Ryzen AI Max PRO 400: what 192 GB enables ?

#What AMD announces with the PRO 400

In 2026, AMD introduced the Ryzen AI Max PRO 400 series, built on the Zen 5 architecture with RDNA 3.5 graphics and an XDNA 2 NPU. On its official blog, AMD writes that these processors combine “192 GB of system memory and 160 GB of VRAM” for mobile workstations or small desktops aimed at AI, content creation, and simulation. This is a manufacturer announcement: the figures come from AMD, not an independent measurement.

The target audience is professional: AI developers, engineers, and creators working with large datasets. AMD cites HP and Lenovo as OEM partners, and the technical press (ServeTheHome) adds ASUS to the list of confirmed partners.

#192 GB of memory, 160 GB for the GPU: the nuance you shouldn’t miss

This is the point that causes the most confusion: the 160 GB announced for the GPU is not added to the 192 GB of total memory. It is 160 GB TAKEN from the unified 192 GB pool, with the remainder (32 GB) available for the system and CPU. AMD specifies this in its technical note: the Ryzen AI Max+ 495 PRO “supports up to 160 GB of dedicated graphics memory.” So there is no configuration with 192 GB for the CPU AND an additional 160 GB for the GPU.

#What 192 GB really enables (Q4 model sizes)

Using the usual memory estimates for Q4 quantization (weights only, excluding context: 70B ≈ 40 GB, 32B ≈ 19–20 GB), 160 GB of GPU memory leaves room for models well beyond 70B. AMD states, in a technical footnote, that the Ryzen AI Max+ 495 PRO “supports up to 160 GB of dedicated graphics memory, capable of running a model with more than 300 billion parameters in 4-bit quantization.”

The calculation shows one simple thing: loading a 300B model in Q4 nearly saturates all GPU memory, leaving very little room for the context (KV cache) and system overhead. A long context on a model this size therefore isn’t guaranteed even if the model technically fits in memory.

#Why memory bandwidth will hold back what 192 GB promises

This is the point marketing announcements don’t highlight: memory capacity is not generation speed. In the vast majority of cases, an LLM’s generation throughput is limited by memory bandwidth—the model must be reread from memory for every token. The formula to remember, presented as a theoretical estimate rather than a measurement: tokens/s ≈ bandwidth (GB/s) ÷ active weight size (GB).

The PRO 400 advertises 273 GB/s of bandwidth with LPDDR5X-8533, according to ServeTheHome. Applying the formula to a 300B model in Q4 (≈ 160 GB of weights): 273 ÷ 160 ≈ 1.7 tokens/s. This is a theoretical upper-bound estimate, not an actual measurement published by AMD or an independent tester—but it gives an order of magnitude incompatible with fluid conversation. On a dense 70B model (≈ 40 GB), the estimate rises to about 6.8 tokens/s; on a 32B model (≈ 20 GB), to about 13.6 tokens/s.

#The three chips in the PRO 400 lineup

The three chips share the same maximum memory capacity of 192 GB; only the CPU cores, frequency, and number of graphics units differ. For local LLM use, the 495 PRO (40 CU) has the most compute power for processing the prompt and managing the context, even though memory bandwidth — and therefore generation speed — remains identical across all three models.

#“Ryzen AI Halo” Developer Kit and PRO 400 OEM PC: two distinct products

AMD announced two different offerings on the same day, which several press outlets have conflated. The “Ryzen AI Halo” is, according to AMD, “the first compact AI development platform designed by AMD,” intended for local prototyping without depending on the cloud. Its first version is based on a Ryzen AI Max+ 395—the previous-generation chip, not the PRO 400—with up to 128 GB of unified memory, sold exclusively through Micro Center with preorders opening in June 2026. It is not until the third quarter of 2026 that AMD plans to upgrade this same developer kit to a PRO 400 Series chip, with up to 192 GB of memory and 160 GB of VRAM.

The Ryzen AI Max PRO 400 discussed here is a separate product: a component that manufacturers (HP, Lenovo, ASUS) integrate into their own business PCs, at their own prices, with no direct connection to Micro Center’s sales channel for the developer kit. To avoid confusion when purchasing: the “Halo” kit is an AMD machine you order directly; the PRO 400 is a chip found in third-party branded PCs, whose availability and pricing depend on each manufacturer.

#Availability: what is confirmed and what isn't

AMD announces availability through OEM partners “including HP and Lenovo” in the third quarter of 2026; ServeTheHome adds ASUS to the list of confirmed partners. As of this writing, no public price or specific consumer-machine model has been verified for France: these are professional PCs and mobile workstations, not (at this stage) mini PCs affordable on an individual’s budget like the previous generation.

The first concrete example documented by the tech press: in early September 2026, Lenovo unveiled its ThinkCentre X Ultra, a mini workstation featuring “up to” a Ryzen AI Max+ PRO 495 with an integrated Radeon 8065S GPU—but in the announced configuration, it offers 128 GB of RAM, not the chip's maximum 192 GB. Tom's Hardware reports a planned November 2026 release, with an “expected” starting price of $3,669, while specifying that this price “surely does not” apply to the high-end configuration. The takeaway: AMD's announced 192 GB is a chip limit, not a guarantee of a configuration at an entry-level price—you need to check each machine's exact specifications before expecting the maximum memory at the lowest price.

#Against the Ryzen AI Max+ 395: what really changes

The previous generation, the Ryzen AI Max+ 395 (Strix Halo), tops out at 128 GB of unified LPDDR5X-8000 memory, with theoretical bandwidth of 256 GB/s (256-bit bus)—it is the chip used in mini PCs such as the GMKtec EVO-X2, already available to consumers. The PRO 400 adds 64 GB of memory (192 versus 128) and slightly higher theoretical bandwidth (273 versus 256 GB/s, or about 7%), but at this stage remains positioned in professional machines rather than affordable mini PCs.

• Compare with the Ryzen AI Max+ 395 (128 GB)
• Complete guide: Ryzen AI Max+ 395 for local LLMs
• Compare with a dedicated GPU
• Understanding the context window and its memory cost
• MoE explained: why an “expert” model saves memory
• Source: official AMD announcement
• Source: ServeTheHome analysis
• Source: Tom's Hardware on the Lenovo ThinkCentre X Ultra

#Buy now or wait?

If you need to load a dense 70B to 120B model with a comfortable context window without spending on a multi-GPU workstation, the PRO 400 addresses a real use case: its memory capacity is real and verified by AMD. If your goal is to run a 300B model in a smooth interactive conversation, the bandwidth estimate above (around 1 to 2 tokens/s) suggests waiting for independent measurements before committing a professional budget: the capacity is sufficient, but the speed remains unproven.

You already have a Ryzen AI Max+ 395 (128 GB)

The additional 64 GB of gain does not justify an immediate repurchase for most 30B–70B uses, except when you specifically need a very long context or models above 120B.

You are targeting large MoE models

This profile gets the most out of large memory with modest bandwidth: few active parameters per token, so throughput is less penalized than with a dense model of the same size.

You need a consumer machine today

The PRO 400 is advertised for professional PCs and workstations; make sure a configuration fits your budget before waiting for a more affordable variant.