Intel Problems Roadmap Replace, Targets for ‘Scheduled Predictability’

Intel held an investor webinar the day before today, with the chip massive operating to mission consistency and self belief amid slipping roadmaps and marketplace percentage. On the match, Intel essentially eager about the place it stands with 4 (!) generations of Xeon chips, however the corporate additionally shared a few updates on Gaudi, the deep finding out processor advanced by way of Intel-owned Habana Labs.

Sapphire Rapids (4th-generation)

The 4th-generation (and latest) Xeon CPU, in accordance with the Intel 7 procedure and geared up with Intel’s performance-focused “P-Cores,” introduced previous this 12 months. Sandra Rivera, govt vice chairman and normal supervisor of the Information Middle and AI Staff at Intel, mentioned that Intel is on target to send 1,000,000 devices of Sapphire Rapids by way of mid-year and celebrated the range of Sapphire Rapids-based designs (200+) which can be these days delivery – according to Intel, essentially the most ever on the time of embargo carry.

No new phrase at the “Max” collection CPUs (i.e., Sapphire Rapids with HBM), however we’re anticipating them to seem within the exascale Aurora machine at Argonne Nationwide Lab, in addition to in Los Alamos Nationwide Lab’s Crossroads machine and a brand new supercomputer hosted by way of Argentina’s Nationwide Meteorological Carrier. Spoiler alert: Intel didn’t discuss any coming near near Max collection CPUs within the webinar, and didn’t discuss HBM plans for coming near near generations when it introduced the Max collection rebrand, both. It stays unclear how dedicated the corporate is to offering HBM-equipped Xeon CPUs previous Sapphire Rapids.

Emerald Rapids (Fifth-generation)

“Later this 12 months, we predict to ship our Fifth-gen Xeon Scalable Processor, Emerald Rapids,” Rivera mentioned, maintaining up an Emerald Rapids chip as noticed within the header. A This autumn time-frame was once indicated for Emerald Rapids, which, like Sapphire Rapids, is a P-Core CPU in accordance with the Intel 7 procedure. Rivera mentioned that quantity validation is underway for Emerald Rapids and that it’s these days being sampled to shoppers. Emerald Rapids options an build up in cores (as much as 64, in comparison to Sapphire Rapids’ 60; core density was once a significant emphasis for Intel all the way through the webinar) and, in step with Rivera, options upper performance-per-watt in the similar energy envelope as its predecessor. Rivera mentioned that the supply of Emerald Rapids this 12 months displays Intel “getting again to scheduled predictability.”

Granite Rapids and Sierra Woodland (future-generation)

“Following Fifth-gen Xeon might be Granite Rapids and Sierra Woodland,” Rivera mentioned. “Those two processors might be delivered on our next-generation high-performing platform, which stocks the similar base structure and provides shoppers portability between the 2 merchandise.” Each Granite Rapids (P-Core) and Sierra Woodland (efficiency-focused, hyperscaler-targeted E-Core) might be in accordance with the Intel 3 procedure, Intel’s second-generation EUV node. Intel 4, the primary to leverage EUV, might be used to make PC chips like Meteor Lake, because of arrive later this 12 months.

Granite Rapids (which options even additional greater core density and depend, together with additional progressed performance-per-watt) is slated to be the primary P-Core chip in accordance with Intel 3. Lisa Spelman, company vice chairman and normal supervisor of Intel Xeon, mentioned that Intel is “construction the quickest reminiscence interface on the earth for Granite Rapids,” touting a brand new form of DIMM known as “multiplexer mixed rank” that allowed for an 80% height bandwidth build up over current-gen festival. Rivera mentioned that the primary devices of Granite Rapids are “wholesome” and that it’s being sampled to shoppers nowadays.

Sierra Woodland, in the meantime, would be the debut of the E-Core way, with Rivera calling it the “lead car for the Intel 3 procedure.” The chip will characteristic 144 E-Cores according to socket and, in step with Intel, could have “best-in-class” performance-per-watt. Intel demoed the chip booting on Linux all the way through the webinar, appearing the 144 cores tackling a procedure. “Our first buyer for Sierra Woodland has already won silicon,” Rivera mentioned, “and we predict to be sampling to further shoppers within the coming months.”

Intel expects to send Sierra Woodland to shoppers within the first part of 2024, adopted “in a while after” by way of Granite Rapids. This can be a remarkably swift roadmap for Intel: if all is going in step with plan, Intel could have begun delivery 3 other generations of Xeon processors in a span of lower than a 12 months and a part (January 2023’s Sapphire Rapids release to H1 2024’s Sierra Woodland release).

Clearwater Woodland (future-generation)

If that isn’t sufficient, Intel additionally unveiled the title and trajectory for one in every of its subsequent-generation Xeon processors: Clearwater Woodland. In line with the P-Core “Rapids” names and the E-Core “Woodland” names, Clearwater Woodland is an E-Core processor. Intel says that it is going to be manufactured at the Intel 18A procedure: the “fruits of [Intel’s] 5 nodes in 4 years technique,” Rivera mentioned, and the method by which Intel plans to “reach procedure management.” Rivera added that Intel is “firmly on tempo” to ship Clearwater Woodland in 2025.

Gaudi

Intel’s AI-focused Gaudi chips were quietly gaining steam for a number of years. Amid the snowballing pastime in generative AI packages like ChatGPT, Intel took the chance at its webinar to tout Gaudi’s position in its product lineup. “For the ones very huge language fashions, [Gaudi] is in point of fact the most powerful product that we need to compete and win in that marketplace,” Rivera mentioned, contrasting Gaudi with extra HPC-positioned merchandise like GPUs.

Intel mentioned that Gaudi2 – the current-generation Gaudi chip – demonstrated 2× deep finding out inference and coaching functionality in comparison to in style GPUs. Additional, Rivera mentioned that Intel had labored with Hugging Face to allow the 176 billion-parameter BLOOMZ type with Gaudi2 Accelerators, with Hugging Face discovering a three× speedup relative to aggressive GPUs.

Intel additionally shared {that a} cluster in accordance with 512 Gaudi accelerators – which additionally integrated 256 Xeon CPUs – delivered 1TB/s bisectional bandwidth and had a scaling performance of 97% (“Significantly upper than the trade moderate for any datacenter cluster,” according to Rivera).

Gaudi3 is on target for subsequent 12 months, with Intel’s roadmap graphic lining Gaudi3 up kind of with the release of Sierra Woodland.

Max Collection GPUs (née Ponte Vecchio)

Intel’s datacenter- and HPC-oriented “Max collection” GPUs lately hit a vital bump within the roadmap when the corporate scrapped Rialto Bridge, the successor to the much-delayed current-generation Max collection GPU, and repositioned its plans for the hybrid CPU-GPU Falcon Shores “XPU” as an iterative GPU-only unencumber. Intel – which termed this within the webinar as a “streamlining” of the GPU roadmap – nonetheless took the chance to spotlight Ponte Vecchio, with Rivera announcing that “Max collection GPUs are turning in as much as 50% higher functionality for physics packages as opposed to aggressive merchandise.”

Apparently, when Intel revised its GPU roadmap, we had some questions in regards to the MareNostrum 5 machine on the Barcelona Supercomputing Middle, which was once prior to now slated to characteristic Rialto Bridge GPUs in a fairly small partition. On this presentation, Rivera seemed to point out that Intel has shipped Max collection GPUs to BSC – a conceivable clue that Ponte Vecchio is taking where of Rialto Bridge in MareNostrum 5.