Coming to the top of the primary quarter of 2023, Intel’s Information Heart and AI group is discovering itself at an fascinating inflection level – for causes each good and unhealthy. After repeated delays, Intel is lastly transport their Sapphire Rapids CPUs in excessive volumes this quarter as a part of the 4th Era Xeon Scalable lineup, all of the whereas its successors are developing in a short time. However, the GPU facet of the enterprise has hit a tough spot, with the surprising cancelation of Rialto Bridge – what would have been Intel’s subsequent Information Heart GPU Max product. It hasn’t all been excellent news prior to now few months for Intel’s beleaguered information heart group, nevertheless it’s not all unhealthy information, both.
It’s been simply over a 12 months since Intel final delivered a wholesale replace on its DCAI product roadmaps, which had been final refreshed at their 2022 buyers assembly. So, given the sheer significance of the excessive margin group, in addition to every part that has been happening prior to now 12 months – and will probably be happening over the subsequent 12 months – Intel is holding an investor webinar immediately to replace buyers (and the general public at giant) on the state of its DCAI product lineups. The occasion is being handled as an opportunity to recap what Intel has achieved over current months, in addition to to put out an up to date roadmap for the DCAI group masking the subsequent couple of years.
The high-level message Intel is trying to challenge is that the corporate is lastly turning a nook of their essential information heart enterprise section after some notable stumbles in 2021/2022. Within the CPU house, regardless of the repeated Sapphire Rapids delays, Intel’s successive CPU tasks stay on observe, together with their first all E-core Xeon Scalable processor. In the meantime Intel’s FPGA and devoted AI silicon (Gaudi) are equally coming alongside, with new merchandise hitting the market this 12 months whereas others are taping-in.
Sapphire Rapids: 4th Era Xeon Scalable Transport in Quantity
Following what can solely be described as a protracted growth course of for Intel’s subsequent technology Xeon Scalable processors, Sapphire Rapids lastly started transport in quantity over the previous few months. The Q1’23 (ed: or is that Q5’22?) launch of the product has come later than Intel would have ever appreciated, however the firm is lastly in a position to put the event course of behind them and benefit from the fruits of transport the large chips in excessive volumes.
At this level Intel isn’t quoting exact cargo numbers – again at launch, the corporate stated it anticipated to make it to one million items in file time – however the firm is doubling-down on their claims that they’ll be capable of product the big, complicated chips in excessive sufficient volumes to fulfill buyer demand. Constructed on the Intel 7 course of, the ultimate iteration of what began as Intel’s 10nm line, Intel is benefitting from that well-tuned course of. On the similar time, nevertheless, the 4th Era Xeon Scalable lineup contains Intel’s first chiplet-based Xeon design, so it’s nonetheless not the best launch.
Apart from assembly buyer demand, Intel’s primary level is that each one of their main clients are adopting the long-awaited chips. That is largely unsurprising provided that Intel nonetheless holds the vast majority of the information heart CPU market, however given the investor viewers for immediately’s bulletins, it’s additionally unsurprising to see Intel explicitly calling consideration to this. Apart from a generational enchancment in CPU core structure, Sapphire Rapids additionally delivers every part from DDR5 to PCIe 5/CXL assist, so there is no such thing as a scarcity of curiosity in changing older Ice lake and Cascade Lake (3rd & 2nd Gen Xeon Scalable) {hardware} with one thing newer and extra environment friendly.
Intel, in fact, is trying to fend off arch-rival AMD from taking much more market share on this house with their EPYC processors, which are actually on to their 4th technology (9004 sequence) Geona components. There are a number of demos slated to be run this morning showcasing efficiency comparisons; Intel is eager to point out buyers that they’re transport the superior silicon, particularly as AMD has the benefit when it comes to core counts. So count on Intel to give attention to issues like their AI accelerator blocks, in addition to comparisons that pitch an equal variety of Sapphire Rapids (Golden Cove) and Geona (Zen 4) CPU cores towards one another.
Emerald Rapids: On Monitor for This autumn’23, Will Be 5th Era Xeon Scalable
Diving into the way forward for Intel’s product roadmap, the primary disclosure from immediately’s occasion is an replace on the standing of Emerald Rapids, the architectural successor to Sapphire Rapids. Intel’s earlier roadmap had chips primarily based on the structure slated to reach in 2023, a launch cycle that has been more and more referred to as into query given Sapphire Rapids’ delay to 2023. However certain sufficient, Intel nonetheless expects to ship the subsequent technology of Xeon processors later this 12 months, in This autumn.
In line with Intel, Emerald Rapids chips are already sampling to clients. On the similar time, quantity validation is already underway as nicely. As Emerald Rapids is a comparatively simple successor to Sapphire Rapids, Intel is trying to keep away from the lengthy validation interval that Sapphire Rapids required, which will probably be essential for making up for misplaced time and getting the subsequent Xeon components out by the top of this 12 months.
Provided that that is an investor assembly, Intel isn’t providing a lot in the way in which of technical specs for the next-generation chips. However the firm is confirming that Emerald Rapids will function in the identical energy envelope as Sapphire Rapids – bettering on the platform’s total performance-per-watt effectivity. In actual fact, the truth that Emerald Rapids will use the identical LGA 4677 platform as Sapphire is being handled as a significant promoting level for Intel, who will probably be totally leveraging the drop-in compatibility that may afford. Clients will be capable of swap out Sapphire for Emerald of their present designs, permitting for simple upgrades of already-deployed programs, or within the case of OEMs, shortly bringing Emerald Rapids programs to the market.
Intel has beforehand disclosed that Emerald Rapids will probably be constructed on the Intel 7 course of. Because of this the majority of any efficiency/effectivity positive factors must come from architectural enhancements. That stated, Intel can be touting “elevated core density”, so it seems like Emerald can even supply larger core counts than Sapphire, which topped out at 60.
As a part of the webinar, Intel additionally confirmed off an uncapped Emerald Rapids chip. Primarily based on the sheer quantity of silicon on the package deal and the multi-tile configuration (every tile is definitely over 700mm2), we imagine that is possible the highest-end XCC configuration. Which at two tiles, is a big design change from Sapphire Rapids, which used 4 smaller tiles for its XCC configuration. Which works to point out that although Sapphire and Emerald are socket-compatible and utilizing the identical platform, Intel is not restraining itself from making modifications below the hood (or on this case, below the IHS).
Lastly, following within the footsteps of the product naming scheme they’ve used for the final a number of years now, Intel is formally naming Emerald Rapids because the 5th Era Xeon scalable household. So count on to see the official identify used instead of the code identify for the majority of Intel’s bulletins and disclosures going ahead.
Granite Rapids: Already Sampling, to Ship In 2024 With MCR DIMM Help
Following Emerald Rapids, in 2024 Intel will probably be transport Granite Rapids. This will probably be Intel’s next-generation P-core primarily based product. Like Emerald, Granite has been beforehand disclosed by Intel, so immediately’s announcement is an replace on their progress there.
In line with Intel, Granite Rapids stays on observe for its beforehand introduced 2024 launch. The half is predicted to launch “carefully following” Sierra Forest, Intel’s first E-core Xeon Scalable processor, which is due in H1’24. Regardless of being no less than a 12 months out, Granite Rapids is already to the purpose the place the primary stepping is up and working, and it’s already sampling to some Intel clients.
As famous in earlier disclosures, Granite Rapids is a tile-based structure, with separate compute and I/O tiles – an evolution from Sapphire Rapids, which even in its tiled kind is actually a whole SoC in every tile. Granite Rapids’ compute tiles are being constructed on the Intel 3 course of, Intel’s second-generation EUV node, having been pulled in from Intel 4 in its earliest incarnation. In the meantime we nonetheless don’t have vital official data on the I/O tiles.
Together with upgrades to its CPU structure, Intel can be disclosing for the primary time that Granite Rapids can even include a notable new reminiscence function: MCR DIMM assist. First revealed by SK hynix late final 12 months, Multiplexer Mixed Ranks (MCR) DIMMs basically gang up two units/ranks of reminiscence chips with a view to double the efficient bandwidth to and from the DIMM. With MCR, Intel and SK hynix are aiming to get information charges equal to DDR5-8800 (or larger) speeds, which might be a big boon to reminiscence bandwidth and throughput, as that is typically briefly provide with immediately’s many-core chips.
As a part of immediately’s presentation, Intel is displaying off an early Granite Rapids system utilizing MCR DIMMs to realize 1.5TB/second of reminiscence bandwidth on a twin socket system. Primarily based on Intel’s presentation, we imagine this to be an 8 12 channel reminiscence configuration with every MCR DIMM working on the equal of DDR5-8800 speeds.
As an apart, it’s value noting that because the farthest-out P-core Xeon in Intel’s roadmap, there’s a notable lack of point out of Excessive Bandwidth Reminiscence (HBM) components. HBM on Sapphire Rapids was used as the premise of Intel’s choices for the HPC market, and whereas that wasn’t fairly a one-off product, it’s shut. Future HPC-focused CPUs had been being developed as a part of the Falcon Shores challenge, which was upended with the change to Intel’s GPU schedule. So at the moment, there’s not a brand new HBM-equipped Xeon on Intel’s schedule – or no less than, not one they wish to speak about immediately.
Sierra Forest: The First E-Core Xeon and Intel 3 Lead Product, Transport H1’24
Shifting gears, we’ve got Intel’s forthcoming lineup of E-core Xeons. These are chips that will probably be utilizing density-optimized “effectivity” cores, which had been launched by Intel in late 2021 and have but to make it to a server product.
Sierra Forest is one other earlier Intel disclosure that the corporate is updating buyers on, and is maybe an important of them. Using E cores in a Xeon processor will considerably enhance the variety of CPU cores Intel can supply in a single CPU socket, which the corporate believes will probably be extraordinarily essential for the market going ahead. Not solely will the E core design enhance total compute effectivity per socket (for massively threaded workloads, no less than), however it is going to afford cloud service suppliers the flexibility to consolidate much more digital machine cases on to a single bodily system.
Like Granite Rapids, Sierra Forest is already up and working at Intel. The corporate accomplished the power-on course of earlier within the quarter, getting a full working system up and working inside 18 hours. And although it’s the primary E-core Xeon, it’s already secure sufficient that Intel has it sampling to no less than one buyer.
As beforehand disclosed, regardless of the E-Core/P-Core break up, Sierra Forest and Granite Rapids will probably be sharing a platform. In actual fact, they’re sharing a complete lot extra, as Sierra can even use the identical I/O tiles as Granite. This enables Intel to develop a single set of I/O tiles after which basically swap in E-core or P-core tiles as wanted, making for Sierra Forest or Granite Rapids.
And for the primary time, we’ve got affirmation of what number of E-cores that Sierra will supply. The Xeon will ship with as much as 144 E-cores, over twice as many cores as discovered on immediately’s P-core primarily based Sapphire Rapids processors. There are not any additional architectural disclosures on the E-cores themselves – it was beforehand confirmed that it’s a post-Gracemont structure – so extra particulars are to come back on that entrance. Gracemont positioned its E-cores in quads, which if that holds for the CPU structure utilized in Sierra Forest, would imply we’d be taking a look at 36 E-core clusters throughout your complete chip.
With Sierra Forest up and working, this additionally implies that Intel has wafers to point out off. As a part of her portion of the presentation, Lisa Spelman, Intel’s CVP and GM of the Xeon product lineup, held up a completed Sierra Forest compute tile wafer to underscore Intel’s progress in manufacturing their first E-core Xeon CPU.
Talking of producing, Intel has additionally confirmed that Sierra Forest is now the lead product for the Intel 3 node throughout your complete firm. Which suggests Intel is trying to make a large leap in a really brief time frame with respect to its Xeon product lineup, shifting from Intel 7 on Emerald Rapids in This autumn’23 to their second-generation EUV course of no later than Q2’24. Sierra does get the good thing about merchandise primarily based on Intel 4 (the corporate’s first-generation EUV course of) coming first, however this nonetheless makes Sierra’s progress crucial, as Intel 3 is the primary “full service” EUV course of for Intel, providing assist for Intel’s full vary of cell libraries.
Of the entire Xeon processor architectures outlined immediately, Sierra is arguably an important for Intel. Intel’s opponents within the Arm house have been providing excessive density core designs primarily based on the Neoverse structure household for a number of years now, and arch-rival AMD goes the identical course this 12 months with the deliberate launch of its Zen 4c structure and related EPYC “Bergamo” processors. Intel expects an essential subset of their clients to give attention to maximizing the variety of CPU cores over rising their total socket counts – thus making information heart CPU income extra carefully observe core counts than socket counts – so Intel wants to fulfill these calls for whereas warding off any opponents desirous to do the identical.
Clearwater Forest: Second-Gen E-core Xeon In 2025 on Intel 18A Course of
Lastly, in an all-new disclosure for Intel, we’ve got our first particulars on the half that may succeed Sierra Forest as Intel’s second-generation E-core Xeon processor. Codenamed Clearwater Forest, the follow-up E-core half is scheduled to be delivered in 2025, inserting it not more than 18 months after Sierra Forest.
Much like how Sierra is Intel’s first Intel 3 half, Clearwater Forest is slated to be the primary Xeon produced on Intel’s 18A course of – their second-generation RibbonFET course of, which final 12 months was moved up in Intel’s schedule and will probably be going into manufacturing within the second half of 2024.
At two years out, Intel isn’t disclosing anything in regards to the chip. However its announcement immediately is to substantiate to buyers that Intel is dedicated to the E-core lineup for the long-haul, in addition to to underscore how, on the again of the 18A course of, that is the purpose the place Intel expects to re-attain course of management. In the meantime, Intel has additionally confirmed that there gained’t be any Xeons made on their early 20A course of, so Clearwater Forest will probably be Intel’s first RibbonFET-based Xeon, interval.
Lastly, it’s value noting that with the most recent extension to Intel’s CPU roadmap, P-core and E-core Xeons are remaining distinct product traces. Intel has beforehand commented that their clients both need one core or the opposite on a CPU – however not each on the similar time – and Clearwater Forest maintains this distinction.
Xeon Scalable Generations | ||||||
Date | AnandTech | Codename | Abbr. | Max Cores |
Node | Socket |
Q3 2017 | 1st | Skylake | SKL | 28 | 14nm | LGA 3647 |
Q2 2019 | 2nd | Cascade Lake | CXL | 28 | 14nm | LGA 3647 |
Q2 2020 | third | Cooper Lake | CPL | 28 | 14nm | LGA 4189 |
Q2 2021 | Ice Lake | ICL | 40 | 10nm | LGA 4189 | |
Q5 2022 | 4th | Sapphire Rapids | SPR | 60 P | Intel 7 | LGA 4677 |
This autumn 2023 | fifth | Emerald Rapids | EMR | >60 P | Intel 7 | LGA 4677 |
H1’2024 | sixth? | Sierra Forest | SRF | 144 E | Intel 3 | ? |
2024 | Granite Rapids | GNR | ? P | Intel 3 | ||
2025 | seventh? | Clearwater Forest | CWF | ? E | Intel 18A | ? |
? | Subsequent-Gen P | ? | ? P | ? |
AI Accelerators & FPGAs: Capturing Market Share At All Ends
Whereas the majority of immediately’s presentation from Intel is concentrated on their CPU roadmap, the corporate can be briefly referring to the roadmaps for his or her FPGA and devoted AI accelerator merchandise.
Before everything, Intel is anticipating to qualify (PRQ) 15 new FPGAs throughout the Stratix, eASIC, and Agilex product traces this 12 months. There are not any additional technical particulars on these, however the merchandise, and their successors, are within the works.
In the meantime, for Intel’s devoted AI acceleration ASICs, the corporate’s Habana Labs division has lately tapped-in their next-generation Gaudi3 deep studying accelerator. Gaudi3 is a course of shrink of Gaudi2, which was first launched again within the spring of 2022, shifting from TSMC’s 7nm course of to a 5nm course of. Intel isn’t attaching a supply date to the chip for its investor crowd, however extra particulars will probably be coming later this 12 months.
All informed, Intel is projecting the marketplace for AI accelerators to be no less than a $40 billion market alternative by 2027. And the corporate intends to sort out the market from all sides. Which means CPUs for AI workloads which can be nonetheless finest served by CPUs (basic laptop), GPUs and devoted accelerators for duties which can be finest served by extremely parallel processors (accelerated laptop), after which FPGAs bridging the center as specialist {hardware}.
It’s fascinating to see that, even supposing GPUs and different extremely parallel accelerators ship the most effective efficiency on giant AI fashions, Intel doesn’t see the entire addressable marketplace for AI silicon being dominated by GPUs. Somewhat they count on the 2027 market to be a 60/40 break up in favor of CPUs, which given Intel’s a lot stronger place in CPUs than GPUs, will surely be to their benefit. Definitely, CPUs aren’t going wherever even for AI workloads (if nothing else, one thing wants to organize the information for these GPUs), however will probably be fascinating to see if Intel’s TAM predictions maintain true in 4 years, particularly given the eye-watering costs that GPU distributors have been in a position to cost in recent times.