Following on from the SoC disclosure at Scorching Chips, Qualcomm has this week introduced the formal launch of its new Centriq 2400 household of Arm-based SoCs for cloud functions. The highest processor is a 48-core, Arm v8-compliant design made utilizing Samsung’s 10LPE FinFET course of, with 18 billion transistors in a 398mm2 design. The cores are 64-bit solely, and are grouped into duplexes – pairs of cores with a shared 512KB of L2 cache, and the highest finish design can even have 60 MB of L3 cache. The total design has 6 channels of DDR4 (Supporting as much as 768 GB) with 32 PCIe Gen Three.zero lanes, assist for Arm Trustzone, and all inside a TDP of 120W and for $1995.
We lined the design of Centriq extensively in our Scorching Chips overview, together with the microarchitecture, safety and new energy options. What we didn’t know have been the precise configurations, L3 cache sizes, and some different minor particulars. One key metric that semiconductor professionals are interested by is the affirmation of utilizing Samsung’s 10LPE course of, which Qualcomm states gave them 18 billion transistors in a 398mm2 die (45.2MTr/mm2). This was in comparison with Intel’s Skylake XCC chip on 14nm (37.5MTr/mm2), however we also needs to add in Huawei’s Kirin 970 on TSMC 16FF (55MTr/mm2). At the moment Qualcomm is releasing all this info, together with a extra detailed block diagram of the chip.
The chip has 24 duplexes, basically grouped into units of 4. Connecting all of them is a bi-directional segmented ring bus, with a mid-silicon bypass to hurry up cross-core transfers. This ring bus is ready with 250 GBps of combination bandwidth. Proven within the diagram are 12 segments of L3 cache, which implies these are shipped with 5 MB every (though there could also be extra for yield redundancy). This provides a metric of 1.25 MB of L3 cache per core, and for the SKUs beneath 48 cores the cache is scaled accordingly. Qualcomm additionally integrates its inline reminiscence bandwidth compression to boost the workflow, and gives a cache high quality of service mannequin (as defined in our preliminary protection). Every of the six reminiscence controllers helps a channel of DDR4-2667, with assist as much as 768GB of reminiscence and a peak combination bandwidth of 128 GB/s.
|Qualcomm Centriq 2400 Collection|
|AnandTech.com||Centriq 2460||Centriq 2452||Centriq 2434|
|Base Frequency||2.2 GHz||2.2 GHz||2.Three GHz|
|Turbo Frequency||2.6 GHz||2.6 GHz||2.5 GHz|
|L3 Cache||60.zero MB||57.5 MB||50 MB|
|PCIe||32 PCIe Three.zero|
|TDP||120 W||120 W||110 W|
Beginning with the chips on provide, Qualcomm will initially present three completely different configurations, beginning with 40 cores at 2.Three GHz (2.5 GHz turbo), as much as 46 and 48 cores each at 2.2 GHz (2.6 GHz turbo). All three chips are considerably equal, binned relying on energetic duplexes and cache, with $1995 set for the highest SKU (different costs weren’t revealed). Qualcomm is aiming to assault present x86 cloud server markets on three metrics: efficiency per watt, general efficiency, and value. In that regard it supplied three distinct comparisons, one for every chip:
- Centriq 2460 (48-core, 2.2-2.6 GHz, 120W) vs Xeon Platinum 8180 (28-core, 2.5-Three.eight GHz, 205W)
- Centriq 2452 (46-core, 2.2-2.6 GHz, 120W) vs Xeon Gold 6152 (22-core, 2.1-Three.7 GHz, 140W)
- Centriq 2434 (40-core, 2.Three-2.5 GHz, 110W) vs Xeon Silver 4116 (12-core, 2.1-Three.zero GHz, 85W)
Qualcomm supplied some SPECint_rate2006 comparisons between the chips, displaying Centriq both matching or successful in efficiency per thread, beating in efficiency per watt, and as much as 4x in efficiency per greenback. It must be famous that the info for the Intel chips have been interpolated from different Xeon chips.
One attention-grabbing bit of knowledge from the launch was the ability consumption outcomes supplied. As a server or cloud CPU scales to extra cores, there’ll undoubtedly be conditions the place not all of the cores are at all times drawing energy, both because of how the algorithm works or the system is ready on information. Usually the TDP values are given as a measure of energy consumption, regardless of the precise definition of thermal dissipation necessities – a 120W chip doesn’t at all times draw 120W, in different phrases. To this finish, Qualcomm supplied the typical energy consumption of the 120W Centriq 2460 whereas working SPECint_rate2006.
It reveals a median energy consumption of 65W, peaking slightly below 100W for hmmer and h264ref. The opposite attention-grabbing level is the 8W idle energy, which is indicated as for less than when C1 is enabled. With all idle states enabled, Qualcomm claims below 4W for the complete SoC. Qualcomm was eager to level out that this consists of the IO on the SoC, which requires a separate chipset on an Intel platform.
Any time an Arm chip comes into the enterprise house, ideas instantly flip to high-performance, and Qualcomm is eager right here to level out that whereas performant, their predominant objective is to cloud companies and hyper-scale, resembling scale-out conditions, micro-services, containers, and instance-based implementations. On the launch in San Diego, they rolled out quotes from Alibaba, Google, HPE, and Microsoft, all of whom are working carefully with Qualcomm for deployment. Demonstrations on the launch occasion included NoSQL, cloud automation, information analytics with Apache Spark, deep studying, community virtualization, video and picture processing, compute-based bioinformatics, OpenStack, and neural networks.
On the software program facet, Qualcomm is working with quite a lot of companions to allow and optimize their software program stacks for the Falkor design. At Scorching Chips, Qualcomm additionally acknowledged that there are plans within the works to assist Home windows Server, primarily based on work carried out with their Snapdragon on Arm initiative, though this appeared to be lacking from the presentation.
Additionally as a teaser, Qualcomm gave the title of its next-generation enterprise processor. The following design might be referred to as the Qualcomm Firetail, utilizing Saphira cores. (Qualcomm has already trademarked each of these names).
Qualcomm Centriq is now transport (for income) to key prospects. We must be on the checklist for assessment samples after they develop into out there.
- Analyzing Falkor’s Microarchitecture: A Deep Dive into Qualcomm’s Centriq 2400 for Home windows Server and Linux