The SoC: 48 Falcor Cores, DDR4, PCIe

Two decades ago, when processors were a single core with external memory controllers, external caches, and external IO, routing was comparatively easier than what we have today. Now we have many core systems, multiple cache levels of different varieties, more IO than we can shake a stick at, and it all has to communicate with each other in a low power, low latency and high bandwidth way using a variety of interfaces. For the Centriq 2400, Qualcomm is implementing a number of enterprise requirements as well as integrating its own developed fabric.

For those that have been following our Intel/AMD coverage of late, we discussed how internal coherent fabrics are changing: Intel has moved from a ring-bus topology to a per-core networking mesh, and AMD uses its scalable Infinity Fabric within a die, between dies, between sockets, and from GPUs to memory. In the mobile space, coherent fabrics like ARMs CCI/CCN are typically all the rage, and ARM allows its partners to modify and tune those IPs as they need to (and most do). Rather than using off-the-shelf IP, Qualcomm has stated that its new interconnect is homegrown.

The Qualcomm System Bus (QSB) is a proprietary protocol based, bidirectional segmented ring bus. While Qualcomm shows a ring bus in the image above, we are told that the segmented ring bus might not exactly look like a ring inside the chip – by creating a segmented core-to-core design, it means the cores might not be in a ring at all, with some elements sprouting from off shoots and cores having more than one direction to travel. If Qualcomm were to share a false-color die shot, this would likely be visible. The QSB also allows for multicast on read as well as shortest path routing, which again sounds more like a mesh based networking implementation. Qualcomm quotes a >250GB/s aggregate bandwidth for the QSB.

On the Fabric is everything the system needs: cores, cache, memory, PCIe and IO.

The Centriq family will implement a pair of Falkor cores into a Falkor ‘Duplex’, where each core with have a private L1 cache and a shared L2 cache with ECC. We’ll cover the Falkor design in the next few pages.

For the L3 cache, Qualcomm has not quoted a size but has said that it will scale with the number of cores on the chip. In the above slide it states that it is a distributed unified cache, which can be confusing. Ultimately the cache is fully accessible from all cores, unless a QoS policy is in play, but the cache is likely segmented to allow for the relevant QoS policy tags to bind certain regions to certain cores/VMs. Despite it saying unified, it means that there will be partitions of the L3 around the QSB interconnect. The L3 will be with ECC as well.

Memory controllers are also accessed from the QSB interconnect, with the Centriq 2400 supporting six memory channels of up to DDR4-2667 at up to 2 DIMMs per channel. Support will include RDIMM and LRDIMM, which would suggest up to 1.5TB of LRDIMM support per socket using 128GB LRDIMMs, similar to Intel’s premium memory offerings.

Connectivity comes via 32 lanes of PCIe 3.0, which falls below that offered by Intel (32-44, fewer when chipset level Quick Assist or 10GbE is being used), AMD (128 PCIe in 1P or 2P), X-Gene 3 (42), or Cavium. We probed Qualcomm on features such as NVMe, NVMe RAID, and fall-over support, although in our limited time briefing there was not time to cover it – we might hear more while we are at Hot Chips this week.

Qualcomm has designed the chip as a true SoC such that it doesn’t need a chipset. We’ve confirmed that this is on-die connectivity, rather than via a multi-chip package add-in. The information we have states that the chip will support the usual array of SATA, USB, I2C, UART, GPIO and DMA, although how much of anything has not been stated.

Enterprise Features: Security, QoS, and Secure Boot Getting Intimate with Falkor: The Duplex and Power Management
POST A COMMENT

43 Comments

View All Comments

  • DanNeely - Sunday, August 20, 2017 - link

    Mobile CSS needs fixed. Bulleted lists need to wrap, instead of overflowing into horizontal scroll as they currently do on Android. Reply
  • twotwotwo - Sunday, August 20, 2017 - link

    The things I most wonder about the chip itself are, how much worse will single-thread speed be, and how much better will throughput/$ be?

    For serial speed, there's sort of a sliding scale of "good enough"; you can certainly find uses for chips that are slower than Intel's large cores, but as single-thread perf gets worse and worse, more types of app become tricky to run on it because of latency. So you want latency stats for typical enterprisey Java or C# (or, heck, Go) Web app or widely used databases or infrastructure tools. That's also a test of how well compilers and runtimes are tuned for ARMV8, the chip, and live with may cores, but since early customers will have to deal with the ecosystem that exists today, that's reasonable.

    For cost-effective throughput I guess we need to have an idea of at least price and power consumption, and parallel benchmarks that will hit bottlenecks a single-threaded one might not, like memory bandwidth. And the toughest comparison is probably against Intel's parts in the same segment, Xeon D and their server Atom chips. Something that makes it harder to win big on throughput/$ is that the CPU's cost and power consumption are only a piece of the total: DRAM, storage, network, and so on account for a lot of it. Also, the big cloud customers Qualcomm wants to win probably aren't paying the same premiums as you and I are to Intel.

    Then, aside from questions about the chip itself, there are questions about the ecosystem and customers. There are the questions above of how well toolchains and software are tuned. Maybe the biggest question is whether some big customer will make the leap and do some deployments on lots of slower cores. It might be a strategic long-term bet for some big cloud company that wants more competition in the server chip space, but I bet they have to be willing to lose real money on the effort for a generation or two first.
    Reply
  • name99 - Monday, August 21, 2017 - link

    Intel sells 28 core CPUs that run at 2.1 GHz (and turbo up to 3.8GHz but see below).
    Hell, they sell 16 core systems that run at 2.0 GHz and only turbo up to 2.8 GHz.

    Remember QC is not TRYING to sell these to amateurs, or even as office servers. They are targeted at data warehouse tasks where the job they're doing will be pretty well defined, and it's expected for the most part that ALL the cores will be running (ie when the work load lightens, you shut down entire dies and then racks, you don't futz around with just shutting down single cores).
    For environments like that, turbo'ing is of much less value. QC doesn't have to (and isn't) targeting the entire space of HPC+server+data warehouse, just the part that's a good match to what they're offering.
    Reply
  • Threska - Tuesday, August 22, 2017 - link

    Well there is the small developer virtualization market like Ansible. Reply
  • prisonerX - Monday, August 21, 2017 - link

    A preoccupation with single thread performance is the domain of video game playing teenagers and not terribly important, neither is the "latency" you refer to. This sort of high-core, efficient processor is going to be used where throughput and price/power/performance ratio (ie, all three, not just any one of those), are the key metric.

    Latency is mostly irrelevant since processing will be stream oriented and bandwidth limited rather than hamstrung by latency (thus features such as memory bandwidth compression). Gimmicks like "turbo" (which should be called by its proper name: "throttling") and favoring single thread performance are counterproductive in this mode. Being able to deploy many CPUs in dense compute nodes is what is required and memory, storage and networking are minor parts here.

    I don't know why you think compilers or runtimes is a concern, 99.9% of code is common across archs, so if you've supported a lot of x86 cores your code is going to function well for a lot of ARM cores with a small amount of arch specific configuration. The compilers themselves, namely GCC and LLVM are mature as is their support for ARM.

    Finally the new ARM CPUs don't have to beat Intel, just stay roughly competitive, because the one thing the tech industry hates more than a monopoly is a monopoly that has abused its monopoly powers, and Intel is it. Industry is itching for an alternative, and near enough is good enough.
    Reply
  • deltaFx2 - Thursday, August 24, 2017 - link

    "Industry is itching for an alternative": While this is true, is the industry truly interested in an alternative ISA, or alternative supplier? Because there is one now in the x86 space, and is very competitive, and in some metrics better than Intel. Also, your argument about single threaded performance being irrelevant in servers is false. A famous example of this is a paper in ISCA by google folks arguing in favor of high IPC machines (among other things). They also note that memory bandwidth is not as critical as latency. Now this is specific to google, but in plenty of other cases too, unless you have a very lopsided configuration, bandwidth doesn't get anywhere near saturation. There are also plenty of server users who provide extra cooling capacity to run at higher than base frequencies because it's cheaper than scaling out to more nodes. Obviously, your workloads should scale with freq.

    "Finally the new ARM CPUs don't have to beat Intel," -> change intel to AMD. AMD is hugely motivated to compete on price and has the performance to match intel in many workloads. And AMD's killer app is the 1P system, exactly where Qualcomm intends to go. You also have to add the cost of porting from x86->ARM (recompile, validation, etc). Time is money and employees need to be paid. So the question is, why ARM? More threads/socket? Nope. More memory/socket? Nope. More perf/thread? Probably not based on the architecture described but we'll see. More connectivity then? Nope. Lower absolute power? Maybe. Lower cost? I suspect AMD's MCM design is great for yields. And there's the porting cost if you're not already on ARM.

    There's a lot more work to be done and money to be spent before ARM becomes competitive in the mainstream server space. QC has the deep pockets to stick it out, but I am not sure about cavium.
    Reply
  • Gc - Sunday, August 20, 2017 - link

    Confusing terminology: prefetch vs. fetch

    Prefetch heuristics predict *future* memory addresses based on past memory access patterns, such as sequential or striding patterns, and try to prefetch the relevant cache lines *before* a miss occurs, attempting to avoid the cache miss or at least reduce the delay. A memory fetch to satisfy a cache miss is not a prefetch.

    Slide: "Hardware Prefetch on L1 miss."
    Text: "An L1 miss will initiate a hardware prefetch."
    The initial fetch is after the miss, so the initial fetch is not a 'prefetch'. I assume this means that it is not only fetching the missed cache line but also triggering the prefetchers to fetch additional cache lines.

    Slide: "Hardware data prefetch engine Prefetches for L1, L2, and L3 caches"
    Text: "If a miss occurs on the L1-data cache, hardware data prefetchers are used to probe the L2 and L3 caches."
    The slide is saying that data is prefetched at all three levels of cache. I'm not sure what the text is saying. Probing refers to querying the caches around the fabric to see which if any holds the requested cache line. This is part of any fetch, not specific to prefetching. Maybe the text is trying to say that the prefetchers not only remember past addresses and predict future addresses, but also remember which cache held past addresses and predicts which cache holds fetched and prefetched addresses?
    Reply
  • YoloPascual - Monday, August 21, 2017 - link

    Inb4 fanless data centers near the equator. Reply
  • KongClaude - Tuesday, August 22, 2017 - link

    'however Samsung does not have much experience with large silicon dies'

    I don't remember the actual die size for the DEC Alpha's that Samsung fabbed back in the day, the Alpha was a fairly large CPU even by todays standard. Would they have let go of that knowledge or is Alpha being relegated to low-volume/not much experience?
    Reply
  • psychobriggsy - Tuesday, August 22, 2017 - link

    We should also consider that GlobalFoundries licensed Samsung's 14nm after digging their own 14nm hole and failing to get out of it, and right now AMD are making 486mm^2 Vega dies on that process. The process doesn't have a massive maximum reticle size however, IIRC it's around 700mm^2, whereas TSMC can do just over 800mm^2 on their 16nm. Reply

Log in

Don't have an account? Sign up now