Gráfica Intel Arc Discrete GPUs

E existe mesmo um decoder card com vários gpus DG1, o @Nemesis11 que sabe qual é, eu pesquisei e não achei :D

EDIT: É a Intel SG1 com 4x DG1 no mesmo PCB. Nem sei se isso chegou a sair pro mercado mesmo que só vejo fotos antigas de um render (?) do PCB pela net.
Depende do que consideres chegar ao mercado. Acho que o produto era só para Big Clouds. Pelas fotos, até existem diversos modelos. A placa é a XG310. Antes dela existiram a VCA e VCA2 onde eram colocados APUs e eram usados os GPUs integrados para encoding.

System Spec
dimension¾ longth, 28mm X 34mm)
Package TypeFCBGA
TDP23W
OpenGL4.1
Internal Thermal SensorYes
L3 cache size16MB
TDP150W
GPU chip4
PCIe adapterGen3 x16
thermalPassive
GPU memory128-bit width,8GB per GPU chip,LPDDR4,68.25GB/s
2133MHz/ 4267MT/s(max)
Operating Frequency4267MT/s
GPU memory capacity8GB per GPU, total 32GB
OSCentOS,Debian OS
SWHost BIOS –Default Production BIOS will need Upgrading
SG1-M IFWI
Telemetry – MCU on card

LOrI0bV.jpg


mcjNioj.jpg


wY6nPqZ.jpg



Anyway.............mais fotos de uma ES:
L2XXp75.jpg


TcDB6rC.jpg


IkGffjN.jpg


https://videocardz.com/newz/intel-arc-alchemist-engineering-sample-pictured-once-again
 
No outro dia vi no Reddit que a única ES funcional 'conhecida' do Larrabee foi vendida no eBay recentemente, tipo peça de hardware de coleção.

Lembrei-me, e muito, desta nova investida da Intel nos GPU que deu origem a este tópico, não sei porquê... :D
 
O PCB respira boa qualidade, aquela fila de condensadores SMD e tal, muito nice.

8 + 6 que acho meio weird, por min devia ter sido tudo corrido a 8 pins mal surgiu isso para padronizar nas fontes antes do novo standart

Nada de backplate, imagino que só as custom vão ter.
 
Samuel-i-knew-it-would-happen.gif



Chinese GUNNIR is now selling discrete desktop graphics card with full Intel DG1 GPU​

The DG1 GPUs were initially only available to software developers but in the end, they have made their way into the discrete DIY desktop market under the Iris Xe brand.

Intel developed two DG1 SKUs: Iris Xe and Iris Xe Max, the latter featuring the full implementation of the silicon exclusively available for laptops. This is no longer true though, as GUNNIR is now selling Iris Xe Max Index V2 graphics card with 96 Execution Units (up from 80 available with Iris Xe non-MAX).
Intel-Xe-MAX-Hero.jpg

The company offers as many as three variants of DG1 based cards, which are now all listed on Taobao with the following pricing:
  • GUNNIR DG1 (low-profile): 569 RMB (90 USD)
  • GUNNIR Iris Xe Index V2: 639 RMB (101 USD)
  • GUNNIR Iris Xe MAX Index V2: 699 RMB (110 USD)
GUNNIR-IRIS-INDEX-V2-768x467.jpg
GUNNIR-IRIS-INDEX-MAX-V2-768x469.jpg


It should be noted that Irix Xe graphics cards will not work with all systems. There is no support for any AMD platform, and Intel support is limited only to the Coffee/Comet Lake series (9th Gen and 10th Gen). The Iris Xe models are compatible with Intel B460, B365, and H310C motherboards.
https://videocardz.com/newz/chinese...desktop-graphics-card-with-full-intel-dg1-gpu
 
GUNNIR é um belo nome......

Anyways, além de só serem compatível com esses Chipsets da Intel, não era preciso uma BIOS diferente, para se conseguir usar esses GPUs?
Para Devs é "ok", mas para o publico geral parece-me muito pouco atractiva.
 
Supostamos sim, mas a originalidade e desenrasque chinês não é igual ao nosso, é maior, bem maior.

Diria que entre a Gunnir e uma série de fabricantes chineses de boards, para o mercado chinês, ainda deve de representar uns milhares de unidades para compensar o investimento.
Internet cafes na China ainda é uma cena actual!

EDIT: para clarificar

A utilização era suposto estar restrita aos OEM, mas nada impede que seja comercializado desta forma, tirando a "imposição da Intel", mas diria que tirando o caso chinês também não compensaria em mais lado nenhum.
 
Última edição:

E aparentemente um update ao sucessor destas, Ponte Vecchio e aparentemente a Intel também vai ter um "Super APU"

Intel Technology Roadmaps and Milestones​


  • Visual Compute Roadmap and Strategy
  • Intel Arc Graphics Timing and Roadmap Update – AXG expects to ship more than 4 million discrete GPUs in 2022. OEMs are introducing notebooks with Intel Arc graphics, code-named Alchemist, for sale in the first quarter of 2022. Intel will ship add-in cards for desktops in the second quarter and workstations by the third quarter. Architecture work has begun on Celestial, a product that will address the ultra-enthusiast segment.

  • Super Compute Roadmap and Strategy
  • Ponte Vecchio – AXG is on track to deliver Ponte Vecchio GPUs for the Aurora supercomputer program later this year. Ponte Vecchio achieved leadership performance results with up to 2.6x more performance compared with the leading market solution on a complex financial services workload.
  • Arctic Sound-M – Arctic Sound-M brings the industry’s first hardware-based AV1 encoder into a GPU to provide 30% bandwidth improvement and includes the industry’s only open-sourced media solution. The media and analytics supercomputer enables leadership transcode quality, streaming density and cloud gaming. Arctic-Sound M is sampling to customers and will ship by mid-2022.
  • Falcon Shores – Falcon Shores is a new architecture that will bring x86 and Xe GPU together into a single socket. This architecture is targeted for 2024 and projected to deliver benefits of more than 5x performance-per-watt, 5x compute density, 5x memory capacity and bandwidth improvements.1
https://www.intel.com/content/www/us/en/newsroom/news/intel-technology-roadmaps-milestones.html?


Está explicado os 300M€ que a Intel Federal teve de pagar à instituição que ia receber o Supercomputador Aurora, o Argonne National Lab (era para ter sido entregue originalmente em 2018 ainda com o Phi/Kngihts, depois reformulado a entrega para 2020 já com o Ponte Vecchio), apesar desta instituição ter recebido o Borealis no final do ano passado
HPCwire: What is Borealis? Give us the basics.

Wisniewski:
Borealis is our testbed for evaluating, testing, and debugging the Aurora system and the new technologies going into it. It’s a two-rack mini-system, so it’s small compared to the 100+ racks we’ll have at Argonne – but it will have more compute power than many HPC centers, and we expect it to rank on the TOP500 when it’s completed.
https://www.hpcwire.com/2021/12/16/...-exascale-readiness-for-aurora-supercomputer/
 
Existem alguns rumores que os drivers ainda estão verdinhos, o que não surpreende. Sinceramente espero que ninguém esteja à espera de grande coisa nesta primeira geração, mas quando sairem as Battle Mage a coisa ja deve estar mais refinada e pronta para efetivamente serem concorrência.
 

Intel Arc Update: Alchemist Laptops Q1, Desktops Q2; 4mil GPUs Total for 2022​



Celestial Architecture Under Development: Targeting Ultra Enthusiast Market

Arc_Roadmap_575px.png


Celestial is now under development, and at this point, Intel is expecting it to be their first product to address the ultra-enthusiast market (i.e. the performance crown). GPUs based on the Celestial architecture are expected in the “2024+” timeframe; which is to say that at this far out Intel doesn’t seem to know for sure if they’ll be 2024 or 2025 products.
Covering the gap between Alchemist and Celestial will be Battlemage, the second of Intel’s Arc GPU architectures. Battlemage now has a 2023-2024 release date, with Intel expecting the architecture to improve performance over Alchemist to the point where Battlemage will be competitive in the enthusiast GPU market – but not quite reaching the ultra-enthusiasts that Celestial will.
https://www.anandtech.com/show/1726...ptops-q1-desktops-q2-4mil-gpus-total-for-2022



Intel’s Arctic Sound-M Server Accelerator To Land Mid-2022 With Hardware AV1 Encoding​


The announcement of Arctic Sound-M follows a hectic, and ultimately sidetracked set of plans for Intel’s original GPU server hardware. The company initially commissioned their Xe-HP series of GPUs to anchor the traditional server market, but Xe-HP was canceled in November of last year. Intel didn’t give up on the server market, but outside of the unique Ponte Vecchio design for the HPC market, they did back away from using quite so much dedicated server silicon.
Intel_Arctic_Sound_M_678x452.jpg

In the place of those original products, which were codenamed Arctic Sound, Intel is instead coming to market with Arctic Sound-M. Given the investor-focused nature of today’s presentation, Intel is not publishing much in the way of technical details for their forthcoming server accelerator, but we can infer from their teaser videos that this is an Alchemist (Xe-HPC part), as we can clearly see the larger Alchemist die mounted on a single-slot card in Intel’s teaser video. This is consistent with the Xe-HP cancellation announcement, as at the time, Intel’s GPU frontman, Raja Koduri indicated that we’d see server products based on Xe-HPG instead.
https://www.anandtech.com/show/1726...or-to-land-mid2022-with-hardware-av1-encoding

Também estava a achar estranho a designação Artic Sound, pois esta já tinha sido cancelada, mas por alguma estranha razão resolveram usar a mesma designação para um produto diferente.
 

ISSCC 2022: How Intel brings 63 tiles into one package for Ponte Vecchio​


The base tile has an area of 646 mm² and provides the infrastructure for PvC. This also includes the memory controller, the Fully Integrated Voltage Regulators (FIVR), the associated power management and the 16-lane PCIe 5.0 or CXL host interface. The base tile consists of 17 layers and is manufactured in Intel 7.

The 47 active tiles have a common chip area of 2,330 mm². Including the thermal tiles, it is 3,100 mm². The entire package measures 77.5 x 62.5 mm (4,844 mm²). As an OAM module, Ponte Vecchio is connected to the rest of the system via 4,468 pins
isscc-2022-intel-pvc-1_1920px.png

We already reported on the structure of the compute tiles in summer 2021. Some information about the RAMBO cache is new. This SRAM consists of four banks of 3.75 MB each and thus comes to 15 MB per tile. With four RAMBO tiles on two base tiles, we get 120 MB of additional cache. This is connected to the fabric at 1.3 TB/s per chip. A short reference to the connection of the compute tiles: This takes place via the fabric at 2.6 TB/s each.
isscc-2022-intel-pvc-2_1920px.png

ct the two base tiles and the HBM2E chips. The compute and RAMBO tiles are powered by the base tile using Power TSVs in various array configurations with thick metal layers for distribution.

Another component of the base tile are 144 MB of L3 cache and a memory fabric (MF) with a complex geometric topology and a bandwidth of 4,096 bytes per clock cycle. The memory banks are located directly below the compute tiles and represent an extension of the L1 and L2 caches. Bidirectionally, they each have 64 bytes per clock.

The Foveros Die Interconnect (FDI) is used to connect the compute and RAMBO tiles to the base tile. Each FDI link is divided into eight groups. Each group in turn consists of 800 lanes for the compute tiles. The energy efficiency of the FDI is 0.2 pJ/B. A single link thus achieves 2.8 GT/s.
isscc-2022-intel-pvc-4_1920px.png

https://www-hardwareluxx-de.transla...sl=de&_x_tr_tl=en&_x_tr_hl=pt-PT&_x_tr_pto=sc

@Nemesis11

hot-shots2-charlie-sheen.gif


artigo traduzido pelo Google do original em alemão.
 
Back
Topo