AMD's 192GB Ryzen AI Halo vs Nvidia: What the Memory Advantage Buys

The machine is AMD's refreshed Ryzen AI Halo, with 192GB of unified memory and up to 160GB available to its GPU, against 128GB in Nvidia's DGX Spark. Capacity lets larger models and more agents fit. Both run at 273GB/s, though, so generation speed does not improve. Neither vendor has published prefill comparisons, so treat claims about prompt-processing advantage as unmeasured.

What AMD announced

The machine is the refreshed AMD Ryzen AI Halo developer platform. HotHardware reported on October 13, 2026 that it now uses the Ryzen AI Max+ PRO 495, a 16-core, 32-thread chip (codenamed Gorgon Halo) with an integrated 40-core Radeon 8065S GPU on RDNA 3.5 and an NPU rated up to 55 TOPS. The headline is memory: 192 GB of unified LPDDR5X-8533, of which up to 160 GB can be allocated to the GPU. HotHardware gave the starting price as $7,299 and said it is sold through Micro Center.

The chip itself was not new. ServeTheHome covered the Ryzen AI Max PRO 400 series in May 2026: the same 256-bit memory bus as the previous generation, faster memory, and capacity raised from 128 GB to 192 GB using newly available 24 GB LPDDR5X packages. AMD's pitch, per HotHardware, is support for 300-billion-parameter and larger models at 4-bit quantisation, and running several large agents at once, each with its own KV cache.

This is a different product from AMD's Instinct MI455X, the data centre accelerator with 432 GB of HBM4 that Phoronix reported launching in July 2026. The Halo is a desktop box.

The Nvidia comparison, precisely

The natural rival is Nvidia's DGX Spark. Nvidia's hardware documentation lists 128 GB of LPDDR5x unified memory on a 256-bit interface at 273 GB/s, and support for models up to 200 billion parameters, or 405 billion with two Sparks connected.

So the claim holds on one axis. A single Halo can give its GPU 160 GB, and a single Spark has 128 GB in total, shared with the operating system. ServeTheHome called 160 GB of GPU memory enough for a 300B model at FP4, a first for a single-chip system other than a Mac Studio.

On the other axis, the two are level. ServeTheHome puts the Halo's bandwidth at 273 GB/s, the same figure Nvidia lists for the Spark. Two Sparks also pass a single Halo on capacity. And the description of AI 365's video on the Halo says Nvidia keeps important strengths in prompt processing and power efficiency, consistent with ServeTheHome's warning that the extra memory is unlikely to move performance and that pre-fill bottlenecks remain.

Why capacity and bandwidth do different jobs

Local inference has two phases with two different bottlenecks.

Capacity decides what can run at all. Weights take roughly parameters times bits per weight divided by eight. A 300B model at 4-bit is about 150 GB of weights alone, which is why it needs the Halo's 160 GB allocation and leaves little room for context. Below capacity, a model does not load, or it spills to slower memory and becomes unusable for interactive work.

Bandwidth decides how fast it generates. Each new token reads the active weights from memory. At 273 GB/s, a dense model that reads 150 GB per token has a theoretical ceiling under 2 tokens per second. Capacity let the model in; bandwidth makes it slow.

Compute decides how fast it reads the prompt. Prompt processing is parallel and compute-bound. Agents send long prompts, so this phase often decides how long a user waits before the first token, and adding memory does nothing for it.

Where 192 GB genuinely helps

The extra memory pays off in workloads that are capacity-bound but not bandwidth-bound.

Large mixture-of-experts models

A mixture-of-experts model has a large total parameter count but activates a small fraction per token. Total size needs capacity; the bytes read per token stay small, so generation speed stays usable on 273 GB/s. This is the class of model where 160 GB of GPU memory turns into models the Spark cannot hold.

Several models and several agents

Agent systems often keep a planner, a coding model, and an embedding model resident at once, and every concurrent agent needs its own KV cache. Those caches grow with context length. Spare capacity here means fewer reloads and more agents working in parallel, which is the use AMD emphasised.

Long context

The KV cache scales with context length and concurrency. Room for long contexts is useful, with the caveat that long prompts still pay the compute-bound prefill cost on every turn they are reprocessed.

What it does not fix

A bigger box does not make long prompts cheap. Every token an agent sends must be processed, and on a bandwidth-limited and prefill-limited machine that cost lands as latency on every turn. The most effective lever is the same on a $7,299 workstation as on a laptop: send less. Retrieval that hands the model the relevant passages instead of whole documents shortens prefill, shrinks the KV cache, and frees capacity for the model itself. RDK does this by indexing local vaults, docs, and code as encrypted private chunks that agents search before they call any model.

Should you buy one

Buy for a named workload. If the model you need only fits in more than 128 GB, and it is a mixture-of-experts model or you accept slow generation, the Halo is one of very few single desktop systems that can hold it. If your models fit in 128 GB, the extra capacity buys headroom, not speed. If prompt processing time is the pain, more memory is the wrong fix. Look for measured prefill and generation speeds on your own model and context length, not launch-day claims.

Frequently asked questions

What is the AMD Ryzen AI Halo?
It is AMD's compact developer workstation for local AI. The refreshed version, reported by HotHardware on October 13, 2026, uses the Ryzen AI Max+ PRO 495 with a 40-core Radeon 8065S GPU and 192 GB of unified LPDDR5X-8533 memory, up to 160 GB of which can serve as GPU memory. HotHardware reported a starting price of $7,299 through Micro Center.
Does the Ryzen AI Halo beat Nvidia DGX Spark?
On memory capacity, yes: 192 GB with up to 160 GB for the GPU against the Spark's 128 GB total, per AMD and Nvidia figures. On memory bandwidth they are level at 273 GB/s, so generation speed on models that fit both is similar in principle. Nvidia is described as stronger in prompt processing, and two connected Sparks exceed one Halo's capacity.
Can a 192GB machine run a 300B parameter model?
It can load one at 4-bit, which ServeTheHome and AMD both note. The weights alone are about 150 GB, so little room remains for context. Speed depends on the architecture: a dense 300B model would generate very slowly at 273 GB/s, while a mixture-of-experts model reads far fewer bytes per token and can be usable.
Why does memory matter more than compute for local AI?
Capacity decides which models load at all, and bandwidth caps how fast tokens are generated, because each token reads the active weights from memory. Compute matters most for prompt processing, which agents stress with long prompts. A machine needs all three, but for local inference memory is usually the first wall.