Processador Intel CPU 2018-2021 Roadmap Leaks Out

[Dark Kaeser]

A despropósito:

From HPE Cray Shasta Slide.
2020 : AMD Milan + Nvidia Volta Next : Perlmutter.
2021 : AMD Trento + AMD MI200 : Frontier.
2022 : Intel Sapphire Rapids + Intel Ponte Vecchio : Aurora.
2023 : AMD Genoa + AMD Next-Gen : El Capitan.

https://twitter.com/KOMACHI_ENSAKA/status/1340526846530621442

 

[Nemesis11]

Isso a ser verdade quando lançariam os 7nm, em 2024?

O Charlie fala em fins de 2023 para o Granite Rapids.
Mas isto é em relação a produção em volume e dificilmente o Granite Rapids será o primeiro produto a 7nm. É um shrink/evolução do Sapphire Rapids e esse parece que irá usar 2 dies enormes.

Isto se tudo se mantiver como está. É partir do principio que não irá haver mais atrasos, ou possíveis decisões de port para outro processo de fabrico da Intel ou TSMC.
Ainda falta muito tempo para se ter a certeza que é aquilo que irá acontecer.

 

[Nemesis11]

A Intel lá conseguiu lançar os novos Atoms a 10 nm, para o mercado consumidor, 1 ano e 3 meses depois do anuncio desta arquitectura nova (Tremont).

https://www.anandtech.com/show/16388/intel-launches-jasper-lake-tremont-atom-cores-for-all

É possível que sejam interessantes para alguns mercados. NAS, Mini PCs, Media Centers, etc. No Apollo e Gemini Lake a Intel já eram mais interessantes que outros Atoms do Passado.
Como neste Atom, cada módulo é constituído no mínimo por 4 Cores, mesmo as versões Dual Core vão ter acesso a 1.5 MB de L2 e 4 MB de L3. Apenas os outros 2 Cores estarão desactivados.

https://www.anandtech.com/show/16388/intel-launches-jasper-lake-tremont-atom-cores-for-all

O Slide de comparação da Intel é que é um bocado............"manhoso".
Pegam no melhor SKU de 6W de TDP e comparam-no com uma AMD Stoney Ridge de 6W, que saiu a meio de 2016 e um Mediatek de 5W, que saiu no inicio de 2018 e que usa um Core ARM A73 que saiu em 2016.
A Intel podia mostrar um pouco mais de confiança.

 

[MGD]

Já os benchmarks de jogos foi tudo jogo que sempre beneficiou a Intel, e mesmo assim com vantagens de 4% apenas, e sem especificar Cooling, PL, etc.

 

[Dark Kaeser]

Intel Reports Q4 2020 Earnings: 2020 Delivers A Profitable Pandemic

Looking at Intel’s individual business groups, for the year both the client and data center groups did very well for themselves. Client revenue was up 8% to $40.1B, coming in on the back of higher laptop sales. And that growth really is all from laptops; desktop revenue was down for the year, and even “adjacencies” (Wi-Fi adapters and the like) were down a bit versus 2019.

Meanwhile data center revenue was up 11% to $26.1B, with Intel coincidentally reporting that they shipped 11% more data center chips than in 2019. On the whole, data center growth was driven by cloud and communication service providers, both of whom ramped up their hardware purchases to meet service demands during the pandemic, while enterprise and government sales dipped on the year. Meanwhile it’s interesting to note that on both a quarterly and yearly basis, Intel’s ASPs for the data center group are down; despite the volume, they face an increasing amount of competition.

https://www.anandtech.com/show/1644...gs-2020-delivers-a-rather-profitable-pandemic

Q: Stacy: Intel said for a year that 7nm was great, and then in July it wasn't and had a delay. Now you say after a brief look at the data, it's now on track? What are you seeing that gives you that confidence, and what does that mean for other nodes?

Pat Gelsinger: Bob and the team have been working for six months on this. I'm looking at thoroughly analyzed data. I looked at that data and came to the same conclusion as Bob and his team. We're also pausing to look deeper into the roadmap for more clarity in the future. We also have adjustments in leadership too, also making adjustments in product leadership team over time. One of my favorite engineers Glenn Hinton announced he was coming back to the company today. The key leaders are coming back to Intel. Leadership, roadmap, a few more weeks of analysis will allow us to put Intel on a path.

https://www.anandtech.com/show/16441/intel-reports-q4-2020-earnings-analyst-qa-transcript

 

[DiogoAntunes]

https://www.techpowerup.com/278110/intel-xeon-sapphire-rapids-lga4677-x-processor-sample-pictured

É para um mercado especifico , mas é monstro.

 

[Nemesis11]

De facto, processadores para o mercado de Anime, são mesmo muito específicos.

Agora mais a sério, o mais provável é que por baixo, estejam 2 dies enormes. O próprio Ice Lake-SP, que irá sair antes, deverá ser constituído por 2 dies, que não devem ser pequenas, especialmente quando comparadas com os chiplets da AMD.
Adicionando a isso, é quase garantido que esse Sapphire Rapids, terá SKUs com HBM integrado no package.

Já agora, ontem vi uma comparação entre ES do Intel Ice Lake-SP e o AMD Milan e a Intel vai continuar com sérios problemas no mercado servidor:
https://browser.geekbench.com/v5/cpu/compare/6288580?baseline=6198746

A comparação é entre 2 Processadores Ice Lake-SP com 36 Cores, contra 1 AMD Milan de 64 Cores.
De referir que se forem ver resultados da versão final do AMD Milan, os resultados são bem melhores que este ES. No caso da Intel não se sabe se haverá melhoria de performance para a versão final.
Aquele AMD, com aqueles clocks, deve andar entre os 225 a 280 W de TDP. Os slides do Ice Lake-SP falavam em 270 W de TDP (Aquele sistema tem 2).

É bom que esse Sapphire Rapids não sofra atrasos e que venha com melhores resultados de performance e TDP.

 

[Dark Kaeser]

Há mais imagens...

FPGA?

https://t.bilibili.com/488021238176910289?tab=2

 

[Nemesis11]

Há mais imagens...

Ena......

Atirando assim para o ar, 4 Dies "coladas" via EMIB?

Gostava de saber a área de cada die e a área total. Assim a "olho", a área total parece enorme.

Interessante, mas não vejo que se possam tirar grande conclusões da presença desse FPGA, sem algum leak que explique a sua presença. Pode ser para muita coisa. Se calhar até pode ser só para dev desses ES.

 

[Dark Kaeser]

Não seja por isso.

Alguém iria fazê-lo

using CPU dimensions of 56.6mm x 77.6mm
Each chiplet die size ≈ 418.8mm²

https://twitter.com/_rogame/status/1357630681111740418

Portanto 4x418.8mm² = +1600mm²

 

[JPgod]

São 4 dies ao nivel do que se encontra nas gráficas high end... va lá vai

 

[Nemesis11]

Portanto 4x418.8mm² = +1600mm²

São mesmo 4 enormes "bolachas". Apenas por termo de comparação, o actual Epyc Zen2 tem uma área total de 1008 mm2. Se o Epyc Zen 3 usar a mesma IOD, terá uma área total de 1086 mm2.

Parece que cada die nesse Saphiire Rapids terá 14 Cores (poderá ter mais, mas esses são os activos) e esse sample tem 7 Cores activos por die.
Parece que estará mesmo a usar EMIB.

Since we're allegedly looking at a 28-core CPU, each die should only have seven active cores. Unless Intel decided to pack enormously large caches and loads of special-purpose accelerators into its Sapphire Rapids chiplets, we can gather that each chiplet integrates a significantly higher number of cores (14, according to YuuKi_AnS), but at this point many of them are disabled. In any case, the image shows that Intel’s Xeon Scalable ‘Sapphire Rapids’ CPU uses four dies presumably interconnected using EMIB technology. Since it is inherently easier to make smaller chiplets than one large die, it should be easier for Intel to ramp up the production of its next-gen CPU for data centers and supercomputers

A FPGA será uma Altera Max 10.

The Sapphire Rapids sample apparently carries four dies and an Altera Max 10 FPGA (which is probably used for startup/initialization purposes).

https://www.tomshardware.com/news/sapphire-rapids-pictured-lga4677

São 4 dies ao nivel do que se encontra nas gráficas high end... va lá vai

Com uma die daquele tamanho, é mesmo muito provável que terá 1 ou mais Cores inactivos, mesmo nas versões de topo.


Por ultimo, alguém se lembra deste Slide da Intel, sobre o primeiro Epyc?

 

[Dark Kaeser]

Hmmm....

Intel “Raptor Lake” desktop and mobile series appear on a leaked roadmap

Intel Raptor Lake Desktop

We have already been told that Raptor Lake is basically a refresh of Alder Lake, which makes a lot of sense. It was previously rumored that “Meteor Lake” might launch immediately after Alder Lake, but the former is already expected at 7nm SuperFin architecture, which seems too soon, considering Intel’s problems with this node. Intel Raptor Lake would, on the other hand, still use 10nm Enhanced SuperFin.

https://videocardz.com/newz/intel-raptor-lake-desktop-and-mobile-series-appear-on-a-leaked-roadmap

Assim sendo parece que a Intel decidiu introduzir o Raptor entre o Alder e Meteor, que supostamente será adiado? Cancelado?

Agora a pergunta que todo o mundo certamente está a pensar:

Alguém na Intel achou mesmo boa ideia introduzir o Raptor e a seguir lançar o Comet?

Já foi há algum tempo, mas, como é que isso acabou da última vez?

 

[Nemesis11]

Desculpem, mas alguém tinha que pegar neste meme. Era obrigatório.

Vamos ver se o Raptor Lake não será a 10++nm e o Meteor a 10+++nm..........and son on.

 

[Torak]

Com um pouco de azar, ainda fazem o backport para 14nm++++

 

[Dark Kaeser]

Eu escrevi Comet em vez de Meteor
Estou a ficar afectado com o Whiskey Lake

 

[Dark Kaeser]

EagleなStreamのSPRはTDP 300Wかー。Whitley ICXがTDP 270Wだと考えれば、それほど悪くないのかな。

https://twitter.com/KOMACHI_ENSAKA/status/1370881695730126849

A plataforma Eagle Stream, que vem substituir a Whitley do Ice Lake Xeon, e que vai servir os Sapphire Rapids Xeon, vai ter aparentemente TDP até 300w.
Acho que se falava dos 10nm SuperFIN ou Enhanced Superfin, e até 56 cores acima dos 38/40 do Ice Lake, ainda assim

 

[Dark Kaeser]

Micron Quits 3D XPoint, Puts Factory Up for Sale

Micron announced today (Tuesday) that it is ending its involvement in 3D XPoint, the non-volatile memory technology it developed in collaboration with Intel nearly a decade ago. As it ceases all R&D on 3D XPoint, Micron is seeking a buyer for the Lehi, Utah, fab where the memory is manufactured.

Micron expects a higher ROI for the new investments it will be making. The company reported it was losing about $400 million a year on 3D XPoint.

The Lehi fab was at one time jointly owned by Micron and Intel, but Micron bought Intel out of their share in 2019. Market analyst Dan Olds of Intersect360 Research told HPCwire that while Micron supplied Intel with wafers to use in their Optane products, that hasn’t produced enough end user sales, assumedly, to make the product profitable.

“Reading between the tea leaves of the official announcement, the market just hasn’t embraced 3D XPoint to the degree that it makes sense for Micron to continue production and the partnership with Intel,” Olds said.

He added that while Micron is shopping its factory to several interested parties, “Intel [is] probably at the head of the list.”

https://www.hpcwire.com/2021/03/16/micron-quits-3d-xpoint-puts-factory-up-for-sale/

 

[Nemesis11]

Li muito na diagonal:

We're Keeping the Fabs: Please Stop Asking​

Since the start of the year, previous Intel CEO Bob Swan and new Intel CEO Pat Gelsinger have both done the equivalent of shouting from the rooftops when it comes to Intel’s market advantages. At the top of that list is Intel’s vertical integration between manufacturing and chip design, enabling Intel to control the process from top to bottom more closely than any of its chip competitors. This unique proposition labels Intel as an IDM (integrated device manufacturer) or having an IDM model (integrated device manufacturing model), and the only company that can even come close to Intel in this regard is Samsung. Intel’s plus over its competitor here is that it can apply its own designs to a scale that Intel is often attributed to.

In recent months there have been comments from analysts and investors about the potential for Intel to spin out its manufacturing plants and facilities into a separate business, similar to how AMD detached its manufacturing facilities into a new company called GlobalFoundries. The benefits of this move would allow Intel to segregate the losses between the two sides of the business, and showcase the core Intel product teams in a better light by comparison to the manufacturing arm. This sort of path has significant pitfalls, most importantly that Intel’s manufacturing main customer is Intel. GlobalFoundries had this problem initially, but Intel’s foundry arm is on a much larger scale.

So to that end, Pat Gelsinger today is hoping to put those rumors to rest, more than he stated on Intel’s 2020 financial call. Not only is Intel going to retain its manufacturing facilities, but it is going to embrace a new era of manufacturing, which is going to be called IDM 2.0.

Build (7nm)​

Today Intel will be announcing a $20 billion (USD) investment in two new manufacturing facilities (fabs) in Arizona, set to come online for production in 2024. Gelsinger will point out today that Intel is ready to break ground, increasing the number of factories on Intel’s Ocotillo campus (Chandler, AZ) from four to six.

These new fabs will be on leading edge process node technologies, with Intel extending cooperation with the State of Arizona, as well as the current administration’s target of improving semiconductor manufacturing inside the country. All parties involved seem raring to go, and Intel will be seeking to outfit its production facilities with the equipment necessary to enable leading edge manufacturing, including using Extreme Ultra Violet (EUV) technology.

Expand (TSMC)​

Alongside building new manufacturing facilities, Intel today will reaffirm its roadmap to use a mixture of internal and external process node manufacturing depending on the product capabilities. Intel already makes heavy use of external partners, such as TSMC, and already spends more than $7 billion USD at TSMC annually. But today’s announcement will double down on ensuring that Intel is ready to use the right process at the right time for the right products.

Productize (IFS)​

The two new manufacturing plants / fabs also filter under this Expand section, however this area is more focused on how Intel will keep those fabs fully occupied. Other semiconductor manufacturers in the market, such as TSMC, Samsung, GlobalFoundries, SMIC etc, they all have what are called foundry services which allow customers to build silicon using their manufacturing technology. With today’s announcement, the company is ready to enable its external Intel Foundry Services (IFS) to new customers. IFS will be a standalone company, with unique access to Intel's current and future offerings.

Collaboration with IBM, New Intel Events for Engineers​

Alongside the core IDM 2.0 path, Intel also has announcements relating to the future of its R&D roadmap, as well as its outreach to engineers and commercial partners.

The collaboration with IBM on process node development and development of next-generation logic is the wildcard of today’s announcements. The two companies are set to work together on foundational technologies to move the needle on both semiconductor performance and semiconductor efficiency. The collaboration will scale through to the ecosystem, with a significant nod towards key US government initiatives.

https://www.anandtech.com/print/165...ndry-services-ibm-collaboration-return-of-idf

 

[Dark Kaeser]

Hmmm

Build 7nm?

Espero que não seja Fab 42 v3.0

Spoiler

Intel will invest $7 billion to finish a factory it started in 2011​

Construction was paused in 2014, and Fab 42 has sat empty since.​

Andrew Cunningham - 2/8/2017, 10:35 PM

https://arstechnica.com/information...llion-to-finish-a-factory-it-started-in-2011/

E como era mais ou menos expectável com a introdução do Raptor Lake...

Intel’s 2023 CPU product roadmap: “It includes Meteor Lake for client and ‘Granite Rapids’ for data center. Both Meteor Lake and Granite Rapids will have compute tiles built on Intel 7nm. […] For our 2023 roadmap, we will also leverage our relationship with TSMC to deliver additional leadership CPU products for our client and data center customers. This is the power of our new IDM 2.0 model combined with a modular approach to design and Intel’s industry leading packaging technologies.”

— Pat Gelsinger

EDIT: repare-se no importante pormenor da imagem

TILE

as fatias chegaram aos CPU client da Intel.