Coding Horizon

192GB Just Broke The Local AI Memory Wall Open

Sources for every figure, name and benchmark this video states. Checked 2026-09-03.

The platform

AMD’s Ryzen AI Max PRO 400 series, codenamed Gorgon Halo, was announced at Computex and ships in systems from Q3 2026. Three parts, all sharing the same memory ceiling:

Part Cores / threads Cache iGPU CUs NPU TDP
Ryzen AI Max+ PRO 495 16 / 32 80MB Radeon 8065S 40 55 TOPS 45W to 120W
Ryzen AI Max PRO 490 12 / 24 76MB Radeon 8050S 32 50 TOPS 45W to 120W
Ryzen AI Max PRO 485 8 / 16 40MB Radeon 8050S 32 50 TOPS 45W to 120W

Sources:

The 300 billion parameter claim

AMD’s own framing is capacity, not throughput: the platform has “enough space to load up a 300B FP4 parameter model, a first for any single SoC system”. FP4 is four bit, which is the precision the claim depends on. AMD has published no inference throughput figures for a 300B model at that precision, and has not said which 300B class architectures it means.

The weight arithmetic behind it is straightforward and is the reason 192GB matters where 128GB does not: 300 billion parameters at four bits is half a byte each, so roughly 150GB for the weights alone, before the KV cache, the runtime and the operating system are given anything. That is why the claim is physically plausible at 192GB and not at 128GB.

Sources:

What it is being compared against

Machine Memory Bandwidth Power
Ryzen AI Max+ PRO 495 192GB unified, up to 160GB as graphics LPDDR5X-8533 45W to 120W
GeForce RTX 5090 32GB GDDR7 1,792 GB/s 575W
RTX PRO 6000 Blackwell Workstation 96GB GDDR7 ECC 1,792 GB/s 600W
Mac Studio, M3 Ultra up to 512GB unified as announced over 800 GB/s —
Instinct MI300X 192GB HBM3 5.3 TB/s 750W

Sources:

Caveat: the 512GB Mac Studio is no longer orderable

Apple’s 512GB M3 Ultra configuration was real at announcement in March 2025 and the figures quoted above are Apple’s own. Some time before March 2026 Apple quietly removed it, and the M3 Ultra Mac Studio now tops out at 256GB, with the 96GB to 256GB upgrade rising from $1,600 to $2,000. Reporting attributes the change to the DRAM and NAND shortage.

The comparison the video draws still holds in direction, because 256GB is still above 192GB, but 512GB describes the machine as launched rather than the machine as sold today.

What 128GB Strix Halo already measures

The video cites two separate sets of figures, and the spread between them across the wider body of testing is the point rather than a problem: the same model on the same silicon moves substantially with the backend.

A Strix Halo laptop, Ryzen AI Max+ 395 with Radeon 8060S integrated graphics. Models run from Qwen3-4B (3.98 GiB) through gpt-oss-20b (11.27 GiB) to gpt-oss-120b (59.02 GiB). gpt-oss-120b generated at 35 to 40 tokens per second, reported as 40 t/s on the Vulkan backend. Qwen3-4B at Q8_0 reached 41 to 44 t/s depending on backend.

A 128GB Ryzen AI Max+ 395 guide. Qwen3-Coder 30B-A3B at Q4_K_S measured 98.51 t/s, Qwen3-30B-A3B 100.04 t/s, and gpt-oss-120b at MXFP4 55.57 t/s. The same source measures memory bandwidth at roughly 215 GB/s against a 256 GB/s theoretical peak, a gap it describes as typical for LPDDR5X under mixed CPU and GPU load.

Other measurements of the same model, for the spread. gpt-oss-120b has also been reported at roughly 31 t/s and roughly 34 t/s on Ryzen AI Max+ 395 systems. Architecture matters more than parameter count on this hardware: a dense 70B runs at around 5 t/s while a 30B mixture of experts runs at 70 to 100 t/s on the same machine, because token generation is bandwidth bound and a sparse model reads far fewer weights per token.

Why 192GB is not expected to be a speed jump

The compute does not move: the flagship is still 40 RDNA 3.5 compute units and still Zen 5, the same counts as Strix Halo’s top part. The memory data rate moves from LPDDR5X-8000 on Strix Halo to LPDDR5X-8533 here, a step of under seven per cent in data rate. The capacity moves by fifty per cent, from 128GB to 192GB. That asymmetry is the whole argument: it is a capacity release, not a throughput release.

The software position

The video’s claim that ROCm does not yet have CUDA’s gravity is specific and checkable. gfx1151, which is the Strix Halo graphics target, is listed only as Preview in AMD’s ROCm compatibility matrix rather than as a fully supported target. Running it today generally means overriding the target version, and first class gfx1151 support is being targeted at ROCm 8.0. llama.cpp and Ollama work through ROCm with that override, and LM Studio’s Vulkan backend works without it.

Not chased to a primary source