AMD Gorgon Halo 192GB for Local AI (2026): +50% Memory, +6.6% Bandwidth — What the Extra 64GB Actually Buys

Every piece of coverage about AMD's new Halo part leads with the same number: 192GB. It is the right number to lead with and the wrong number to buy on. The specification that decides how fast your local model actually answers barely moved.

This guide does the arithmetic nobody has published yet, then turns it into a purchase decision. Everything here is checkable from published spec sheets and independently measured benchmarks, and where a number is a vendor claim or a projection we say so. If you want the previous-generation version of this decision, our Strix Halo mini PC guide for local AI covers the 128GB machines you can buy today, and the mini PC for AI hub maps the whole category.

What Gorgon Halo actually changes (and what it doesn't)

Three deltas, side by side. This is the entire story of the generation:

One correction worth making before we go further, because several aggregator pages have it wrong: Gorgon Halo is Zen 5, not Zen 6, and the iGPU is the same RDNA 3.5 generation as Strix Halo. The Radeon 8065S is a rebadge with higher clocks, not a new architecture. Anyone expecting an RDNA 4 or UDNA iGPU here will be disappointed, and anyone expecting the ROCm maturity story to have changed will be disappointed twice.

Why bandwidth, not capacity, sets your tokens/sec

Here is the mental model that makes the whole purchase obvious. To generate one token, a dense transformer reads every single weight in the model. Not some of them. All of them. So the ceiling on tokens per second is simple division:

Now use the measured number rather than the marketing one. Independent testing published on llm-tracker.info (page dated 2025-05-17, measuring the 395, not the 495) recorded a peak of 212 GB/s on rocm_bandwidth_test against the 256 GB/s theoretical figure — about 83% efficiency, which is normal for LPDDR5X on a 256-bit bus. Scale that same efficiency onto Gorgon Halo's 273 GB/s and you get roughly 226 GB/s effective.

Run the division on a dense model that actually uses the new capacity. A dense model quantized to 4 bits that fills 150GB reads 150GB per token:

• 226 GB/s ÷ 150GB = ~1.5 tok/s. A sentence takes most of a minute. Nobody will use this twice.
• Fill 96GB instead (the 128GB machine's ceiling): ~2.2 tok/s. Still unusable.
• A 70B dense model at Q4, about 40GB: ~5 tok/s. Borderline — roughly half of comfortable reading speed.

That is the uncomfortable conclusion hiding under the 192GB headline: on dense models, the memory you are paying extra for is memory you cannot usefully fill. The 128GB part already has more capacity than its bandwidth can service. Adding 64GB more, with 6.6% more bandwidth to serve it, widens that gap rather than closing it.

If this framing is new to you, our guide to how much RAM local AI actually needs works through the sizing side, and the unified memory and VRAM entries cover why a shared pool behaves differently from a discrete card's dedicated memory. The VRAM requirements calculator will tell you what fits; this section is about what moves.

The 64GB is for MoE models and KV cache — here's the proof

There is one class of model where extra capacity translates directly into usable capability, and it is the reason Gorgon Halo exists at all: mixture of experts.

An MoE model stores many expert subnetworks but routes each token through only a few of them. All the weights must be resident in memory, but only the active subset is read per token. That breaks the tie between model size and generation speed — exactly the thing dense models cannot do.

The measured contrast on Strix Halo hardware is stark. All four rows below are llama.cpp results from the same llm-tracker.info test bench on a 128GB Ryzen AI Max+ 395:

Read the third row against the fourth. Llama 4 Scout occupies 58 GiB and generates at 20.23 tok/s. A dense model of identical footprint would manage 3–4 tok/s on the same box. Same hardware, same bandwidth, roughly 5–6x the throughput — purely because Scout reads about 17B active parameters per token instead of 109B. That single comparison is the entire economic case for 192GB.

The second lever is KV cache. Long contexts consume memory that has nothing to do with model weights, and unlike weights, cache is re-read selectively rather than in full. A 100B-class model with a 128K context window and several concurrent sessions can add tens of gigabytes of cache on top of the weights. If your work is long-document analysis or agents with large scratchpads, the extra 64GB is buying context headroom, and that is a real purchase.

Concretely, the models that justify it are the large sparse ones: see our hardware guides for DeepSeek-V4-Flash, Kimi K2.6, GLM-5.2 and Qwen 3.6, plus Llama 4 Maverick and Llama 4 Behemoth on the capacity end. The models that do not justify it are dense mid-size workhorses like Qwen 3 72B, DeepSeek-R1 70B and Gemma 3 27B — all of which sit comfortably inside 128GB and gain nothing from a bigger pool. Our guide to running large MoE models on small GPUs goes deeper on exploiting sparsity with less hardware.

The three Gorgon Halo machines you can actually buy (and their prices)

"Actually buy" is doing generous work in that heading. Three machines are announced; one has a price.

Minisforum MS-S1 MAX-P495

Announced at IFA 2026 on 2026-09-05 and the most interesting of the three for desk use. Per HotHardware's spec write-up: a 3.5-litre aerospace-grade aluminium chassis, pure copper cold plate, six heat pipes, dual turbine blowers and phase-change thermal interface material, sustaining 120W continuous and 160W peak. Connectivity is unusually serious for a mini PC — dual 10GbE, two USB4 at 40Gbps plus two USB4 V2 at 80Gbps, HDMI 2.1 FRL, a full-length PCIe 4.0 x16 slot, and dual M.2 PCIe 4.0 slots supporting up to 16TB with RAID 0/1. Up to 160GB of the 192GB pool can be allocated strictly as graphics memory.

Price: unannounced. Minisforum's own product page still reads "Price Reveal Coming Soon" and takes early-access signups rather than orders. Several outlets have circulated a figure around €7,000; that is reporting and estimation, not an announced price, so treat it as needs verification.

Framework Desktop 192GB

A mini-ITX board and chassis, listed on frame.work as a 192GB configuration "coming soon" at 273 GB/s — the source of the 6.6% bandwidth figure. No announced price and no announced date. The comparable 128GB Ryzen AI Max+ 395 configuration currently lists at $3,449, up from $1,999 at its February 2025 launch, which tells you most of what you need to know about LPDDR5X pricing in 2026 (our DRAM shortage guide covers why). Framework has signalled the 192GB part will cost meaningfully more. Figures above $4,500 in circulation are analyst expectations, not prices.

HP ZBook Ultra G3a 16

The first 192GB AI laptop, expected October 2026. HP's 2026-09-15 press release states the machine supports "up to 300B-parameter LLMs locally with up to 192GB unified memory and 160GB dedicated to the VRAM," and says pricing would be provided closer to availability. A $7,449 figure for the 192GB/512GB configuration subsequently surfaced in HP's own cost calculator and was reported by VideoCardz and TechRadar. Note that is a 512GB SSD and an FHD panel at that price.

So: one aspirational laptop price, and two boxes you cannot yet pay for. Here is the machine that does have a price, today, and that will match a 192GB box token-for-token on every model both can hold:

The GMKtec EVO-X2 is the 128GB Ryzen AI Max+ 395 box at $2,199–$3,649 depending on memory and storage — note that the 64GB to 128GB step alone is roughly $1,300, which is LPDDR5X pricing rather than a vendor markup, and is a preview of what 192GB will cost. Full write-up in our Ryzen AI Halo review.

Interrogating the "300 billion parameters" claim

HP's headline number deserves a careful answer rather than either applause or a dismissal, because it is half true in a way that matters.

The capacity claim is correct. A 300B-parameter model at 4-bit quantization needs roughly 150GB for weights. 160GB is allocatable as graphics memory. It fits, with about 10GB to spare. HP has not overstated the arithmetic.

The claim is silent about throughput, and that is where it misleads. Apply the formula from earlier: a dense 300B model reads all 150GB per token, and at ~226 GB/s effective that is about 1.5 tokens per second. You would wait roughly ten minutes for a 900-token answer. The model loads. It does not usefully run.

If the 300B model is sparse, the picture inverts. A 300B MoE with 25B active parameters reads about 12.5GB per token at Q4, which lands near 18 tok/s — genuinely interactive, and a capability no 128GB machine can offer at all because the full 150GB of weights must still be resident. Frontier-scale sparse models are exactly what a 160GB allocation is for.

Gorgon Halo vs DGX Spark vs RTX Spark vs Mac Studio M5

Unified-memory boxes only look interchangeable. They differ on the axis that decides everything: capacity per dollar, bandwidth per dollar, and ecosystem.

Two things in that table should stop you. First, Gorgon Halo's 273 GB/s is identical to the DGX Spark's — AMD has caught up to NVIDIA's GB10 on memory bandwidth while offering 50% more capacity. Second, an RTX 5090 has 6.6x the bandwidth of Gorgon Halo at a fraction of the price, and 1/6th the usable memory. That is the entire trade in one line: unified memory boxes buy capacity with bandwidth. See RTX 5090 vs RTX 3090 for how that looks inside the discrete-GPU tier.

Where DGX Spark ($4,699–$7,999) still wins decisively is prefill — the compute-bound phase where the model reads your prompt. On gpt-oss-120b, hardware-corner's comparison of llama.cpp results (2025-10-15) put DGX Spark at 1,723.07 tok/s prompt processing against Strix Halo's 339.87 — a ~5x gap — while token generation came in at 38.55 vs 34.13 tok/s, a 13% difference. That asymmetry is the whole story: generation is bandwidth-bound and nearly tied; prefill is compute-bound and NVIDIA wins it outright. If you paste 50-page documents into a model all day, prefill is your bottleneck and Gorgon Halo is the wrong box. Full breakdown in DGX Spark vs Strix Halo.

The consumer RTX Spark ships fall 2026 from a rumoured ~$2,899 with up to 128GB, which will make the CUDA path considerably cheaper than DGX Spark. It is not orderable yet, so it belongs in your "wait" column rather than your cart.

And if bandwidth is what you actually need, Apple sells it. The Mac Studio M5 Ultra (From $5,499) runs 512GB at 1.2 TB/s — 4.4x Gorgon Halo's bandwidth and 2.7x its capacity. It costs accordingly, and note Apple states the 512GB configuration ships late October 2026. There is no CUDA, but MLX is a mature path. See the Apple Silicon for AI hub and our Strix Halo vs Mac Studio head-to-head for the previous-generation version of this comparison.

Should you wait for 192GB, or buy 128GB today?

Four branches. Find yours.

Buy 128GB now

If your models are dense, or MoE that fits in about 90GB at Q4, or you want the machine this quarter. You will lose nothing in tokens per second, because on any model both machines can hold, a 6.6% bandwidth delta is invisible. The GMKtec EVO-X2 at $2,199–$3,649 is the pick; Framework's 128GB Desktop at $3,449 is the modular alternative. Our best mini PC for AI roundup covers the rest of the field.

If you would rather spend the same money on bandwidth than capacity, the Mac Studio M5 Max (From $2,499) holds the same 128GB at 614 GB/s — roughly 2.2x the tokens per second on anything that fits both. The Mac Mini M5 Pro (From $1,699) is the cheaper 64GB entry point if 70B-class is your ceiling.

Wait for 192GB

If you specifically run 200B+ sparse models, or need very long context on something above 100B, and you can tolerate an unannounced price and date. The capability is real and no 128GB box can substitute for it. Be clear-eyed about the cost signal though: the EVO-X2's 64GB-to-128GB step is already ~$1,300, and Framework's 128GB configuration went from $1,999 to $3,449 on memory pricing alone.

Buy neither

If prefill throughput or CUDA compatibility is your constraint, unified memory is the wrong category entirely. Go DGX Spark for CUDA in a small box, or a discrete GPU for raw speed: an RTX 5090 ($1,999–$2,199) at 1,792 GB/s, a used RTX 3090 ($699–$999) for cheap 24GB, or an RTX 5060 Ti 16GB ($429–$479) as the budget entry. See the AI GPU buying guide and RTX 5090 vs RTX 5060 Ti. And if your real workload is small models and always-on agents rather than 100B-class inference, a GMKtec M8 ($389–$459), GMKtec M6 Ultra ($429–$549) or MAGICNUC AS1 ($229–$299) does that job for a tenth of the money — see the AI on a budget hub.

The two-box question

HotHardware's MS-S1 write-up relays a Minisforum claim that, in their words, "a dual-unit 192GB cluster can run a massive 235B parameter model at a respectable 16 tokens per second." That is a vendor claim relayed by a publication, not an independent measurement, and we would not spend money on it unverified. Clustering unified-memory boxes over Ethernet generally costs you more in interconnect latency than it returns in capacity — our Mac mini cluster guide shows how that actually behaves in practice, and NVIDIA's pairing approach covers the higher-bandwidth alternative.

Power, heat, and what it costs to run

This is Gorgon Halo's most underwritten advantage. The MS-S1 sustains 120W with 160W peaks for the entire system. A discrete-GPU build with an RTX 5090 draws 575W for the card alone and wants a 1000W+ PSU.

At a US average around $0.17/kWh, an always-on 140W box costs roughly $17/month. The same duty cycle on a 500W build is closer to $60/month — about $500/year of difference, which is a meaningful fraction of the purchase-price gap over three years. It also means no dedicated circuit, no PSU upgrade, and a machine quiet enough to sit on the desk instead of in a closet. Our local AI electricity cost guide has the full model, and quiet AI PC builds covers the acoustics side. For a small team running one shared box, that operating profile matters more than peak throughput — see setting up a private AI server for a team.

Software: ROCm, Vulkan, and the llama.cpp reality on RDNA 3.5

Same iGPU architecture means the same software maturity story, so set expectations from the 395's measured behaviour rather than from AMD's roadmap.

On Strix Halo, llm-tracker.info's testing found Vulkan beat HIP/ROCm on prompt processing by 2.6x — 884.20 tok/s versus 343.91 on Llama 2 7B Q4_0 with Flash Attention — while token generation was nearly identical (52.73 vs 50.88). More pointedly, the HIP backend hung outright on Llama 4 Scout where Vulkan completed the run. On a 40-CU RDNA 3.5 iGPU in 2026, llama.cpp with the Vulkan backend is the default you should assume, not the fallback.

Practical notes for either generation:

• Use GGUF quants and llama.cpp/Ollama first. vLLM on RDNA 3.5 is possible but not the path of least resistance.
• Enable Flash Attention. It materially reduces KV cache footprint, which is most of why you bought the capacity.
• Q4 is the sweet spot. Below Q4 quality degrades faster than memory savings justify — see our quantization primer.
• Set the graphics memory allocation deliberately. The 160GB ceiling is a setting, not a default. Leave the OS about 32GB.
• Speculative decoding is the one free lunch here. It trades spare compute for generation speed, and a 40-CU iGPU starved of bandwidth is exactly the compute-rich, bandwidth-poor profile it helps.

One honest caveat on every performance number in this article: no independent LLM benchmarks of the 495 have been published as of 2026-09-27. Every Gorgon Halo figure above is either a published specification or our own projection from measured 395 results scaled by the 6.6% bandwidth delta, and we have labelled which is which. Review units are with reviewers; when real numbers land, this guide gets updated.

The bottom line

Gorgon Halo is a capacity upgrade wearing a performance headline. 192GB is 50% more memory served by 6.6% more bandwidth and exactly the same 40 compute units, which means it loads models the 128GB part cannot load and generates tokens at the same speed on everything both can hold.

That makes the decision unusually clean. Look up the active parameter count of the models you actually run. If total parameters far exceed active parameters, or you need very long context above 100B, 192GB is a real upgrade and worth waiting and paying for. If they are dense, or sparse-but-small, the correct purchase is a 128GB box today — the GMKtec EVO-X2 at $2,199–$3,649 — and the 64GB you did not buy would have sat empty because nothing could stream it fast enough to matter.

The mistake to avoid is not buying too little memory. It is buying memory on the assumption that more of it makes inference faster. It never has.