ServerNews 2023 Review: AI Everywhere, AI All Around
Intel Corporation
NVIDIA
Amazon Web Services
Microsoft
Google/DeepMind
SiPearl
Fujitsu
Loongson
The IT industry fully recovered from the pandemic and raced to catch up. ServerNews highlights the most interesting announcements, deals, and events of 2023 and early 2024, covering CPUs, accelerators, data centers, memory, networks, HPC, and the Russian market. The year saw intense competition between Intel and AMD, the rise of Arm and RISC-V, and a surge in AI accelerators amid a shortage.
In 2023, Intel released two Xeon Scalable generations (Sapphire Rapids and Emerald Rapids) and Xeon Max with HBM2e memory, while AMD countered with EPYC Genoa-X and Bergamo, and plans for Turin. Arm processors gained ground with AWS Graviton4, Microsoft Cobalt 100, and AmpereOne, while RISC-V also advanced. The AI accelerator market was dominated by NVIDIA's H100, with a shortage that benefited niche cloud providers like CoreWeave and Lambda Labs. Microsoft introduced its own Maia 100 accelerator, and Google launched TPU v5p and v5e. AMD's Instinct MI300 series competed with H100, and software optimization became increasingly crucial. The article also covers memory, storage, networking, HPC, and the Russian market.
- Abbreviations
- CPU = Central Processing Unit — центральный процессор
- APU = Accelerated Processing Unit — ускоренный процессорный блок
- HPC = High-Performance Computing — высокопроизводительные вычисления
- TOPS = Tera Operations Per Second — тера операций в секунду
- L3 cache = Level 3 cache — кэш 3 уровня
- ISA = Instruction Set Architecture — архитектура системы команд
- UCIe = Universal Chiplet Interconnect Express — универсальный интерконнект для чиплетов
- RISC-V = Reduced Instruction Set Computer - V — RISC-V (открытая архитектура набора команд)
- BMC = Baseboard Management Controller — контроллер управления материнской платой
- RoT = Root of Trust — корень доверия
- TPU = Tensor Processing Unit — тензорный процессор
- LLM = Large Language Model — большая языковая модель
- XPU = Accelerated Processing Unit (Intel's term) — ускоренный процессорный блок
- CUDA = Compute Unified Device Architecture — вычислительная архитектура NVIDIA
- oneAPI = One Application Programming Interface — единый программный интерфейс
- HBM = High Bandwidth Memory — высокопропускная память
- HBM2e = High Bandwidth Memory 2e — высокопропускная память 2e
- HBM3 = High Bandwidth Memory 3 — высокопропускная память 3
- DDR5 = Double Data Rate 5 — двухскоростная память типа 5
- PCle = Peripheral Component Interconnect Express — взаимосвязь периферийных компонентов
- NVLInk = NVIDIA Link — связь NVIDIA
- LPDDR5X = Low Power Double Data Rate 5X — низкопотребляющая память типа 5X
- POWER = Performance Optimization With Enhanced RISC — архитектура POWER
- UEFI = Unified Extensible Firmware Interface — унифицированный интерфейс прошивки
- DARPA = Defense Advanced Research Projects Agency — Агентство перспективных оборонных исследовательских проектов
- SDB = Software Defined Base — программно-определяемая база
- HPC = High-Performance Computing — высокопроизводительные вычисления
Source: ServerNews —
original