LecturaInvestigaciónRadarMarco de inversión
Iniciar sesión / Registrarse
← Volver a la lectura
NoticiasTema · 美光内存架构之争Material4 min de lecturaEditorial de DisconfirmAI

Chip Architecture Allocates Memory by Workload: Chipstrat Analyst Austin Lyons Argues Micron's 'More Memory is Better' Answer Misses the Point

After attending a Micron Technology product webinar, Chipstrat analyst Austin Lyons pointed out that Micron executives' response to the question of "despeccing"—claiming that "more memory equals better system performance"—avoided the real issue: how total memory demand and structure will change when training, prefill, and decoding use different chips and different types of memory. Citing examples such as Google's TPU 8i/8t, Nvidia's Rubin CPX, and Positron's Asimov, he noted that customers are already precisely configuring memory according to workloads. Since HBM consumes three times the wafers per bit compared to DDR5 and currently has lower margins, migrating bits from HBM to LPDDR/GDDR could improve rather than harm Micron's profits.

2026-10-27 23:30 GMT+8

Micron held its Technology and Product Webinar last week. Chipstrat analyst Austin Lyons watched it and recorded his analysis, with the core judgment being: Micron gave an answer that was correct but missed the point regarding the memory demand question most concerning investors.

The starting point of the question came from analyst Vivek Arya. He asked a crucial question for memory manufacturers: today, memory accounts for more than half of an accelerator's bill of materials. Customers are discussing "despeccing," i.e., using less or cheaper memory. Could inference accelerators eventually be equipped with less DRAM or different types of DRAM? If so, Micron's bit sales would decline.

Micron's Scott DeBoer answered that system performance improves with increased memory capacity and bandwidth. Customers will choose the minimum specification that meets their product and business needs. Micron's job is to increase supply and drive bandwidth higher through products like vertical HBM. He concluded with: "Inherently, more memory does equal better system performance."

Iniciar sesión / Registrarse
LecturaInvestigaciónRadarMarco de inversión

Lyons believes this answers the wrong question. DeBoer is right—more memory does help performance—but Arya didn't ask "is more memory better?"; he asked "what does it mean for memory suppliers when customers configure memory precisely according to workload?"

Lyons uses Google's chip design to illustrate that customers are already doing this. Publicly available specifications for Google Cloud's TPUs show: the inference chip TPU 8i has less compute but more memory; the training chip TPU 8t has more compute but less memory bandwidth because training is primarily compute-intensive. This is not hypothetical; it is differentiated products already in the market.

He further breaks down inference into two stages: prefill processes prompts and is compute-intensive; decode generates output token by token and is memory-intensive. Disaggregated inference runs these two stages on separate chips. This means prefill chips may not need as much HBM, while decode chips need more—and racks often contain more decode chips than prefill chips. Whether total memory demand decreases or offsets depends on the specific ratio.

Chipmakers are already differentiating memory configurations based on workload. AMD's MI455X carries 432GB of HBM4, and Nvidia's Rubin carries 288GB. Nvidia also announced the Rubin CPX, a prefill chip using 128GB of GDDR7 instead of HBM. Startup Positron's Asimov chip skips HBM entirely, using LPDDR5X with configurable capacities from 288GB to 2.3TB, allowing customers to choose capacity based on workload.

Here arises a dimension often overlooked: not just total volume, but structure (mix). GDDR7 and LPDDR are still DRAM, and Micron sells both. However, their profit margins differ from HBM. Lyons cites Micron's own data: HBM consumes three times the wafers per bit of DDR5, and "this trade ratio only increases with future generations of HBM." Therefore, every bit migrated from HBM to LPDDR or GDDR due to architectural optimization releases a significant amount of DRAM wafers.

The key counter-intuitive judgment here: Lyons points out that HBM currently has lower margins than Micron's other DRAM products. Thus, bits migrating from HBM to LPDDR/GDDR might not be bad news; they could actually boost profits. This means simply tracking the total volume of HBM bits is insufficient to judge Micron's profitability trajectory; one must simultaneously track changes in demand structure.

Lyons concludes that memory demand now depends on how each accelerator is sized for its workload. Forecasting demand means tracking the direction of accelerator architecture—how much HBM does a training chip carry? Does prefill use HBM or GDDR? Is the KV cache placed in HBM, SRAM, or LPDDR? These architectural choices determine how many bits the industry needs and what kind of bits.

He also mentioned that Micron co-designs custom base dies with customers starting from HBM4E, so it should understand these trends more thoughtfully than the answer given at the meeting. Lyons believes Micron indeed has answers, they just haven't been fully articulated.

The second part of the article raises another question—who captures the value when customers put their own logic into the base die—but that section is paywalled and unreadable.

Implications for investors: When evaluating memory companies like Micron, accelerator architecture roadmaps should be treated as core variables for demand forecasting, rather than just looking at HBM shipment volumes or total DRAM bits. Architectural differentiation may lead to growing total bit volumes but improved margins, or flat total bit volumes but declining profits due to structural changes. This is a mechanism requiring continuous tracking, not a one-time event.

Fuentes

  • newsletter.chipstrat.com↗
Tema

美光内存架构之争

【背景】美光科技(Micron Technology)正面临关于AI加速器内存需求的核心争论:随着芯片设计趋向“按负载分配内存”(right-sizing),传统的“更多内存等于更好性能”逻辑是否依然成立?

4 pasos1 piezasActualizado 2026-10-28 13:09

Cómo llegó hasta aquí

  1. 2026-08-18据BofA全球研究报道,英伟达正在测试将Rubin Ultra加速器的HBM配置从计划的1TB降低至192GB-288GB,原因是HBM4e供应限制和资格认证延迟。finance.yahoo.com ↗
  2. 2026-10-21美光首席技术官Scott DeBoer在网络研讨会上公布2027年技术路线图,称4F² DRAM设计存在根本问题,并指出新AI内存类型每比特消耗更多硅片,导致常规内存供应进一步收紧。aciliszili.com ↗
  3. 2026-10-24美光公布2026财年第四季度营收为542.3亿美元,同比增长379%,并指引下一季度营收为615亿美元。公司强调其26项战略客户协议旨在打破传统内存周期波动。top1markets.com ↗

Lo que aún está pendiente

  • 英伟达Rubin Ultra的最终量产配置是否会维持低内存版本,还是随HBM4e供应恢复回归高容量?
  • 美光能否证明在客户进行‘内存右尺寸’(right-sizing)优化后,其HBM市场份额和利润率仍能保持增长?
  • 美光批评的4F² DRAM设计若被长鑫存储(CXMT)和三星成功商业化,将对美光的3D DRAM路线构成多大竞争压力?
Leer el tema completo →
Continuar investigando →