Your Memory is Getting its OWN CPU?? - WAN Show February 19, 2021

Episode Summary

Executive Summary: The episode centers on Samsung’s AI-enabled HBM2 memory breakthrough, NVIDIA’s move to curb Ethereum mining on the RTX 3060, and Facebook’s decision to block news links in Australia and Canada. It also features a long, humorous analysis of XQC’s PC build video, highlighting how stream culture, chat noise, and hardware complexity affect DIY building.

Main Topics: Samsung HBM2 memory with integrated AI processors (Priority: 5/5): The hosts unpack Samsung’s claim that its HBM2-based processing-in-memory design can accelerate AI/data-center workloads by moving computation closer to memory, potentially improving performance-per-watt without changing other hardware or software. NVIDIA RTX 3060 mining limiter and CMP cards (Priority: 5/5): They discuss NVIDIA’s software-based restriction on Ethereum mining for the RTX 3060 and the launch of mining-specific CMP cards, debating whether this really preserves GPU supply for gamers and whether the mining cards are actually more efficient or practical. Watching XQC watch the LTT PC build (Priority: 4/5): A major segment reacts to XQC’s stream response to Linus building a PC, mixing hardware commentary with meta-humor about reaction videos, streamer culture, chat delays, and the difficulties of communicating with high-volume creators. Counterfeit LTT merchandise investigation (Priority: 4/5): The hosts describe buying fake Linus Tech Tips merchandise to gauge the scale of the problem, concluding the scam products are low quality and likely low-volume, with few reviews and short-lived storefronts. Facebook blocking news in Australia and Canada (Priority: 4/5): They analyze Facebook’s response to proposed link-tax laws by blocking news links, noting the collateral damage to charities and emergency services and the broader tension between platforms and publishers over algorithmic transparency and revenue sharing. Xbox Series X FPS Boost and nostalgia gaming (Priority: 2/5): They briefly cover Microsoft’s FPS Boost updates for backward-compatible titles and pivot into broader nostalgia/remaster talk about older Blizzard and console games, noting the appeal of revisiting classics with better performance.

Key Arguments: Samsung’s AI-in-memory approach is not just buzzwording; it could materially reduce latency by doing processing inside memory banks rather than across the CPU-memory boundary. The claimed improvement likely applies to the memory subsystem, not the whole system, even if marketing language suggests system-wide gains. NVIDIA’s RTX 3060 mining limiter is a targeted attempt to steer miners toward CMP hardware and away from gaming GPUs, but miners may still prefer standard cards if they are more efficient or easier to resell. Facebook’s news block is a strategic response to legislation that could force it to pay publishers and reveal algorithmic changes, which it sees as core competitive secrets. The XQC build illustrates that even a knowledgeable creator can get tripped up by PC assembly, especially when live chat introduces noise, delay, and conflicting advice. Counterfeit merch exists but appears to be a low-scale, low-quality operation rather than a major business threat. Good platform policy and tech gains often come with tradeoffs: Samsung may sacrifice capacity for compute, Facebook sacrifices news distribution, and NVIDIA may not actually solve GPU shortages.

Data Points: Performance improvement: over 2x - Samsung claims the HBM2 PIM design can deliver more than double system performance in certain AI/data-center workloads. Power reduction: over 70% - Samsung claims the memory innovation can reduce power consumption significantly, likely for the memory subsystem. Memory capacity change: 8 GB stack to 6 GB stack - Host explains that adding PCUs to HBM2 stacks reduces usable capacity compared with conventional stacks. Compute unit throughput: 1.2 teraflops - Samsung paper title references a programmable computing unit using bank-level parallelism for machine learning applications. Process node: 20 nm - The in-memory DRAM design is described as being fabricated in a 20-nanometer DRAM process. Mining hash-rate reduction: ~50% - NVIDIA’s RTX 3060 driver is designed to detect Ethereum mining and cut hash rate roughly in half. CMP 90HX efficiency: 86 MH/s at 320 W - The top mining-specific card discussed as part of NVIDIA’s CMP lineup. RTX 3080 efficiency: 100 MH/s at 250 W - Used as a comparison showing the gaming GPU can outperform the mining card on both speed and efficiency. Number of games mentioned for FPS Boost: 5 - Microsoft update adds boost to five previous-gen titles in the segment. WAN show audience interaction: 3-month trial - Remote HQ sponsor offer included a free three-month trial. Merch haul: a few hundred dollars - They spent this much on counterfeit Linus merch for the investigation. XQC build RAM: 128 GB - The streamed PC build featured 128 gigabytes of memory. XQC build CPU choice: Ryzen 9 5950X / high-end AMD CPU context implied - The build discussion centers on an expensive, top-tier AMD setup (speaker mentions the motherboard/CPU as premium components). Audience scale concern: 200 to 400 people - Linus describes meeting roughly this many people at trade shows and forgetting names due to overload.

Pivotal Quotes: "This is what Musk warned us about. First, you're going to get AI memory. Next, your computer's going to build its own legs and arms and take over." — Linus: Joking introduction to Samsung’s AI-enabled memory announcement. "If I de-bolted the motherboard a little bit, a tiny bit, and I loosened it and I pushed the whole motherboard to the side and then re-bolted it, it would have given me a bigger play zone to get the screws in." — XQC: Moment from the PC build stream showing his improvised explanation for case fitment issues. "The value of reporting. There's a value to news media. And it's clear that business has become unsustainable in a way." — Linus: Reflection on why some regulation around news distribution is understandable, even if Facebook’s reaction is harsh.

Implications: The episode suggests PC hardware is becoming more specialized and embedded with compute, while platform fights over content and creator culture keep shifting power toward large tech companies and live-stream ecosystems.