Intel has the most CPU cores in a processor
At the start of the in-house event Intel Innovation 2023, Intel boss Pat Gelsinger pulled two surprises out of the hat. The “Sierra Forest” server processor, which has been announced for 2024 for some time now, will have up to 288 processor cores instead of the previously expected maximum of 144.
And following Nvidia’s example, Intel wants to build an AI supercomputer with its own technology that customers can rent. For this computer with Xeon processors and 4,000 Gaudi2 computing accelerators, Intel named the company Stability AI as a pilot customer, which, among other things, develops the open source AI image generator Stable Diffusion.
Intel im Plan
Pat Gelsinger is under enormous pressure to lead Intel out of the crisis and wants to keep big promises on time. With the strategy “five nodes in four years” – i.e. the introduction of five generations of new manufacturing technology (Intel 7, Intel 4, Intel 3, Intel 20A, Intel 18A) within four calendar years – Intel wants to return to TSMC as the technically leading chip manufacturer. (Contract) manufacturers move past and win external customers for the extremely expensive new chip fabs of the foundry division.
“Five nodes in four years”: The manufacturing processes Intel 7, Intel 4, Intel 3, Intel 20A and Intel 18A appear in rapid succession.
Gelsinger sees his company “on track” in this regard and proves this with specifically planned new Intel processors.
It is particularly important for Intel to get its Xeon brand server processors back on track. Headwinds are blowing from three directions: the unfavorable situation in the global economy including increasing sanctions against the growth market of China, the growing competitive pressure from AMD (Epyc) and ARM (AWS Graviton, Ampere Altra) and the reallocation of the budgets of important major customers from cloud servers to AI machines .
Gelsinger can’t change anything about the global economy, but he consciously emphasizes the cooperation with the Chinese customer Alibaba Cloud and recently warned the US government against excessive delivery restrictions.
Intel assembles the Xeon “Sierra Forest” from several chiplets; When it comes to PC processors, Intel prefers to call them “tiles”.
Intel wants to score points against AMD with two new Xeon generations: On December 14th – just within the promised time frame of 2023 – the Xeon-SP of the fifth generation Emerald Rapids is scheduled to launch, which compared to the 4 (Sapphire Rapids) probably only brings minor improvements and also comes from “Intel 7”. Granite Rapids from the “Intel 3” manufacturing technology will follow in the second half of 2024.
Intel’s “Sierra Forest” with up to 288 E cores will compete against multi-core cloud server processors with ARM technology and the Epyc 98×4 “Bergamo” recently introduced by AMD with up to 128 Zen 4c cores in the first half of 2024. The Compute Chiplets come from the “Intel 3” manufacturing technology. As early as 2025, Intel will add “Clearwater Forest” with CPU chiplets including Intel 18A technology, emphasized Gelsinger.
This schedule has been known for a long time, but what is new is the higher number of cores at Sierra Forest. Until now, Intel had mentioned a maximum of 144 cores. In the new version, Intel uses two compute chiplets with a total of 288 cores; The storage controllers are likely to be located in two other areas.
Intel’s Sierra Forest processor uses four dies in its full configuration.
The AI server Supermicro SuperServer SYS-820GH-TNR2 with eight Gaudi2 modules.
Intel is proud that the acquired subsidiary Habana achieved quite good results with its Gaudi2 computing accelerator in the AI benchmark GPT-J from the MLPerf Inference v3.1 Datacenter suite, which is managed by the MLCommons consortium – by the way, under the leadership of David Kanter, who website realworldtech.com operates.
Based on these current benchmark results, Gelsinger calls the Gaudi2 “the only viable alternative on the market for AI computing needs.” By this he means the Nvidia H100 “Hopper”, which is currently bringing Nvidia record revenues and which potential buyers are literally fighting over because they hope to be able to train their respective AI models more quickly and launch them onto the market.
In other words: Intel also wants to benefit from the current AI boom and the Gaudi2 is the best of its own products. The Supermicro SuperServer SYS-820GH-TNR2 with eight Gaudi2 modules and two Xeon-SP Gen3 is available with 256 GB of RAM from around 141,000 US dollars, less than half the price of an Nvidia DGX H100 with eight times H100.
Gelsinger initially did not provide any further details about the Gaudi2 supercomputer. Competitor Cerebras recently began building several AI supercomputers with its own Wafer Scale Engine 2 (WSE-2) together with a cloud provider from the United Arab Emirates. Nvidia has been running the Selene with A100 for years (which is also high in the top 500) and is currently equipping the Eos with 4608 H100.
Waiting for Ponte Vecchio and Gaudi3
Intel Data Center GPU Max alias Ponte Vecchio
(Image: c’t/Carsten Spille)
Intel actually has another powerful AI accelerator on offer, namely the Data Center GPU Max 1550 with 128 GB of HBM2E RAM, also known as Ponte Vecchio, which was presented in January. But Supermicro only laconically writes “Coming Soon” to the SYS-821GV-TNR equipped with it. Clearly, Ponte Vecchio is not a “realistic alternative” to AI accelerators currently sold.
And Gelsinger himself announced a few months ago – long after the “launch” of Ponte Vecchio, which also powers the 2-EFlops supercomputer Aurora – that Gaudi3 was virtually on the home stretch for 2024. The tape-out for Gaudi3 production with TSMC N5 took place in the first quarter of 2023.
But according to current knowledge, Intel is also continuing to work on the accelerators with
Update September 19, 2023 6:31 p.m
Image added to Sierra Forest.
To the home page
#Intel #CPU #cores #processor