New E-Core Only Xeons in 2024

It’s no secret that Intel’s enterprise processor platform has been stretched in latest generations. In comparison with the competitors, Intel is chasing its multi-die technique whereas counting on a producing platform that hasn’t supplied one of the best out there. That being stated, Intel is quoting extra shipments of its newest Xeon merchandise in December than AMD shipped in all of 2021, and the corporate is launching the following era Sapphire Rapids Xeon Scalable platform later in 2022. Past Sapphire Rapids has been considerably underneath the hood, with minor leaks right here and there, however at the moment Intel is lifting the lid on that roadmap.

State of Play At present

At the moment out there is Intel’s Ice Lake 3^rd Technology Xeon Scalable platform, constructed on Intel’s 10nm course of node with as much as 40 Sunny Cove cores. The die is giant, round 660 mm², and in our benchmarks we noticed a sizeable generational uplift in efficiency in comparison with the two^nd Technology Xeon providing. The response to Ice Lake Xeon has been combined, given the competitors out there, however Intel has cast forward by leveraging a extra full platform coupled with FPGAs, reminiscence, storage, networking, and its distinctive accelerator choices. Datacenter revenues, relying on the quarter you take a look at, are both up or down based mostly on how clients are digesting their present processor inventories (as said by CEO Pat Gelsinger).

That being stated, Intel has put a considerable amount of effort into discussing its 4^th Technology Xeon Scalable platform, Sapphire Rapids. For instance, we already know that will probably be utilizing >1600 mm² of silicon for the best core rely options, with 4 tiles linked with Intel’s embedded bridge know-how. The chip may have eight 64-bit reminiscence channels of DDR5, help for PCIe 5.0, in addition to a lot of the CXL 1.1 specification. New matrix extensions additionally come into play, together with knowledge streaming accelerators, fast help know-how, all constructed on the newest P-core designs at the moment current within the Alder Lake desktop platform, albeit optimized for datacenter use (which usually means AVX512 help and larger caches). We already know that variations of Sapphire Rapids will probably be accessible with HBM reminiscence, and the primary buyer for these chips would be the Aurora supercomputer at Argonne Nationwide Labs, coupled with the brand new Ponte Vecchio high-performance compute accelerator.

The launch of Sapphire Rapids is considerably later than initially envisioned a number of years in the past, however we count on to see the {hardware} extensively accessible throughout 2022, constructed on Intel 7 course of node know-how.

Subsequent Technology Xeon Scalable

Wanting past Sapphire Rapids, Intel is lastly placing supplies into the general public to showcase what’s arising on the roadmap. After Sapphire Rapids, we may have a platform appropriate Emerald Rapids Xeon Scalable product, additionally constructed on Intel 7, in 2023. Given the naming conventions, Emerald Rapids is prone to be the 5^th Technology.

Emerald Rapids (EMR), as with another platform updates, is predicted to seize the low hanging fruit from the Sapphire Rapids design to enhance efficiency, in addition to updates from the manufacturing. With platform compatibility, it means Emerald may have the identical help relating to PCIe lanes, CPU-to-CPU connectivity, DRAM, CXL, and different IO options. We’re prone to see up to date accelerators too. Precisely what the silicon will seem like nonetheless continues to be an unknown. As we’re nonetheless new in Intel’s tiled product portfolio, there’s an excellent probability will probably be much like Sapphire Rapids, but it surely might equally be one thing new, corresponding to what Intel has deliberate for the era after.

After Emerald Rapids is the place Intel’s roadmap takes on a brand new freeway. We’re going to see a diversification in Intel’s technique on various ranges.

Beginning on the prime is Granite Rapids (GNR), constructed totally of Intel’s efficiency cores, on an Intel 3 course of node for launch in 2024. Beforehand Granite Rapids had been on roadmaps as an Intel 4 node product, nonetheless, Intel has said to us that the development of the know-how in addition to the timeline of the place it would come into play makes it higher to place Granite on that Intel 3 node. Intel 3 is supposed to be Intel’s second-generation EUV node after Intel 4, and we count on the design guidelines to be very comparable between the 2, so it’s not that a lot of a bounce from one to the opposite we suspect.

Granite Rapids will probably be a tiled structure, simply as earlier than, however it would additionally characteristic a bifurcated technique in its tiles: it would have separate IO tiles and separate core tiles, fairly than a unified design like Sapphire Rapids. Intel hasn’t disclosed how they are going to be linked, however the thought right here is that the IO tile(s) can comprise all of the reminiscence channels, PCIe lanes, and different performance whereas the core tiles will be targeted purely on efficiency. Sure, it feels like what Intel’s competitors is doing at the moment, however in the end it’s the proper factor to do.

Granite Rapids will share a platform with Intel’s new product line, which begins with Sierra Forest (SRF) which can also be on Intel 3. This new product line will probably be constructed from datacenter optimized E-cores, which we’re aware of from Intel’s present Alder Lake client portfolio. The E-cores in Sierra Forest will probably be a future era than the Gracemont E-cores we now have at the moment, however the thought right here is to offer a product that focuses extra on core density fairly than outright core efficiency. This permits them to run at decrease voltages and parallelize, assuming the reminiscence bandwidth and interconnect can sustain.

Sierra Forest will probably be utilizing the identical IO die as Granite Rapids. The 2 will share a platform – we assume on this occasion this implies they are going to be socket appropriate – so we count on to see the identical DDR and PCIe configurations for each. If Intel’s numbering scheme continues, GNR and SRF will probably be Xeon Scalable 6^th Technology merchandise. Intel said to us in our briefing that the product portfolio at the moment supplied by Ice Lake Xeon merchandise will probably be lined and prolonged by a mixture of GNR and SRF Xeons based mostly on buyer necessities. Each GNR and SRF are anticipated to have full world availability when launched.

The E-core Sierra Forest targeted on core density will find yourself being in comparison with AMD’s equal, which for Zen4c will probably be referred to as Bergamo – AMD may need a Zen5 equal when SRF involves market.

I requested Intel whether or not the transfer to GNR+SRF on one unified platform means the era after will probably be a novel platform, or whether or not it would retain the two-generation retention that clients like. I used to be informed that it could be excellent to take care of platform compatibility throughout the generations, though as these are deliberate out, it relies on timing and the place new applied sciences must be built-in. The earliest trade estimates (past CPU) for PCIe 6.0 are within the 2026 timeframe, and DDR6 is extra like 2029, so except there are extra reminiscence channels so as to add it’s possible we’re going to see parity between 6^th and seven^th Gen Xeon.

My different query to Intel was about Hybrid CPU designs – if Intel was now going to make P-core tiles and E-core tiles, what’s stopping a mixed product with each? Intel said that their clients choose uni-core designs on this market because the wants from buyer to buyer differ. If one buyer prefers an 80/20 cut up on P-cores to E-cores, there’s one other buyer that prefers a 20/80 cut up. Having a wide selection of merchandise for every completely different ratio doesn’t make sense, and clients already investigating this are discovering out that the software program works higher with a homogeneous association, as an alternative cut up on the system degree, fairly than the socket degree. So we’re not prone to see hybrid Xeons any time quickly. (Ian: Which is an excellent factor.)

I did ask concerning the unified IO die – giving the identical P-core solely and E-core solely Xeons the identical variety of reminiscence channels and I/O lanes may not be optimum for both state of affairs. Intel didn’t actually have an excellent reply right here, other than the truth that constructing them each into the identical platform helped clients synergize non-returnable growth prices throughout each CPUs, whatever the one they used. I didn’t ask on the time, however we might see the door open to extra Xeon-D-like eventualities with completely different IO configurations for smaller deployments, however we’re speaking merchandise which might be 2-3+ years away at this level.

For each Granite Rapids and Sierra Forest, Intel is already working with key ‘definition clients’ for microarchitecture and platform growth, testing, and deployment. Extra particulars to come back, particularly as we transfer via Sapphire and Emerald Rapids throughout this yr and subsequent.