ASUS has built its first Intel Arc graphics card, and the company skipped the gaming market entirely. The new ASUS UGen Intel Arc Pro B70 is a 32GB workstation accelerator aimed squarely at local AI inference, not frame rates, according to reporting from BigGo Finance and other outlets on October 17 and 18, 2026. The card arrives under a brand-new ASUS label, UGen AI, and it signals that the GPU maker sees more near-term money in on-premises AI boxes than in another round of the Radeon-versus-GeForce fight.
The timing matters. Intel has sold Arc Pro B-series workstation cards since earlier this year, but ASUS staying out of the Arc ecosystem through two prior generations made its sudden entry notable. Multiple outlets, including TechPowerUp and WCCFTech, framed the move as ASUS placing a bet on cheaper, higher-capacity local AI hardware at a moment when GPU and memory prices are both climbing fast across the industry.
What ASUS Just Announced
The product is called the ASUS UGen Arc Pro B70, built around Intel’s BMG-G31 die, the full “Big Battlemage” silicon that anchors Intel’s workstation GPU lineup. The card ships with 32GB of GDDR6 ECC memory across a 256-bit bus, delivering 608GB/s of memory bandwidth. Inside the die sit 32 Xe2 cores paired with 256 XMX matrix engines, the units Intel relies on for AI acceleration work. ASUS rates the boost clock at up to 2.8GHz.
Power draw lands at 310W total board power, run through a dual-slot blower cooler, a design choice that favors dense rackmount and workstation chassis over the open-air coolers found on gaming cards, according to specs reported by Hardware Busters. Display output is modest for a card this expensive: one HDMI 2.1b port plus three DisplayPort 2.1 connections, enough for a multi-monitor workstation setup but nothing built for a living room.
On paper, Intel’s own spec sheet puts the Arc Pro B70 at up to 367 TOPS of INT8 inference performance, a figure ASUS carries over in its own marketing. That number is a theoretical ceiling tied to the silicon’s matrix engines, not an independently verified benchmark, and buyers evaluating the card for real workloads should treat it as a ceiling rather than a guarantee.
Why 32GB of VRAM Is the Whole Pitch
VRAM capacity is the entire reason this card exists. Large language models and other generative AI workloads are memory-bound long before they’re compute-bound: a model with billions of parameters needs enough VRAM just to load before a single token gets generated. Consumer gaming GPUs, even high-end ones, routinely cap out at 16GB or 24GB, which rules out running larger open-weight models locally without quantizing them down to the point where quality suffers.
ASUS is leaning into that gap. The company says up to four Arc Pro B70 cards can be combined in a single system, pooling to 128GB of aggregate VRAM, enough headroom, in ASUS’s own framing, to run AI models in the neighborhood of 120 billion parameters locally. That multi-card path matters because it undercuts the usual alternative for 128GB of GPU memory: a single flagship accelerator costing many times more than four B70 cards combined, or a specialized workstation built around one vendor’s high-end silicon.
ASUS has already shown how this plays out in a finished product. At IBC 2026, the company’s ProArt division paired the UGen Intel Arc Pro B70 32GB accelerator with its ExpertCenter Pro ET500I W8 system, pitching the combination as an on-premises AI inference box for enterprise and professional buyers who want generative AI running inside their own four walls rather than routed through a cloud API. That pitch, local deployment with no data leaving the building, is the single biggest differentiator ASUS has over cloud-rental alternatives, and it is becoming a more common sales angle as businesses grow wary of sending proprietary data to third-party inference endpoints.
Software and Framework Support
Intel’s Arc Pro B-series has spent 2026 building out software compatibility, and the B70 inherits that work. ASUS lists support for vLLM through Intel’s XPU backend, llama.cpp via SYCL and Vulkan, Ollama, PyTorch through Intel’s IPEX extension, OpenVINO, oneAPI, and Vulkan itself. That is a meaningfully broader toolchain than Intel’s Arc GPUs supported even a year ago, when developers routinely complained about driver gaps and missing kernel support for anything outside Intel’s own OpenVINO stack.
ASUS has published early performance numbers tied to that stack: 369 tokens per second running vLLM with 50 simultaneous requests, and 188.2 tokens per second on Qwen3 8B. Those figures come from ASUS’s own benchmarking rather than a third-party lab, so they should be read as a vendor’s best-case showing rather than an independently reproduced result. Still, the fact that ASUS is quoting throughput on a real, named open-weight model, Qwen3 8B, rather than only citing abstract TOPS figures, suggests Intel’s software story for Arc Pro has matured enough to support serious benchmarking.
Pricing Is the Open Question
ASUS has not confirmed a price or a shipping date for the UGen Arc Pro B70. That gap is the biggest unknown hanging over the launch. For context, Intel’s own reference Arc Pro B70 launched earlier this year at $949, though US retail listings for that reference design had already drifted closer to $1,300 by the time ASUS’s card surfaced. One European price-comparison listing showed the ASUS-branded card priced around €1,749, but that figure should not be treated as a confirmed ASUS MSRP since it reflects a single retailer listing rather than an official announcement.
That price spread, roughly $949 at reference-design launch versus potential retail pricing well north of $1,700 for a custom board, tells its own story about where GPU pricing sits right now, a gap TweakTown also flagged in its own coverage of the launch. Memory prices have been climbing across the board in 2026, and a 32GB ECC GDDR6 card sits directly in the crosshairs of that trend. Whatever ASUS eventually charges, the card will be competing against that backdrop of rising component costs rather than the relatively calmer pricing environment Intel launched into earlier in the year.
| Workstation GPU | VRAM | Rated AI TOPS (INT8) | Approx. Price |
|---|---|---|---|
| ASUS UGen Arc Pro B70 | 32GB GDDR6 ECC | 367 | Unconfirmed (reference: $949-$1,300) |
| Intel reference Arc Pro B70 | 32GB GDDR6 ECC | 367 | $949 at launch |
| ASUS ProArt Nvidia Blackwell workstation card | 32GB GDDR7 | 3,352 (claimed) | Premium tier, not disclosed |
| ASUS Arc Pro B70 (4-card pool) | 128GB combined | Up to ~1,468 (4x single card) | Unconfirmed, scales with per-card price |
How the B70 Stacks Up Against AMD and Nvidia
The Arc Pro B70 is not trying to out-muscle Nvidia’s top workstation silicon, and ASUS’s own marketing materials make that distinction explicit. The company’s Nvidia Blackwell-based professional offerings carry 32GB of GDDR7 and a claimed 3,352 AI TOPS figure, which dwarfs the B70’s 367 INT8 TOPS rating on paper. Nvidia remains the performance ceiling for premium local AI work, and nothing about the B70 changes that.
What the B70 offers instead is an accessible middle tier: a workstation-grade card with a full 32GB of memory at a price point that, even allowing for the current uncertainty, likely lands well under what equivalent Nvidia RTX Pro or AMD Radeon Pro cards with similar memory footprints command. That positioning echoes an argument shattered.io covered in its breakdown of RTX Pro economics, where buyers increasingly care less about peak throughput and more about dollars spent per gigabyte of usable VRAM.
| Spec | ASUS UGen Arc Pro B70 | Intel Reference Arc Pro B70 |
|---|---|---|
| GPU die | Intel BMG-G31 (Battlemage) | Intel BMG-G31 (Battlemage) |
| VRAM | 32GB GDDR6 ECC | 32GB GDDR6 ECC |
| Memory bus | 256-bit | 256-bit |
| Memory bandwidth | 608GB/s | 608GB/s |
| Xe2 cores / XMX engines | 32 / 256 | 32 / 256 |
| Boost clock | Up to 2.8GHz | Not specified by ASUS |
| AI performance (INT8) | Up to 367 TOPS | Up to 367 TOPS |
| Board power | 310W | Lower, per Intel reference spec |
| Cooling | Dual-slot blower | Blower-style reference cooler |
| Display outputs | 3x DisplayPort 2.1, 1x HDMI 2.1b | Varies by board partner |
| Launch price | Unconfirmed | $949 at launch |
Multi-GPU Local AI: The 128GB Pitch
Four-card configurations are where ASUS wants buyers to pay attention. Pooling four Arc Pro B70 cards into one system yields 128GB of combined VRAM, a figure that places it in the same conversation as dedicated AI workstation products that shattered.io has tracked through 2026, including Nvidia’s own DGX Spark split between 64GB and 128GB configurations. The difference is architectural flexibility: rather than buying into one fixed memory tier from a single vendor, a buyer scaling an Arc Pro B70 system can add cards incrementally as budget allows, moving from one card’s 32GB up toward the four-card ceiling over time.
That scalability is also where the comparison to local LLM setups built around Nvidia’s consumer flagship becomes relevant. shattered.io’s own walkthrough on running LLMs locally on an RTX 5090 highlights just how much VRAM constrains what a single-card setup can realistically host. A 32GB Arc Pro B70 already clears the ceiling most consumer cards hit, and four of them stacked together clear it by a wide margin, assuming software support keeps pace across all four cards simultaneously, which remains an open question given how recent Intel’s Arc Pro multi-GPU tooling is.
Intel’s Arc Pro B-Series in Context
The B70 sits at the top of Intel’s Arc Pro B-series workstation lineup, a family Intel has marketed around three pillars: larger AI models, higher-precision professional workloads, and content creation tasks that benefit from extra VRAM headroom. Intel’s own quick-reference documentation lists the B70 at 32GB of memory, a 256-bit bus, 608GB/s of bandwidth, and 367 INT8 TOPS, figures that match what ASUS is now shipping on its own board. Intel’s documented power range for the series spans 160W to 290W depending on configuration, with ASUS’s custom board pushing to 310W, presumably to support the higher clock speeds the company is targeting.
Intel’s broader Arc graphics push has been a multi-year project with mixed results on the gaming side, where Arc GPUs have struggled to carve out meaningful market share against Nvidia and AMD. The Arc Pro workstation line, by contrast, has found more oxygen by avoiding a head-to-head fight over gaming frame rates and instead chasing a market, local AI inference, that is expanding quickly and where capacity, not raw gaming benchmarks, is the deciding factor for buyers. shattered.io has tracked that broader shift in its look at the smaller players fighting for share outside Nvidia and AMD, and the ASUS Arc Pro B70 is a direct extension of that trend: a case of an Intel silicon design finding a partner willing to put real engineering and marketing behind it, not for gamers, but for IT buyers.
ASUS’s Broader AI Hardware Strategy
ASUS launching an Intel-based AI accelerator is notable given how deeply the company has invested in Nvidia-based AI hardware over the past two years. ASUS’s ProArt division has already shipped Nvidia-based workstation GPUs rated at up to 3,352 AI TOPS, and the company has been one of the most aggressive partners pushing Nvidia’s own AI PC hardware into retail, a push shattered.io covered when Acer’s Predator Atlas 8 built around Intel’s Arc G3 architecture showed Intel graphics gaining real traction in a different corner of the PC market: integrated gaming silicon rather than discrete workstation cards.
The UGen brand itself is new territory for ASUS, separate from its established ROG, TUF, and ProArt lines. Branding an Intel-based AI accelerator under a fresh name rather than folding it into ProArt, where the Nvidia-based workstation cards already live, suggests ASUS wants a clear line between its Nvidia AI hardware and its new Intel-based AI hardware, possibly to avoid confusing buyers who associate ProArt specifically with Nvidia silicon, or possibly to give the Intel line room to grow independently if Intel’s software ecosystem keeps maturing.
The Local AI Inference Market ASUS Is Chasing
Local AI inference, running a model on hardware you own rather than renting GPU time in the cloud, has grown from a hobbyist niche into a real purchasing category over the past 18 months. Enterprises running sensitive data through generative AI models increasingly want that processing to happen on-premises, both for latency reasons and for data governance concerns that cloud APIs can’t fully satisfy. That demand has pulled workstation GPU makers, including Intel, AMD, and Nvidia, into a race to offer the most VRAM per dollar, since memory capacity is the single biggest bottleneck standing between a buyer and running a meaningfully large model locally.
Intel’s Arc Pro B70 entering that race through ASUS, rather than through a reference-design-only launch, gives Intel something it has historically lacked in the GPU space: a major board partner willing to build custom silicon around its chips and sell it through retail and enterprise channels simultaneously. ASUS’s ExpertCenter Pro pairing at IBC 2026 demonstrates that Intel’s workstation push now has at least one partner building complete, sellable systems rather than leaving integrators to pair bare cards with generic chassis themselves.
What’s Still Unconfirmed
Several pieces of this story remain open. ASUS’s exact MSRP and release date have not been announced. Independent, third-party benchmarking of the card against real workloads, rather than ASUS’s own vLLM and Qwen3 8B figures, has not yet happened. Driver maturity across four-card configurations is untested in public reporting. And direct, apples-to-apples comparisons against equivalent AMD Radeon Pro and Nvidia RTX Pro cards with similar VRAM footprints have not been published by any outlet covering the launch.
Those gaps matter because workstation buyers, unlike gaming buyers, tend to make purchasing decisions based on total cost of ownership across a fleet of machines rather than a single card’s specs. Until ASUS confirms pricing and independent reviewers get hands-on time, IT buyers evaluating the B70 against established Nvidia and AMD options are working with an incomplete picture.
Market Impact and Competitive Pressure
The launch lands at a moment when GPU and memory pricing across the industry is under pressure, a dynamic shattered.io has tracked closely in coverage of the ongoing memory chip shortage pushing PC prices higher. A 32GB ECC GDDR6 card entering that market gives ASUS a product that, if priced competitively, could pull budget-constrained AI buyers away from pricier Nvidia options, particularly smaller businesses and research labs that need real VRAM capacity but can’t justify flagship Nvidia workstation pricing.
For Intel, the stakes are different but just as real. Intel’s Arc graphics division has spent years trying to prove it belongs in a market Nvidia and AMD have dominated for decades. A major board partner like ASUS building custom silicon around Intel’s workstation chips, and marketing it specifically around a growing use case like local AI inference, is a credibility signal Intel has needed. If the B70 sells well under the ASUS brand, it could pull in more board partners for future Arc Pro generations, the kind of ecosystem effect that took Nvidia and AMD decades to build and that Intel has struggled to replicate.
Historical Context: Intel’s Long Road in Discrete Graphics
Intel’s modern discrete GPU effort dates back to the first Arc Alchemist cards, which launched facing skepticism about driver quality and gaming performance relative to established Nvidia and AMD options. Battlemage, the architecture underpinning the Arc Pro B70, represented Intel’s second real swing at the discrete GPU market, and its workstation variants have generally earned more positive reception than Intel’s consumer gaming cards because workstation buyers prioritize memory capacity and software stack compatibility over raw frame rates.
ASUS’s absence from Intel’s GPU ecosystem through two prior Arc generations was itself notable, given that ASUS has historically been one of the fastest board partners to adopt new Nvidia and AMD silicon. That ASUS waited for the workstation-focused Battlemage generation, rather than jumping in during the earlier Alchemist generation, suggests the company saw the AI inference angle specifically, not gaming, as the opening worth building a new product line around.
Predictions: Where This Goes Next
Several outcomes look plausible based on how this launch has unfolded and how similar workstation GPU launches have played out in 2026:
- ASUS will likely confirm pricing within weeks rather than months, given that the card has already appeared on European price-comparison sites like Idealo, suggesting retail channels are already stocking units ahead of an official announcement.
- Expect independent reviewers to focus heavily on the four-card, 128GB configuration rather than the single-card setup, since that multi-GPU scaling story is the card’s clearest differentiator against single-card Nvidia and AMD alternatives.
- Other ASUS product lines, particularly ProArt, may eventually fold in Arc Pro options if the UGen brand gains traction, rather than keeping Intel and Nvidia AI hardware permanently siloed under separate names.
- Pricing pressure from the ongoing memory shortage will likely push ASUS toward the higher end of plausible pricing, closer to the €1,749 European listing than Intel’s original $949 reference price.
- If the B70 performs well in independent reviews, expect at least one other major board partner, likely Gigabyte or MSI, to announce a competing Arc Pro B70 variant within the next two quarters, following the same pattern Intel’s Arc gaming cards eventually saw once early skepticism faded.
Frequently Asked Questions
What is the ASUS UGen Arc Pro B70?
It is ASUS’s first graphics card built on Intel Arc silicon, a 32GB workstation GPU based on Intel’s BMG-G31 Battlemage die, designed for local AI inference and professional creation workloads rather than gaming.
How much does the ASUS Arc Pro B70 cost?
ASUS has not announced an official price. Intel’s reference Arc Pro B70 launched at $949 earlier in 2026, with US retail listings later climbing closer to $1,300. A European listing showed the ASUS card at roughly €1,749, though that figure is not a confirmed ASUS MSRP.
How much VRAM does the Arc Pro B70 have?
The card carries 32GB of GDDR6 ECC memory on a 256-bit bus, with 608GB/s of memory bandwidth. Four cards can be combined in one system for 128GB of total VRAM.
Can the Arc Pro B70 run large AI models locally?
ASUS positions a single card as suitable for mid-sized models, and a four-card, 128GB configuration as capable of running models in the range of roughly 120 billion parameters, according to the company’s own positioning. Independent third-party verification of that claim has not yet been published.
Is the ASUS Arc Pro B70 a gaming GPU?
No. It is marketed exclusively as a workstation and AI inference card, with a dual-slot blower cooler, limited display outputs, and software support centered on AI frameworks like vLLM, llama.cpp, Ollama, PyTorch, and OpenVINO rather than gaming drivers.
How does the Arc Pro B70 compare to Nvidia and AMD workstation GPUs?
It trails Nvidia’s top Blackwell-based workstation cards on raw AI throughput, with Intel’s 367 INT8 TOPS rating well below Nvidia’s claimed 3,352 AI TOPS on comparable 32GB Nvidia hardware. The B70’s pitch is price-per-gigabyte of VRAM rather than peak performance.
What software does the Arc Pro B70 support?
ASUS lists support for vLLM via Intel’s XPU backend, llama.cpp through SYCL and Vulkan, Ollama, PyTorch via Intel’s IPEX extension, OpenVINO, oneAPI, and Vulkan.
When will the ASUS Arc Pro B70 ship?
No official release date has been confirmed as of October 18, 2026. The card’s appearance on European retail price-comparison sites suggests availability may follow soon, but ASUS has not announced a firm date.