Nvidia의 파트너들이 간접적으로 Blackwell B200 GPU의 지연을 확인(2024.08.10)

2024.08.10

Nvidia partners indirectly confirm Blackwell B200 GPU delay — offer interested parties liquid-cooled H200 instead | Tom's Hardware (tomshardware.com)

 

Nvidia partners indirectly confirm Blackwell B200 GPU delay — offer interested parties liquid-cooled H200 instead

Nvidia의 파트너들이 간접적으로 Blackwell B200 GPU의 지연을 확인하며, 관심 있는 고객들에게 대신 액체 냉각 방식의 H200을 제공하고 있습니다.

 

Nvidia's possible delay of Blackwell GPUs for AI and HPC applications will not have a dramatic impact on AI server makers and the AI server market, according to Supermicro, a major server supplier. The comment by Charles Liang (via SeekingAlpha), chief executive of Supermicro, can be considered confirmation that Nvidia's next-generation B100 and B200 processors are indeed delayed as they need rework.

 

Supermicro의 주요 서버 공급업체인 Charles Liang CEO에 따르면, Nvidia의 AI 및 HPC 애플리케이션용 Blackwell GPU의 지연은 AI 서버 제조업체와 AI 서버 시장에 큰 영향을 미치지 않을 것이라고 합니다. 이는 Nvidia의 차세대 B100 및 B200 프로세서가 실제로 지연되었으며, 재작업이 필요하다는 것을 확인하는 발언으로 볼 수 있습니다. (출처: SeekingAlpha)

 

"We heard Nvidia may have some delay, and we treat that as a normal possibility," Liang said during the company's earnings call with financial analysts and investors. He stated, "When they introduce a new technology, new product, [there is always a chance] there will be a push out a little bit. In this case, it pushed out a little bit. But to us, I believe we have no problem to provide the customer with a new solution like H200 liquid cooling. We have a lot of customers like that. So, although we hope better deploy in the schedule, that's good for a technology company, but this push out overall impact to us. It should be not too much."

 

"우리는 Nvidia가 일부 지연될 수 있다는 이야기를 들었으며, 이를 정상적인 가능성으로 받아들이고 있습니다."라고 Liang은 회사의 실적 발표 회의에서 금융 분석가 및 투자자들에게 말했습니다. 그는 "Nvidia가 새로운 기술이나 신제품을 도입할 때, 언제나 약간의 지연이 발생할 가능성이 있습니다. 이번 경우에는 약간 밀렸습니다. 그러나 우리는 H200 액체 냉각과 같은 새로운 솔루션을 고객에게 제공하는 데 문제가 없다고 믿습니다. 많은 고객들이 그것을 좋아합니다. 그래서 우리는 일정 내에 더 잘 배포되기를 바라지만, 기술 회사로서 이는 좋은 일입니다. 하지만 이번 지연이 우리에게 미치는 전체적인 영향은 크지 않을 것입니다."라고 덧붙였습니다.

 

Nvidia's B100 and B200 GPUs are the industry's first products to use TSMC's CoWoS-L packaging with a Super Carrier interposer. It enables the building of system-in-packages up to six times the reticle size by using active or passive local silicon interconnect (LSI) bridges integrated into a redistribution layer (RDL) interposer (instead of a silicon interposer in the case of CoWoS-S used for H100).

 

Nvidia의 B100 및 B200 GPU는 TSMC의 CoWoS-L 패키징과 Super Carrier 인터포저를 사용하는 업계 최초의 제품입니다. 이는 능동 또는 수동 로컬 실리콘 인터커넥트(LSI) 브리지를 재분배층(RDL) 인터포저에 통합하여 시스템 인 패키지를 리티클 크기의 최대 6배까지 구성할 수 있게 해줍니다(H100에 사용된 CoWoS-S의 실리콘 인터포저 대신).

 

The placement of the bridge dies demands exceptional precision, particularly for the bridges between the two main compute dies, as they are essential for maintaining the 10 TB/s chipset-to-chipset interconnect. A significant design issue is rumored to involve these bridge dies, necessitating their redesign. Analysts from SemiAnalysis imply that there could be a coefficient of thermal expansion (CTE) mismatch between the GPU chiplets, LSI bridges, the RDL interposer, and motherboard substrate, which causes warpage and failure of the whole SiP. It has never been officially confirmed, though. Additionally, there are reports of a required redesign of the top global routing metal layers and bumps out of the Blackwell GPU silicon, which means a multi-month delay.

 

브리지 다이의 배치는 특히 두 개의 주요 컴퓨팅 다이 사이의 브리지에 대해 매우 정밀한 작업이 필요합니다. 이는 10 TB/s 칩셋 간 인터커넥트를 유지하는 데 필수적입니다. 주요 설계 문제는 이러한 브리지 다이와 관련이 있다는 소문이 있으며, 이로 인해 재설계가 필요하다는 이야기가 있습니다. SemiAnalysis의 분석가들은 GPU 칩렛, LSI 브리지, RDL 인터포저, 그리고 메인보드 기판 사이의 열팽창 계수(CTE) 불일치로 인해 SiP 전체가 변형되고 실패할 가능성이 있다고 암시하고 있습니다. 이 문제는 공식적으로 확인된 적은 없지만, Blackwell GPU 실리콘에서 상위 글로벌 라우팅 금속층과 범프의 재설계가 필요하다는 보고가 있으며, 이로 인해 수개월의 지연이 발생할 수 있다고 전해집니다.

 

SemiAnalysis reports that Nvidia will still produce 1000W B200 in low volumes in Q4 2024 for HGX servers, as planned, partly due to limited CoWoS-L capacity and partly due to the aforementioned issues. Also, high-end GB200-based NVL36 and NVL72 servers (using 1200W B200) will be available to some customers, albeit in low volumes in the fourth quarter. 

 

SemiAnalysis에 따르면 Nvidia는 2024년 4분기에 계획대로 HGX 서버용 1000W B200을 소량으로 생산할 예정입니다. 이는 제한된 CoWoS-L 생산 능력과 앞서 언급된 문제들 때문입니다. 또한, 고급형 GB200 기반의 NVL36 및 NVL72 서버(1200W B200을 사용하는)도 일부 고객에게 제공될 예정이지만, 4분기에는 소량으로만 제공될 것입니다.

 

To fulfill the demand for lower-end and mid-range AI systems, Nvidia is working on its B200A product featuring a monolithic B102 silicon with 144 GB (four stacks) of HBM3E and packaged using good-old CoWoS-S (or competing technologies from Amkor, ASE, SPIL, or even Samsung). This part is due in the second quarter of 2025. The new model will be available in 700W and 1000W HGX form factors, featuring up to 144GB of HBM3E memory and up to 4 TB/s of memory bandwidth. However, it is essential to note that this offers less memory bandwidth than the H200. However, whether the B102 die has anything to do with the graphics-oriented GB202 processor for graphics cards is unclear. 

 

Nvidia는 하위 및 중간급 AI 시스템에 대한 수요를 충족하기 위해 144GB(4 스택)의 HBM3E를 탑재한 단일 B102 실리콘을 특징으로 하는 B200A 제품을 개발 중입니다. 이 제품은 기존의 CoWoS-S(또는 Amkor, ASE, SPIL, 혹은 Samsung의 경쟁 기술)을 사용해 패키징되며, 2025년 2분기에 출시될 예정입니다. 이 새로운 모델은 최대 144GB의 HBM3E 메모리와 최대 4 TB/s의 메모리 대역폭을 특징으로 하며, 700W 및 1000W HGX 폼 팩터로 제공될 예정입니다. 그러나 이 제품의 메모리 대역폭은 H200보다 적다는 점을 주목할 필요가 있습니다. 또한 B102 다이가 그래픽 카드용 GB202 프로세서와 관련이 있는지는 불확실합니다.

 

We do not know whether Nvidia plans to redesign the original B200, its LSI, or the package itself. Still, SemiAnalysis claims that Nvidia is working on the mid-generation upgrade of the Blackwell series called the Blackwell Ultra. The 'Ultra' lineup will use two chiplets, CoWoS-L packaging, formally called B210 or B200 Ultra. The Blackwell Ultra includes a memory upgrade to up to 288 GB of 12Hi HBM3E and a performance boost in FLOPS of up to 50%. 

 

Nvidia가 원래 B200, LSI, 또는 패키지를 재설계할 계획이 있는지는 알 수 없지만, SemiAnalysis에 따르면 Nvidia는 Blackwell 시리즈의 중간 세대 업그레이드인 Blackwell Ultra를 개발 중이라고 합니다. 'Ultra' 라인업은 두 개의 칩렛과 CoWoS-L 패키징을 사용하며, 공식적으로 B210 또는 B200 Ultra로 불립니다. Blackwell Ultra는 최대 288GB의 12Hi HBM3E 메모리 업그레이드와 최대 50%의 FLOPS 성능 향상을 포함할 예정입니다.

 

Nvidia's B210/B200 Ultra is expected to be available in 1000W and 1200W versions sometime in Q3 2025, so the delay of high-end, high-volume Blackwell GPUs seems significant. Considering the demand for AI servers in general and Nvidia's H100/H200 GPUs, the green company will likely manage to navigate its B200-related issues.

 

Nvidia의 B210/B200 Ultra는 2025년 3분기쯤에 1000W 및 1200W 버전으로 출시될 것으로 예상되므로, 고급형 대량 생산 Blackwell GPU의 지연은 상당히 중요한 문제로 보입니다. 하지만 AI 서버와 Nvidia의 H100/H200 GPU에 대한 수요를 고려할 때, Nvidia는 B200과 관련된 문제를 잘 해결해 나갈 가능성이 큽니다.