Ryzen AI Max+ 495 vs 395: Is the Premium Worth It?

Data freshness: 27 September 2026. Independent benchmarks for the 495 do not exist yet — StorageReview is testing a pre-production unit and says scores wait for October shipping hardware. Every price below is quoted from a named source with its date. Vendor performance claims are labeled as claims. Anything we could not source is marked unknown rather than estimated.

1. 30-second answer

The Ryzen AI Max+ 495 is a capacity SKU, not a speed SKU, and at HP's own prices the premium is only worth it when a model physically will not fit in 128GB. HP's Local AI Value Calculator puts the ZBook Ultra G3a 16 with the 495 and 192GB at $7,449, the same processor with 128GB at $5,999, and a Ryzen AI Max PRO 490 with 64GB at $3,899 (VideoCardz, 2026-09-22). So inside one SKU, 64GB more unified memory costs $1,450 — about $22.66 per extra GB. Meanwhile a 395 desktop with 128GB, the GMKtec EVO-X2, is $2,199 in our database. You are not buying 2x the tokens per second; both generations run on a bandwidth-bound unified-memory platform, which is why VideoCardz, on the same HP pricing report, argues a multi-GPU desktop can beat it on inference speed at similar money.

2. What the 495 actually is

The 495 sits at the top of AMD's Strix Halo PRO stack. Per StorageReview's pre-production lab writeup of the ZBook Ultra G3a 16 (2026-09-15), the 495 configuration pairs 16 cores, 32 threads, a 5.2GHz boost clock, 192GB of LPDDR5X-8533 soldered around the CPU, up to 160GB assignable to the integrated Radeon 8065S, and an NPU rated at 55 TOPS. HP says it co-engineered the system with AMD.

Two claims in that writeup are HP's, not measured: that the configuration runs LLMs up to 300B parameters locally, and that it is the most capable local agentic laptop — a ranking HP derives from its own August comparison of GPU memory capacity across competing laptops. StorageReview notes the same caveat HP does: actual model support and performance depend on the software and workload. The one real throughput number in the article — Gemma 3 27B at 8.96 tokens per second — is from the older ZBook Ultra G1a 14 with 128GB, not from the 495.

The generational contrast, using our own database rows as the 395 side: the 395 generation in our catalog tops out at 128GB of LPDDR5X-8000 on a 256-bit bus at 256 GB/s (GMKtec EVO-X2 and Minisforum MS-S1 MAX spec rows), with a 50 TOPS NPU. The 495 family in our catalog is 192GB parts — LPDDR5x-8533, and on the Chuwi UniBox AI495 Pro, 273 GB/s on the same 256-bit bus. So the real delta is capacity: 192GB versus 128GB of addressable unified memory, plus +5 TOPS of NPU. Per-GPU benchmark deltas between the two SKUs: not yet published.

3. The price math, from HP's own calculator

VideoCardz reported on 2026-09-22 that HP quietly exposed ZBook Ultra G3a 16 pricing through the company's Local AI Value Calculator. These are HP's listed figures, quoted as published — not street prices, and not a configuration that includes the high-end displays (HP's reference $7,449 build ships the FHD panel, not the OLED or DreamColor options).

Three things fall out of that table. First, the 64GB step inside the 495 SKU is $1,450 — $22.66 per added GB — while the processor step from 490 to 495 at 128GB versus 64GB is $2,100 for 64GB, or $32.81 per GB. Second, price per GB falls as capacity rises on the same SKU, so the largest configuration is the efficient one. Third, none of these figures is a measured performance number: buying the 192GB config buys 64GB of headroom, and on a bandwidth-bound platform headroom and speed are not the same axis.

4. Why the SKU upgrade is not a speed upgrade

The reason to hesitate on the premium is structural, and it is the same reason large unified-memory laptops are not inference speed machines. VideoCardz, on the HP pricing story, states plainly that Strix Halo and Gorgon Halo are not the fastest platforms for local AI workloads, largely because of memory bandwidth, and that at around $7,000 it may be cheaper to build a desktop workstation with multiple GPUs and get much higher inference performance. When tokens per second are bound by how fast the shared memory pool can feed the GPU, doubling cores or buying a higher-binned SKU does not double throughput — only a wider or faster memory path does, and here the bus width is the same 256-bit across the family.

So the buyer's real question is not "495 or 395" in the abstract. It is: is my workload capacity-bound or bandwidth-bound? If you are serving a 70B model at interactive speed, you are bandwidth-bound and the 395 already delivers that — a 70B at 4-bit quantization fits comfortably in 128GB. If you are trying to hold a 200B-plus model, a large agent stack, or several simultaneous models resident in VRAM-equivalent unified memory, you are capacity-bound and the 192GB pool is the only thing that works. HP's own calculator prices the third option for you: 490 at 64GB is $3,899, which is the honest floor if your models are small.

5. The 495 and 395 machines in our database

Rows below come from the CompareAIHardware products database as of 2026-09-27. Price column is the MSRP or launch figure on file — only HP's ZBook calculator entries are HP-quoted prices, and none of these is a verified street price. Status is our catalog state: available means on sale, announced means vendor page exists but we have no shipping-price row.

Read that table as the actual price ladder: 495-class desktops with 192GB are the announced tier and mostly have no shipping price on file yet, while the 395 128GB tier already has four available machines between $2,199 and $4,999 MSRP. If your workload fits 128GB, the cheaper end of this table is where the value is. The 192GB entries are the ones worth waiting on for street prices.

Spec gaps we did not fill, deliberately: no bandwidth figure for the ACEMAGIC F9A 495 or the GMKtec EVO-X5 Pro, no TDP for the ACEMAGIC or GMKtec 495 rows, and no price at all for the announced units. Our database has no benchmark rows for any 495 machine, so nothing on this page implies a measured 495-versus-395 speed ratio.

6. Verdict by buyer type

• Running 70B-class models, agents, or a fine-tune that fits in 128GB? Buy the 395. GMKtec EVO-X2 at $2,199 MSRP is the cheapest published entry, Minisforum MS-S1 MAX at $2,399 adds dual 10GbE, and the 256 GB/s memory path is the same one the 495 uses. The 495 would spend extra money for capacity you cannot use.
• Holding a 200B+ model, or several large agents resident at once? Buy the 495 at 192GB. That is the one constraint the 395 cannot satisfy at any price. On HP's calculator the step from 128GB to 192GB is $1,450, and the 192GB config is the cheapest per GB of the three.
• Chasing inference speed, not capacity? Buy neither yet. Per VideoCardz on the same HP story, memory bandwidth makes Strix Halo and Gorgon Halo weak throughput platforms, and a multi-GPU desktop at comparable money can deliver higher inference performance. Wait for October's independent numbers before paying a premium for the top bin.
• Buying a laptop because you need the machine to be portable? The 495 ZBook is the only route to 160GB GPU-assignable unified memory in a thin chassis, and HP's own reference build at $7,449 ships an FHD panel. If you want OLED or DreamColor, your real price is above the calculator figure.
• Cheapest usable box, period? The HP calculator's 490 + 64GB at $3,899 is the honest floor for a 64GB mobile pool, and the $2,199 EVO-X2 undercuts it for 128GB desktop use. Neither is a fast inference machine; both are capacity playthings relative to a multi-GPU tower.

7. Sources (all URLs verified 200 on 2026-09-27)

8. Data freshness

27 September 2026. This page is deliberately a price-and-capacity analysis, not a benchmark comparison, because no independent 495 benchmarks exist: StorageReview's unit carries a pre-production label and its article says benchmarks wait for shipping hardware, with HP expecting October availability. The only measured throughput number we cite (8.96 tokens per second) is from the previous-generation ZBook Ultra G1a 14 and is labeled as such. When October's independent scores land — 495 versus 395 on the same 256-bit bus would settle whether the top bin is worth anything for speed — the verdict sections get rewritten first. Vendor performance claims on this page (300B parameters, "most capable local agentic laptop") are HP's, flagged where they appear, and our database has no benchmark rows for any 495 machine, so nothing here should be read as a measured 495-versus-395 performance ratio.

Frequently asked questions

Is the Ryzen AI Max+ 495 worth it over the 395?

Usually not, on price. The 495 is a capacity-and-headroom SKU, not a throughput upgrade you are paying twice for: on the same HP ZBook Ultra G3a 16, HP's Local AI Value Calculator prices the 495 with 192GB at $7,449 and the 495 with 128GB at $5,999, so the memory step inside one SKU is $1,450 for 64GB more unified memory (about $22.66 per extra GB). Both configurations share the same bandwidth-bound platform, so capacity is what you buy, not speed. Buy the 495 when a model or workload physically does not fit in 128GB — HP claims up to 300B-parameter local models on the 192GB configuration. Otherwise a 395 machine with 128GB (GMKtec EVO-X2 at $2,199 in our database) buys the same class of local-AI capability for a third of the money.

What exactly does the 495 add over the 395?

More unified memory capacity and a higher top-bin position, not more compute. AMD's 495 is the top of the Strix Halo PRO stack; StorageReview's pre-production ZBook Ultra G3a 16 pairs it with 16 cores, 32 threads, a 5.2GHz boost clock, 192GB of LPDDR5X-8533 soldered around the CPU, up to 160GB assignable to the integrated Radeon 8065S, and a 55 TOPS NPU. The 395 generation in our database tops out at 128GB of LPDDR5X-8000 on a 256-bit bus at 256 GB/s. Practical delta: 192GB versus 128GB of addressable unified memory, and +5 TOPS of NPU. Per-GPU benchmark deltas are not published yet (see the freshness section).

How much does 64GB more unified memory actually cost?

HP prices it directly in the ZBook Ultra G3a 16 calculator: 495 + 192GB = $7,449 versus 495 + 128GB = $5,999, a $1,450 difference for 64GB, or roughly $22.66 per GB of extra unified memory. HP also lists a Ryzen AI Max PRO 490 with 64GB at $3,899. Comparing the two ends of that calculator: 192GB at $7,449 works out to about $38.80 per GB, while 128GB at $5,999 is about $46.87 per GB — the bigger memory pool is cheaper per gigabyte, so capacity is where the platform gets more efficient.

Is a 495 laptop worth it versus a multi-GPU desktop workstation?

Probably not if inference throughput is the goal. VideoCardz, reporting HP's pricing, notes that Strix Halo and Gorgon Halo are not the fastest platforms for local AI largely because of memory bandwidth, and that at around $7,000 a multi-GPU desktop can be cheaper while delivering much higher inference performance. The 495 laptop wins on portability, silence, and a single 192GB pool; it loses on tokens per second per dollar. Videocardz also flags the desktop alternative in the same report.

Are there independent benchmarks for the 495 yet?

No production-hardware numbers exist yet. StorageReview has a pre-production HP ZBook Ultra G3a 16 in the lab and states that benchmarks wait for shipping hardware, with HP expecting to ship in October. Its published Gemma 3 27B result of 8.96 tokens per second belongs to the older ZBook Ultra G1a 14 with 128GB, not the 495. Our products database has no benchmark rows for any 495 machine, so every performance figure on this page is a vendor claim or a different product, and independent scores are the next thing to watch.

Which 495 and 395 machines are in your database?

495 family: ACEMAGIC F9A (192GB LPDDR5x-8533, $5,999 MSRP), Minisforum MS-S1 MAX P495 (192GB, announced), GMKtec EVO-X5 Pro (192GB, announced), Chuwi UniBox AI495 Pro (192GB LPDDR5x-8533 on a 256-bit bus, 273 GB/s, announced). 395 family: GMKtec EVO-X2 (128GB, $2,199 MSRP), Minisforum MS-S1 MAX (128GB, $2,399 MSRP), ACEMAGIC F9A 395 (128GB, $4,999 MSRP), Chuwi UniBox AI395 (128GB, announced), BOSGAME M5 (128GB, $3,099 MSRP). Spec rows and MSRP values come from the CompareAIHardware products database; machine-level prices other than HP's ZBook calculator and the vendor pages we could verify are MSRP or launch figures, not street prices.

Related: AI mini PC index, best laptops for AI, best laptop for LLMs, best laptop for Stable Diffusion, MacBook Pro vs PC for AI, DLSS 5 GPU support matrix, all guides.