AMD isporučuje liderski portfolio rješenja za AI za data centre sa AMD Instinct MI300 serijom

AMD je najavio dostupnost AMD Instinct™ MI300X akceleratora – s vodećim memorijskim opsegom za generativni AI¹ i vodećim performansama za obuku i zaključivanje velikog jezičkog modela (LLM) – kao i AMD Instinct™ MI300A ubrzanu procesorsku jedinicu (APU) – koja kombinuje najnovija AMD CDNA™ 3 arhitektura i “Zen 4” CPU-i za isporuku revolucionarnih performansi za HPC i AI radna opterećenja.

“AMD Instinct MI300 serija akceleratora dizajnirani su s našim najnaprednijim tehnologijama, isporučujući liderske performanse i bit će u širokom cloud-u i implementaciji u preduzećima,” rekao je Victor Peng, predsjednik, AMD. „Upotrebom našeg vodećeg hardvera, softvera i pristupa otvorenog ekosistema, provajderi u cloud-u, OEM i ODM-ovi donose na tržište tehnologije koje osnažuju preduzeća da usvoje i implementiraju rješenja zasnovana na umjetnoj inteligenciji.”

MD Instinct MI300X

AMD Instinct MI300X akceleratori su pokretani novom AMD CDNA 3 arhitekturom. U poređenju sa prethodnom generacijom AMD Instinct MI250X akceleratora, MI300X isporučuje skoro 40% više računarskih jedinica², ,5x više memorijskog kapaciteta, 1,7x veći teoretski memorijski propusni opseg³ kao i podršku za nove matematičke formate kao što su FP8; sve je usmjereno na AI i HPC radna opterećenja.

Današnji LLM-ovi nastavljaju da se povećavaju u veličini i složenosti, zahtijevajući ogromne količine memorije i računanja. AMD Instinct MI300X akceleratori imaju najbolji u klasi 192 GB HBM3 memorijskog kapaciteta, kao i 5,3 TB/s vršni memorijski propusni opseg ² kako bi pružili performanse potrebne za sve zahtjevnija AI radna opterećenja. AMD Instinct Platforma je vodeća generativna AI platforma izgrađena na industrijskom standardnom OCP dizajnu sa osam MI300X akceleratora kako bi ponudila vodeći u industriji 1,5 TB HBM3 memorijskog kapaciteta. Standardni dizajn AMD Instinct Platforme omogućava OEM partnerima da dizajniraju MI300X akceleratore u postojeće AI ponude i pojednostave implementaciju i ubrzaju usvajanje servera zasnovanih na AMD Instinct akceleratorima.

U poređenju sa Nvidia H100 HGX, AMD Instinct Platforma može ponuditi povećanje propusnosti do 1,6x kada izvodi zaključivanje na LLM-ovima kao što je BLOOM 176B4 ⁴ i jedina je opcija na tržištu koja može izvesti zaključak za model parametara od 70B, kao što je Llama2, na jednom MI300X akceleratoru; pojednostavljivanje implementacije LLM-a na nivou preduzeća i omogućavanje izvanrednog TCO-a.

AMD Instinct MI300A

AMD Instinct MI300A APU-i, prvi APU za podatkovne centre na svijetu za HPC i AI, koriste 3D pakovanje i 4. generaciju AMD Infinity arhitekture za pružanje liderskih performansi na kritičnim radnim opterećenjima koja se nalaze na konvergenciji HPC-a i AI. MI300A APU kombinuju AMD CDNA 3 GPU jezgra visokih performansi, najnovija AMD "Zen 4" x86 bazirana CPU jezgra i 128 GB nove generacije HBM3 memorije, da isporuče ~1,9x performanse po vatu na FP32 HPC i AI radnim opterećenjima , u poređenju sa AMD Instinct MI250X prethodne generacije⁵.

Energetska efikasnost je od najveće važnosti za HPC i AI zajednice, međutim ova radna opterećenja su izuzetno intenzivna za podatke i resurse. AMD Instinct MI300A APU-i imaju koristi od integracije CPU i GPU jezgara u jednom paketu koji isporučuje visoko efikasnu platformu, a istovremeno pruža i performanse računara za ubrzavanje obuke najnovijih AI modela. AMD postavlja tempo inovacija u energetskoj efikasnosti sa ciljem kompanije 30×25, sa ciljem da isporuči 30x poboljšanje energetske efikasnosti u serverskim procesorima i akceleratorima za AI-obuku i HPC od 2020-2025⁶.

Prednost APU-a znači da AMD Instinct MI300A APU-ovi imaju objedinjene memorijske i keš resurse dajući korisnicima GPU platformu koja se lako može programirati, visokoučinkovito računanje, brzu AI obuku i impresivnu energetsku efikasnost za napajanje najzahtjevnijih HPC i AI radnih opterećenja.

Specifikacije proizvoda

AMD Instinct™	Architecture	GPU CUs	CPU Cores	Memory	Memory Bandwidth (Peak theoretical)	Process Node	3D Packaging w/ 4th Gen AMD Infinity Architecture
MI300A	AMD CDNA™ 3	228	24 “Zen 4”	128GB HBM3	5.3 TB/s	5nm / 6nm	Yes
MI300X	AMD CDNA™ 3	304	N/A	192GB HBM3	5.3 TB/s	5nm / 6nm	Yes

---

 

^{1 MI300-05A: Calculations conducted by AMD Performance Labs as of November 17, 2023, for the AMD Instinct™ MI300X OAM accelerator 750W (192 GB HBM3) designed with AMD CDNA™ 3 5nm FinFet process technology resulted in 192 GB HBM3 memory capacity and 5.325 TFLOPS peak theoretical memory bandwidth performance. MI300X memory bus interface is 8,192 and memory data rate is 5.2 Gbps for total peak memory bandwidth of 5.325 TB/s (8,192 bits memory bus interface * 5.2 Gbps memory data rate/8).The highest published results on the NVidia Hopper H200 (141GB) SXM GPU accelerator resulted in 141GB HBM3e memory capacity and 4.8 TB/s GPU memory bandwidth performance.The highest published results on the NVidia Hopper H100 (80GB) SXM5 GPU accelerator resulted in 80GB HBM3 memory capacity and 3.35 TB/s GPU memory bandwidth performance.}

^{2 MI300-15: The AMD Instinct™ MI300X (750W) accelerator has 304 compute units (CUs), 19,456 stream cores, and 1,216 Matrix cores.
The AMD Instinct™ MI250 (560W) accelerators have 208 compute units (CUs), 13,312 stream cores, and 832 Matrix cores.
The AMD Instinct™ MI250X (500W/560W) accelerators have 220 compute units (CUs), 14,080 stream cores, and 880 Matrix cores.}

^{3 MI300-13: Calculations conducted by AMD Performance Labs as of November 7, 2023, for the AMD Instinct™ MI300X OAM accelerator 750W (192 GB HBM3) designed with AMD CDNA™ 3 5nm FinFet process technology resulted in 192 GB HBM3 memory capacity and 5.325 TFLOPS peak theoretical memory bandwidth performance. MI300X memory bus interface is 8,192 (1024 bits x 8 die) and memory data rate is 5.2 Gbps for total peak memory bandwidth of 5.325 TB/s (8,192 bits memory bus interface * 5.2 Gbps memory data rate/8).The AMD Instinct™ MI250 (500W) / MI250X (560W) OAM accelerators (128 GB HBM2e) designed with AMD CDNA™ 2 6nm FinFet process technology resulted in 128 GB HBM3 memory capacity and 3.277 TFLOPS peak theoretical memory bandwidth performance. MI250/MI250X memory bus interface is 8,192 (4,096 bits times 2 die) and memory data rate is 3.20 Gbps for total memory bandwidth of 3.277 TB/s ((3.20 Gbps*(4,096 bits*2))/8).}

^{4 MI300-34: Token generation throughput using DeepSpeed Inference with the Bloom-176b model with an input sequence length of 1948 tokens, and output sequence length of 100 tokens, and a batch size tuned to yield the highest throughput on each system comparison based on AMD internal testing using custom docker container for each system as of 11/17/2023.
Configurations:
2P Intel Xeon Platinum 8480C CPU powered server with 8x AMD Instinct™ MI300X 192GB 750W GPUs, pre-release build of ROCm™ 6.0, Ubuntu 22.04.2.
Vs.
An Nvidia DGX H100 with 2x Intel Xeon Platinum 8480CL Processors, 8x Nvidia H100 80GB 700W GPUs, CUDA 12.0, Ubuntu 22.04.3.
8 GPUs on each system were used in this test.
Server manufacturers may vary configurations, yielding different results. Performance may vary based on use of latest drivers and optimizations.}

^{5 MI300-23: Calculations conducted by AMD Performance Labs as of Nov 16, 2023, for the AMD Instinct™ MI300X (192GB HBM3 OAM Module) 750W accelerator designed with AMD CDNA™ 3 5nm | 6nm FinFET process technology at 2,100 MHz peak boost engine clock resulted in 163.43 TFLOPS peak theoretical single precision (FP32) floating-point performance.
The AMD Instinct™ MI300A (128GB HBM3 APU) 760W accelerator designed with AMD CDNA™ 3 5nm | 6nm FinFET process technology at 2,100 MHz peak boost engine clock resulted in 122.573 TFLOPS peak theoretical single precision (FP32) floating-point performance.
The AMD Instinct™ MI250X (128GB HBM2e OAM module) 560W accelerator designed with AMD CDNA™ 2 6nm FinFET process technology at 1,700 MHz peak boost engine clock resulted in 47.9 TFLOPS peak theoretical single precision (FP32) floating-point performance.}

^{6 Includes AMD high-performance CPU and GPU accelerators used for AI training and high-performance computing in a 4-Accelerator, CPU-hosted configuration. Goal calculations are based on performance scores as measured by standard performance metrics (HPC: Linpack DGEMM kernel FLOPS with 4k matrix size. AI training: lower precision training-focused floating-point math GEMM kernels such as FP16 or BF16 FLOPS operating on 4k matrices) divided by the rated power consumption of a representative accelerated compute node, including the CPU host + memory and 4 GPU accelerators.}