Kunal Khullar is a contributing writer at Tom\u2019s Hardware.\u00a0 He is a long time technology journalist and reviewer specializing in PC components and peripherals, and welcomes any and every question around building a PC. ","collapsible":{"enabled":true,"maxHeight":250,"readMoreText":"Read more","readLessText":"Read less"}}), "https://slice.vanilla.futurecdn.net/13-4-22/js/authorBio.js"); } else { console.error('%c FTE ','background: #9306F9; color: #ffffff','no lazy slice hydration function available'); } Kunal Khullar Social Links Navigation News Contributor Kunal Khullar is a contributing writer at Tom’s Hardware. He is a long time technology journalist and reviewer specializing in PC components and peripherals, and welcomes any and every question around building a PC.
usertests As far as we can tell, practically nothing is different between Strix Halo and Gorgon Halo, so for the 300-series chips already manufactured out there, you would think it would be possible to pair them with 192 GB of memory. You might conclude 24 GB is the largest 32-bit package being made. But based on this story, there are already 128-bit SOCAMM2 modules using 32-bit 64 GB LPDDR5X packages (16x 32 Gb dies). So it should be possible to make a 512 GB Strix/Gorgon Halo product. 16-layer packages are definitely more expensive to make per-gigabyte than 8-layer though. "Nobody needs 192-512 GB with a product this slow" And I don't need 64 GB in my trash system. But from what I'm hearing, mixture of experts (MoE) models could run faster than expected on Strix Halo, while benefiting from having their entire contents in the large memory pool. It may not be worth the $4k-5k a 192 GB Strix/Gorgon Halo box would probably cost, but "more memory better" remains technically true. Hopefully, everyone can have a cheap terabyte of memory by 2040. Reply
w_barath I'm so happy to see Tom's excellent deep understanding of the product's place in the market and the needs of that place in the market, taken into account in the thoughtful research invested into this article! The one metric that would significantly improve this Gorgon Halo over Strix Halo would be increased memory bandwidth. Great job reporting! 🌟 Reply
usertests w_barath said: The one metric that would significantly improve this Gorgon Halo over Strix Halo would be increased memory bandwidth. It should have explicit 8533 MT/s support, instead of 8000 MT/s. There's your 6.6% improvement. Reply
w_barath usertests said: It should have explicit 8533 MT/s support, instead of 8000 MT/s. There's your 6.6% improvement. AOOSTAR MAX+ 395 and HP ZBook Ultra G1a have trusted source reviews stating they have 8533MT/s. A couple of them have recently retconn'd their articles to say 8000, which I assume is in response to pressure from AMD saying that 8000 is the official number. Reply
Notton I guess the next question is… Is the AMD memory controller robust enough to use 9600MT/s or 10700MT/s? 8000MT/s wasn't impressive when Strix Halo was released. 8533MT/s isn't impressive whenever Gorgon Halo releases. Reply
DS426 I'm sure someone will be able to use more than 128 GB of shared RAM in their workloads. usertests said: … And I don't need 64 GB in my trash system. But from what I'm hearing, mixture of experts (MoE) models could run faster than expected on Strix Halo, while benefiting from having their entire contents in the large memory pool. … And longer contexts, and keeping more LLM's loaded into memory simultaneously (makes sense in some coding and agentic applications), agents working out into the horizon for longer, and even some combination of all of this. The new 192 GB mem limit is definitely usable, not just marketing and over-the-top halo product specs. Faster mem support would be nice, yes. We'll have to see when the final specs are announced. Reply
w_barath DS426 said: I'm sure someone will be able to use more than 128 GB of shared RAM in their workloads. And longer contexts, and keeping more LLM's loaded into memory simultaneously (makes sense in some coding and agentic applications), agents working out into the horizon for longer, and even some combination of all of this. The new 192 GB mem limit is definitely usable, not just marketing and over-the-top halo product specs. Faster mem support would be nice, yes. We'll have to see when the final specs are announced. The trend since early 2025 has actually been that there are more capable models using less RAM, but the memory bandwidth problem remains. Google has released their TurboQuant paper which is supposed to reduce the Key Value Cache memory bandwidth for the by a factor of 6 for the inference stage, with no significant loss of output quality. That should roughly double overall performance for most models which adopt it, but that doesn't solve the bottleneck, it just reduces its impact. I'm optimistic that this will be the year of competent (long horizon) and performant (>20 TPS) coding models with under 30GB VRAM residency. Qwen3.5-35B-A3B is a sparse MOE model that is nearly there already. When we cross this threshold, AI coding assistance will be truly democratized. That will lead to a 3-pronged leap. First, developer productivity. Second, a leap in AI development pace due to increased developer productivity. And third, as more developers realise the gifts of AI, the higher the proportion of developers and engineers who will turn their focus to improving the systems concepts themselves. You factor those things and the rate of growth will be step change quarterly. Reply
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/pc-components/cpus/SPONSORED_LINK_URL
- https://www.tomshardware.com/pc-components/cpus/amd-ryzen-ai-max-pro-495-apu-could-arrive-with-192gb-of-unified-memory-leaked-passmark-benchmarks-suggest-modest-update-over-strix-halo#main
- https://www.tomshardware.com
- Sony increases prices for refurbished PS5 slims by $100 — PS5 Fortnite bundle is out of stock, marking the end of new $399 consoles
- Grab a 1440p-capable gaming rig with an RTX 5060 for just $1,049 — Save up to 25% on ABS Flux & Cyclone prebuilts that also feature 32 GB of RAM
- Anthropic in early talks to buy DRAM-less AI inference chips from UK startup — Fractile's SRAM architecture reduces need for pricey memory during extreme pricin
- Nemotron Labs: What OpenClaw Agents Mean for Every Organization
- AMD Ryzen AI 5 435G APU breaks cover in early benchmarks — six-core Zen 5 chip goes head-to-head with the Ryzen 5 8600G for budget PC builders
Informational only. No financial advice. Do your own research.