The Prospects for 128 Bit Processors ( John Mashey SGI 1995)

6 hours ago

Physical address sizes increase by 2 bits every 3 years, tracking DRAM density growth, and need to cover a 4:1 to 16:1 range to accommodate future memory sizes.
Virtual memory pressure often exceeds physical memory by at least 4x, driven by memory-mapped files and large file systems with holes, accelerating address space demands.
Extrapolating from current trends, a transition from 64-bit to 128-bit addressing would occur around 2043 for physical memory, or earlier (circa 2020) if virtual memory pressure increases.
Practical concerns like die space, wire delays, and the complexity of wide datapaths (e.g., 1024-bit) make such architectures unlikely without strong justification, with 128-bit being the more realistic near-term goal.
32-bit CPUs are already insufficient for some high-end applications, while 64-bit CPUs are expected to remain adequate for a long period, with affordable options already available.

AI token consumption per user has surged dramatically, with the author estimating a 50x increase in daily usage over three years, driven by advanced models like GPT-4, Sonnet 3.5, and Opus 4.5, and the rise of agentic workflows.
The explosive growth in LLM users (estimated at 1 billion active users) and per-user token usage is fueling massive infrastructure investments, with hyperscalers like AWS, Azure, and GCP committing hundreds of billions in capex for datacenter buildouts.
Despite the planned infrastructure spending, actual deployment faces hard constraints, particularly electrical power shortages and a critical shortage of high-bandwidth memory (HBM DRAM), which is essential for AI chips and takes years to ramp production.

Local on-device AI offers privacy, zero latency, and no API costs, but is largely impractical for most consumer devices.
Consumer hardware faces RAM constraints, with 8-16GB common, leaving only 4-8GB for AI models after OS and apps.
A capable 7B parameter model requires ~5GB, but agentic tasks also need large KV cache, quickly exceeding 8GB RAM.

Soldered LPDDR RAM on x86 NAS eliminates memory upgrades, creating a permanent ceiling for heavy workloads and a single point of failure.
Persistent 1GbE ports in mid-range NAS, with 10GbE locked behind high price premiums, using multi-gigabit as an artificial segmentation tool.
Vendor drive lock-in policies restrict third-party HDD/SSD usage, requiring proprietary drives for full features, increasing costs.

Mag World uses RAM as a proxy to measure computing power across orders of magnitude, from mag 2 (128 bytes) to mag 10+ (16+ GB).
Moore's Law drove exponential growth in transistor density, but the pace has slowed; personal experience shows a 10x improvement every 5–7 years.
First computers varied by generation: mag 3 (TRS-80, 4KB RAM), mag 4 (Apple IIe), mag 5–6 (IBM 386, 2MB RAM), and mag 7–8 (Windows 95, WoW-era PCs).

AI compute demand is exploding, with GitHub commits up ~14x annualized, indicating massive inference growth.
AI companies are compute-starved, leading to usage limits, outages, and a domino effect as users switch between services.
Supply constraints include delayed GB200 rollout due to liquid cooling challenges, labor shortages, and geopolitical tensions.

AI token consumption per user has surged dramatically, with the author estimating a 50x increase in daily usage over three years, driven by advanced models like GPT-4, Sonnet 3.5, and Opus 4.5, and the rise of agentic workflows.
The explosive growth in LLM users (estimated at 1 billion active users) and per-user token usage is fueling massive infrastructure investments, with hyperscalers like AWS, Azure, and GCP committing hundreds of billions in capex for datacenter buildouts.
Despite the planned infrastructure spending, actual deployment faces hard constraints, particularly electrical power shortages and a critical shortage of high-bandwidth memory (HBM DRAM), which is essential for AI chips and takes years to ramp production.

Local on-device AI offers privacy, zero latency, and no API costs, but is largely impractical for most consumer devices.
Consumer hardware faces RAM constraints, with 8-16GB common, leaving only 4-8GB for AI models after OS and apps.
A capable 7B parameter model requires ~5GB, but agentic tasks also need large KV cache, quickly exceeding 8GB RAM.

Soldered LPDDR RAM on x86 NAS eliminates memory upgrades, creating a permanent ceiling for heavy workloads and a single point of failure.
Persistent 1GbE ports in mid-range NAS, with 10GbE locked behind high price premiums, using multi-gigabit as an artificial segmentation tool.
Vendor drive lock-in policies restrict third-party HDD/SSD usage, requiring proprietary drives for full features, increasing costs.

Mag World uses RAM as a proxy to measure computing power across orders of magnitude, from mag 2 (128 bytes) to mag 10+ (16+ GB).
Moore's Law drove exponential growth in transistor density, but the pace has slowed; personal experience shows a 10x improvement every 5–7 years.
First computers varied by generation: mag 3 (TRS-80, 4KB RAM), mag 4 (Apple IIe), mag 5–6 (IBM 386, 2MB RAM), and mag 7–8 (Windows 95, WoW-era PCs).

AI compute demand is exploding, with GitHub commits up ~14x annualized, indicating massive inference growth.
AI companies are compute-starved, leading to usage limits, outages, and a domino effect as users switch between services.
Supply constraints include delayed GB200 rollout due to liquid cooling challenges, labor shortages, and geopolitical tensions.

Hasty Briefsbeta