Zen 4 microarchitecture at the Embedded World 2023 trade show. The new CPUs pack up to 96 cores and are aimed at storage, networking, firewall/security, and industrial applications.

AMD's lineup of EPYC Embedded 9004-series processors consists of ten models with 16, 24, 32, 48, 64, and 96 Zen 4 cores operating at max frequencies of 3.70–4.15 GHz and featuring a nominal thermal design power of 200W–360W. These CPUs are designed for always-on applications, so AMD uses the best silicon for them and tests them vigorously to ensure extended reliability.

(Image credit: AMD)
TOPICS
Anton Shilov
Contributing Writer

Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

  • jeremyj_83
    Too bad there isn't a single socket 16c/32t CPU.
    Reply
  • JamesJones44
    Embedded is a strange name for something that uses 320 watts. Edge might have been a better name.
    Reply
  • Robtechnob
    Embedded may be associated with low power devices in our minds, however the concept has little to do with low power small devices, it's only about embedding computing functionality inside a device. Some machines for industrial applications will have embedded processing that requires a lot of computing power, for example the car you drive has embedded computing and the needs are increasing rapidly as we move towards fully autonomous navigation systems. The increasing computation power draw is already an issue for EVs . Also make note that the embedded EPYCs have a range of specs depending on needs, some are lower power, and there are other embedded processors designed for much lower power needs.
    Reply
  • JamesJones44
    Robtechnob said:
    for example the car you drive has embedded computing and the needs are increasing rapidly as we move towards fully autonomous navigation systems

    The processors used in vehicle ECUs ares low power (were talking cell phone level SoC wattage draw at max). Even the processors used by Lidar and duel camera detections systems are low power compared to the average desktop, these also typically use tensor cores (aka gpu cores) not CPU cores to do the model processing on the collected images. Most cars only produce about 500 watts of power from an alternator, of which most of that power is reserved for the powertrain. You're not putting a 320 watt CPU in an ECU.

    This is also true at the manufacturing level. Most of your PLCs, while not low power, don't need a 96 core general purpose CPU to run a manufacturing line. Most are target tiny core CPUs because they process very small subset of commands (most are ASIC if we want to get specific). Even most Robotic controllers don't use general purpose CPUs and certainly not one that would have 96 cores.
    Reply
  • Robtechnob
    I agree that power consumption is an issue, however it's always an issue no matter what the application may be. Lower power is always better than higher, and it is seen today as a much more important consideration than it was just 10 years ago. As for vehicles, what I wanted say, is that the computing power needs are growing very quickly and it has already become a constraint especially on EVs. Unless something drastically changes with the technology, we're not going to have true self-navigation capabilities on a device as low powered as a cell phone is, however given the high investments going into "AI", we can expect advances that bring down the power needs to more practical levels. As for what kinds of computing devices are used inside emebbed applications, it all depends on the application. In many cases an ASIC will do, in others an FPGA is needed such as with space exploration devices, in other cases a GP CPU or GPU is required - it depends entirely on the needs. The EPYCs in our case are designed for fault tolerant longevity and other abilities such as support for NVDIMM memory. You can google up what Siemens uses the processor for, and also Advantech, these are two design wins mentioned by AMD.

    What started the whole discussion is definitions, for example "edge" is not always well defined, it generally means a kind of application involving communications or data processing (it's a fuzzy thing), but "edge" can include the use of an embedded application, for example computing near a cell tower is an edge device that uses embedded computing.
    Reply
  • JamesJones44
    Robtechnob said:
    What started the whole discussion is definitions, for example "edge" is not always well defined, it generally means a kind of application involving communications or data processing (it's a fuzzy thing), but "edge" can include the use of an embedded application, for example computing near a cell tower is an edge device that uses embedded computing.

    Fair argument. We could flip that and say embedded is also ambiguous. Is it still an embedded system if it has a display and connected to the internet? There are a lot of applications that are embedded that also come with a display and can update themselves (Vehicle Scan tools for example).

    I guess when I think embedded I think power contained and largely self contained in some way. For Edge I feel like it is things not specifically controlled by the central system, but linked to it and possibly remotely controlled by it if needed (Cell Tower example). So my first reaction was it was a strangely named component.
    Reply
  • Robtechnob
    It's true the definitions can get fuzzy. For example we could define an embedded EPYC system as being "embedded" only if it's not inside a server room, but then we have to define what a server room is, and I'm sure you'll agree that what first comes to mind as being a server room can quickly get fuzzy. Maybe a half decent way to describe "embedded" is a device that operates in a mostly self contained environment under conditions that requires a high level of reliability for long periods of time without the need for any maintenance.
    Reply
  • jp7189
    Think "embedded" as in a MRI machine. Tons of power and cooling and image processing needed, but embedded as in it's an integral part of the system that gets certified as a complete system and therefore "upgrades" may not be allowed.
    Reply