Intel ARC/Xe GPU

[The Exiled]

Exclusive: Is Intel’s Arctic Sound already dead?

Intel had already announced the Xe (a.k.a. Gen12) graphics architecture in 2020, and step by step it was also concretized. The portfolio here ranges from integrated ultrabook chips and dedicated cards to genuine workstation and server cards. The one I want to show you today is factually currently at the top end of Intel’s graphics food chain. Intel uses Gen12P5 (Xe HP) ) for Arctic Sound instead of Xe LP as with the DG1. The upcoming DG2 cards, on the other hand, are more likely to resemble Ponte Vecchio (Gen12P71, Xe HPC). With Ponte Vecchio there is namely another Xe-HP version, which offers even more performance than Arctic Sound, is already manufactured in 7nm, consists of eight tiles and is supposed to be suitable for supercomputers.

The really interesting thing about the whole story about Arctic Sound is that they had written about freely scalable variants in advance and this was communicated accordingly. The portfolio, which was already rumored in 2020, included several dedicated cards with a power consumption of 75 watts (without additional, external power supply) up to models with 500 watts, which should then even be supplied by 48 volts instead of 12 volts. This flexible gradation was to be achieved by expansion stages with one, two or four tiles, i.e. coupled individual graphics processors analogous to the chiplets on AMD Ryzen CPUs. But what Intel has now concretized is limited to a total of only two models, if one can believe the internal documents, and also deviates somewhat from the structure presented at the beginning.

But it really does seem by now that Intel has already internally given up on Arctic Sound, at least in its broad variability. Neither card exploits the full potential and comes across as a compromise that has become necessary, but not as a euphorically brandished spearhead that testifies to a new self-confidence. This is a rather modest concession and the obvious realization that they probably wanted more at the beginning than they could then manage in reality. Because the announced 3T or 4T cards are completely missing from the current list. Whether all this is suitable to secure Raja Koduri’s position in the Intel universe in the long run and to establish an unbreakable friendship with Pat Gelsinger, can at least be marked with a big question mark. Xe started with great praise and it will probably be very difficult, at least in some areas, to fully meet their own expectations and not crash land.

Izvor: Igor's Lab

[The Exiled]

[The Exiled]

Intel Ponte Vecchio GPU to be liquid cooled inside OAM form factor

Today, we are getting a new piece of information coming from Igor's LAB, regarding the Ponte Vecchio and some of its design choices. For starters, the GPU will be a heterogeneous design that consists out of many different nodes. Some parts of the GPU will be manufactured on Intel's 10 nm SuperFin and 7 nm technologies, while others will use TSMC's 7 nm and 5 nm nodes. The smaller and more efficient nodes will probably be used for computing elements. Everything will be held together by Intel's EMIB and Foveros 3D packaging. Next up, we have information that this massive Intel processor will be accountable for around 600 Watts of heat output, which is a lot to cool. That is why in the leaked renders, we see that Intel envisioned these processors to be liquid-cooled, which would make the cooling much easier and much more efficient compared to air cooling of such a high heat output. Another interesting thing is that the Ponte Vecchio is designed to fit inside OAM (OCP Accelerator Module) form factor, an alternative to the regular PCIe-based accelerators in data centers. OAM is used primarily by hyper scalers like Facebook, Amazon, Google, etc., so we imagine that Intel already knows its customers before the product even hits the market.

Izvor: Igor's LAB i TechPowerUp

[Odinarius]

Taj tile sustav je fascinantan, ako stvarno bude radio i držao temperature i latencije pod kontrolom, bit će to revolucija. Neobično da su to išli plasirati u okviru ovako kompleksnog proizvoda, razumijem da im je to primarno područje (serveri), ali i izazovi su bitno veći (više dijelova i 600W!).

[The Exiled]

Tiles je Intelov naziv za chiplet dizajn, a kaj se samog pakiranja (1 - 2) tiče, svakako moraju isporučiti Ponte Vecchio GPU, jer su ga uz Sapphire Rapids CPU obećali za Aurora superračunalo.

Aurora was first announced in 2015 and to be finished in 2018. It was expected to have a speed of 180 petaFLOPS which would be around the speed of Summit. Aurora was meant to be the most powerful supercomputer at the time of its launch and to be built by Cray with Intel processors. Later, in 2017, Intel announced that Aurora would be delayed to 2021 but scaled up to 1 exaFLOP. In October 2020, DOE said that Aurora would be delayed again for a further 6 months and would no longer be the first exascale computer in the US. With Aurora planned for mid-2021 and featuring GPUs that were pushed to the end of next year or into 2022, a delay on the supercomputer seemed likely. But Intel kept the possibility open that it might manage to hit its deadline by relying on rival chip foundries. We now know that's not going to happen.

[Odinarius]

Samo što je tiles u ovom izdanju daleko kompliciraniji od chipleta ala AMD, s daleko većim brojem komponenti plus uključili su i 3D stacking (Foveros). Tehnološki kompletno suluda stvar

[The Exiled]

Da, jer je Intel itekako svjestan da sad već opasno zaostaju, posebice za nVidijom koja je glavni HPC/AI/ML izbor, a to je ono kaj Intel cilja s Ponte Vecchio GPU-om. Zato su i išli za toliko kompleksnom arhitekturom koja onda naposljetku spaja nekoliko različitih proizvodnih procesa, nekolicinu pločica (tiles/chiplet) koji pak pojedinačno imaju zasebne funkcije i namjene (Rambo cache, HBM2, Xe GPU, I/O link, EMIB, itd.). Bog-bogova, ako to 2022. sprovedu u djelo kak spada.

[picajzla0707]

Autor The Exiled

Da, jer je Intel itekako svjestan da sad već opasno zaostaju, posebice za nVidijom koja je glavni HPC/AI/ML izbor, a to je ono kaj Intel cilja s Ponte Vecchio GPU-om. Zato su i išli za toliko kompleksnom arhitekturom koja onda naposljetku spaja nekoliko različitih proizvodnih procesa, nekolicinu pločica (tiles/chiplet) koji pak pojedinačno imaju zasebne funkcije i namjene (Rambo cache, HBM2, Xe GPU, I/O link, EMIB, itd.). Bog-bogova, ako to 2022. sprovedu u djelo kak spada.

bog bogova za koga?
koliko od toga ima potencijala za nas smrtnike?

[The Exiled]

Možda trenutno nema za nas direktno, ali ako im ovo uspije, onda imaju jako dobru podlogu za dalje, pa mogu iste ili slične tehnologije u godinama koje dolaze složiti, tj. iskoristiti za desktop i mobilne modele.

Isto kak je AMD odradil sa svojim chiplet dizajnom, Infinity fabric + cacheom i ostatkom popratne skalamerije, koji su im danas sastavni dijelovi CPU i GPU ponude.

Intelu je ovo jako bitno, jer u slučaju da ne zajebu sve po spisku i opet odgode, onda imaju dokaz da stvarno mogu odraditi kaj pričaju i obećavaju.

[Odinarius]

AMD je to definitivno bolje odigrao, išao korak po korak i stekao nužno iskustvo s minimalnim rizikom, a Intel kao i sa 10nm želi sve odjednom, pri čemu je potencijal za greške i odgađanja daleko veći. Kao TSMC koji node smanjuje postepeno. Intel nema izbora nego ići na kamikaze

[The Exiled]

[Disco]

Doslovno iGPU u dGPU paketu...

[The Exiled]

[Disco]

Faul. brisat post.