Processador Curiosidades de hardware

[Dark Kaeser]

Radiador de 360 para watercooling? Isso é para n00bs

Argonne lays the groundwork for its next-generation supercomputer​

Red and blue cables curl in and out of the supercomputer as part of a special water-cooling system that pumps 44,000 gallons of water from beneath the floor. (Image by Argonne National Laboratory.)

The new mechanical room’s pipes range from 4 inches to 30 inches in diameter, forming a complex loop that connects cooling towers, chillers, heat exchangers, a filtration system and other components to cool Aurora. (Image by Argonne National Laboratory.)

e a alimentação?

The electrical room is reinforced to support the weight of an airplane. It has 14 substations and is outfitted with a large ceiling hatch so that equipment can be lowered in by cranes. (Image by Argonne National Laboratory.)

https://www.anl.gov/article/argonne-lays-the-groundwork-for-its-nextgeneration-supercomputer

 

[Rui_Marques]

Há de ter um consumo de energia brutal

 

[Dark Kaeser]

Aquele é o "Aurora" (Intel Sapphire Rapids + Intel Ponte Vecchio) e o consumo é praticamente o dobro do outro Supercomputador capaz de >1ExaFlop o "Frontier" (AMD Trento + AMD Instinct MI250X)

aliás, apesar de isto serem sistemas com cerca de ~9000 CPU + ~38000GPU (o Frontier) e ~20000 CPU e 60000 GPU (o Aurora), o custo do hardware acaba por ser suplantado pelos custos energéticos, precisamente a necessidade de "alimentar" e "arrefecer" estes monstros.

Apesar das diferenças de hardware, o fornecedor de ambos os sistema é a HPE (via aquisição da Cray), neste caso designado "Shasta", o que explica as semelhanças, na imagem o Frontier

 

[JPgod]

60 megawatts já poderia ter a sua própria central nuclear dedicada se os mini reatores fossem uma realidade, damm...

Bonito seria desenvolver um projeto de district heating para reaproveitar este calor todo... Era construir estas coisas "perto" de zonas habitacionais e tal e tinha "free heat" pro inverno...

 

[Nemesis11]

IBM Quantum "Osprey" com 433 Qubits, que a IBM irá usar no IBM Quantum System Two.

Roadmap:

"Cablagem" estranha aquela. Parece mais relojoaria/joalharia, que um computador.

 

[Nemesis11]

Apenas uma micro bateria de 100 mAh recarregavel.

https://www.businesswire.com/news/h...ng-Solution-Based-on-a-100µA.h-Micro-Battery!

 

[Scan_Ferr]

Apenas uma micro bateria de 100 mAh recarregavel.

https://www.businesswire.com/news/home/20221110005719/en/At-ELECTRONICA-ITEN-Present-an-Autonomous-Asset-Monitoring-Solution-Based-on-a-100µA.h-Micro-Battery!

Correcção, são 100uA e não mA

 

[JPgod]

Isso pode ser usado nas board no lugar daquela bateria por exemplo...

 

[Dark Kaeser]

Cerebras Reveals Andromeda, a 13.5 Million Core AI Supercomputer​

Cerebras, the company that builds the world's largest chip, the Wafer Scale Engine 2 (WSE-2), unveiled its Andromeda supercomputer today. Andromeda combines 16 of the wafer-sized WSE-2 chips into one cluster with 13.5 million AI-optimized cores that the company says delivers up to 1 Exaflop of AI computing horsepower, or 120 Petaflops of 16-bit half-precision.

The chips are housed in sixteen CS-2 systems. Each chip delivers up to 12.1 TB/s of internal bandwidth (96.8 Terabits) to the AI cores, but the data is fed to the CS-2 processors via 100 GbE networking spread across 124 server nodes in 16 racks. In total, those servers are powered by 284 third-gen EPYC Milan processors wielding 64 cores apiece, totaling 18,176 cores.

The entire system consumes 500 KW, which is a drastically lower amount of power than somewhat-comparable GPU-accelerated supercomputers.

However, Cerebras says that its implementation scales nearly linearly with GPT-class large language models, like GPT-3, GPT-J, and GPT-NeoX. Andromeda can also process 2.5-billion and 25 billion-parameter models that standard GPU clusters simply can't handle due to memory limitations.

Workload scaling is sub-par on most large systems, leading to a diminishing point of returns, often due to code, memory, fabric and/or networking limitations. However, Cerebras has shown that its CS-2 systems scale nearly linearly via data parallelism with no changes to the underlying code — the company's Andromeda supercomputer began crunching through workloads within ten minutes of being fully connected.

https://www.tomshardware.com/news/cerebras-reveals-andromeda-a-135-million-core-ai-supercomputer

 

[Dark Kaeser]

https://twitter.com/i/web/status/1592491987969937408

Radeon Pro WX5100 (Polaris Pro)

 

[Nemesis11]

Já foi conseguido isso há uns anos atrás, no RPi4.

https://domipheus.com/blog/raspberry-pi-4-pci-express-it-actually-works-usb-sata-gpu/

E com o Compute Module, tornou-se mais simples e já vi todo o tipo de placas Pci-Ex ligadas a ele.
Controladoras de discos:

Placas de rede 10 Gbit:

M.2 NVMe:

Etc...

Há até um site onde é mantido uma listagem com placas Pci-Ex que funcionam ou não com o RPi4/CM4.

https://pipci.jeffgeerling.com

 

[Nemesis11]

Algo bastante fora do comum. O Facebook mudou a forma como "mede a passagem do tempo" e é interessante a complexidade que existe a nível de hardware, para se ter maior precisão.

Antena:

Distribuição do sinal com fibra:

Time Card:

Ligação daquela placa a uma placa de rede:

Another example where PTP can help is to reduce lags in gaming, a notorious pain point for anyone that has played games online. Lag happens because systems are out of sync. And as cloud-based gaming becomes more commonplace, and the games themselves become more graphically intensive, PTP’s ability to mitigate lag will make it an important piece of the future of gaming.

https://engineering.fb.com/2022/11/21/production-engineering/precision-time-protocol-at-meta/
https://engineering.fb.com/2021/08/11/open-source/time-appliance/
https://engineering.fb.com/2022/11/21/production-engineering/future-computing-ptp/

 

[Dark Kaeser]

https://twitter.com/i/web/status/1593237270118961155

 

[Nemesis11]

Fotos da Cerebras ficam aqui sempre bem.

https://www.servethehome.com/a-cerebras-cs-2-engine-block-bare-on-the-sc22-show-floor/

 

[Nemesis11]

Teardown de um cabo óptico 100 Gbit:

Um microcontrolador com 1 Core ARM M3 @ 72 Mhz + 10 KB SRAM + 32 KB Flash.

Mais fotos com mais detalhes e explicações - https://social.afront.org/@kwf/109492743645650432

 

[Dark Kaeser]

Se alguém quiser matar saudades

Olimex AgonLight2 8-bit open-source hardware computer sells for 50 Euros​

Olimex has started to take pre-orders for the AgonLight2, a customized version of the Agon Light 8-bit open-source hardware computer based on the Z80 family and running BBC Basic.

Olimex Agonlight2 (left) vs original Agon light board (right)

• MCU – Zilog eZ80F92 8-bit microcontroller @ 18.423 MHz with 128 KB flash, 256 bytes configuration Flash memory, 8 KB SRAM (See PDF datasheet for details)
• System Memory – 512KB, 10ns, parallel SRAM
• Storage – MicroSD card socket
• Terminal subsystem
  • MCU – ESP32-PICO-D4 dual-core SiP @ up to 240 MHz with 4MB flash
  • Memory – 8MB, 133 MHz, serial PSRAM
  • Video Output – VGA port up to 640×480 resolution, 64 colors
  • Audio – 2x mono audio jacks
  • USB 2.0 Type-C port for power and programming ESP32’s firmware
  • USB Type-A port for keyboard
• Programming – ZDI port for programming the firmware of the CPU
• Expansions
  • Plastic boxed control port with 20 GPIOs
  • UEXT expansion board for Olimex add-on modules
• Power Supply
  • 5V via USB Type-C port or control port; DCDC converter (up to 2A)
  • LiPo battery support with charging circuit, 2-pin connector, and Vbat on control port
• Dimensions – TBD

The main changes made by Olimex are highlighted in bold in the specs above, but I’ll list them again for more clarity:

• Replaced the USB-A power connector with a USB-C connector as it is used in all new phones, tablets and devices as mandated by a new EU directive
• Replaced the PS2 connector with a USB-A connector for connecting USB keyboards without an adapter
• The Linear voltage regulator was replaced with a DCDC converted which delivers up to 2A current
• Added support for LiPo batteries
• Replaced the bare header 32-pin connector with a plastic boxed 34-pin connector, adding two additional signals Vbat and Vin which allow AgonLight to be powered by this connector too.
• Added a UEXT connector for Olimex modules
• A few other small changes and most of the components were changed to Olimex’s component base to lower costs

https://www.cnx-software.com/2022/12/15/olimex-agonlight2-8-bit-open-source-hardware-computer/

O SoC para o I/O é mais poderoso que o chip do sistema em si

 

[Dark Kaeser]

Divulgados dados sobre um daqueles unicórnios... PEZY-SC3 (acelerador many core)

https://twitter.com/i/web/status/1603645651987755008

https://www.pezy.co.jp/wp-content/uploads/2022/12/PEZY-SC3_TD.pdf

 

[JPgod]

Só de caches isso tem 232 MB onboard... damm!

 

[Dark Kaeser]

O tal "lego PC", ele publicou um vídeo

 

[Nemesis11]

Acho que nunca vi uma Docking Station com este formato. É um candeeiro de escritório, com a luz, Hub USB, HDMI, Wireless Charger e Webcam 4K.

• 1 x USB Type-C input (up to 135W power intput)
• 1 x USB 3.1 Gen 2 Type-C (65W passthrough power output for charging a laptop)
• 1 x USB 3.1 Type-C (20W power output)
• 2 x USB 3.1 Gen 1 Type-A (5 Gbps)
• 1 x HDMI 2.0 (4k/60Hz with HDR support)

https://liliputing.com/lenovo-go-desk-station-is-a-lamp-wireless-charger-usb-dock-and-4k-camera/

329$.
Mais interessante seria se, em vez de Docking Station, fosse mesmo um computador, onde a Board ficaria na base. Parece-me perfeitamente fazível.

Não relacionado com isto, mas também fora do comum, uma empresa aqui ao lado, de Barcelona, anunciou um RPU (Randomness Processing Unit).

https://quside.com/quside-unveils-the-worlds-first-randomness-processing-unit/