GUNNIR Packs Two Arc Pro B70 Cards Into One 64GB Workstation [2026]

By Indie Kings | October 2, 2026

Updated October 2, 2026: GUNNIR showed a workstation at Intel Connection 2026 that packs two Intel Arc Pro B70 cards into one liquid cooled chassis, per VideoCardz by WhyCry reporting on September 27, 2026. Each B70 carries 32GB of GDDR6 ECC memory, so the dual GPU system offers 64GB of combined VRAM alongside Intel Core Ultra 200HX Arrow Lake-HX processors.

Intel processor art

Image: Intel processor art. Credit: Intel.

What GUNNIR showed at Intel Connection 2026

GUNNIR demonstrated workstation systems built around Intel Arc Pro B70 graphics cards at Intel Connection 2026 in China, per VideoCardz by WhyCry on September 27, 2026. The headline machine is the larger of the two, and it uses two B70 cards for a stated 64GB of combined VRAM. A smaller workstation places the same GPU family into a more compact design.

The interesting part is not the card count. It is the shape of the machine. Two 230W class accelerators in one chassis force a thermal decision, and GUNNIR answered it with liquid cooling in the larger system. The compact system avoids that problem with a vapor chamber instead, which is the cheaper path when the density is lower [NEEDS VERIFICATION for the exact cooling topology and radiator size on each chassis].

Per VideoCardz, GUNNIR paired the graphics with Intel Core Ultra 200HX processors, which VideoCardz identifies with the Arrow Lake-HX generation. That pairing matters for a practical reason. A 200HX part is a mobile class CPU with a power envelope far below a desktop H series chip, which fits the workstation positioning better than a hot desktop platform would [NEEDS VERIFICATION for which 200HX SKUs GUNNIR offered and their core counts].

Intel Connection is Intel's own regional event for partners, integrators and press, so a board partner showing completed workstations rather than reference cards is a meaningful signal. It means at least one vendor has committed engineering to multi accelerator boxes around the Arc Pro B70, which is more than most GPU launches deliver at launch.

The two systems on the show floor

Both GUNNIR machines share the same broad idea and differ in scale. The larger system carries two Arc Pro B70 cards and liquid cooling. The smaller one uses the same GPU family in a tighter enclosure cooled by a vapor chamber. Per VideoCardz by WhyCry, the larger machine is the one with the 64GB combined VRAM headline.

GUNNIR did not publish model names for either system in the coverage relayed by VideoCardz, so treat any specific SKU you see quoted elsewhere as unconfirmed until the vendor lists it [NEEDS VERIFICATION for system model names, chassis dimensions and availability dates].

System shownGPU countCombined VRAMCooling
GUNNIR large workstation2x Intel Arc Pro B7064GB (32GB per card)Liquid cooling
GUNNIR compact workstationArc Pro B70 family, count not specified in coverageNot specified, per coverageVapor chamber

Note what the table does not claim. The 64GB figure is an addition of two 32GB cards, not a single memory pool. Nothing in the VideoCardz report indicates that the two devices behave as one 64GB graphics card, and that distinction drives most buying mistakes around this hardware.

What 64GB of combined VRAM really means

Each Arc Pro B70 carries 32GB of GDDR6 ECC memory, per Intel's published specifications as relayed by VideoCardz. Two of them therefore total 64GB, and that arithmetic is correct. The wrong conclusion is that software sees one 64GB card.

Per VideoCardz, the B70 is not presented to software as one conventional 64GB graphics card. Instead, Intel supports multi GPU compute through oneAPI, for workloads that are written to use multiple devices. That means a framework or model server has to shard work across both cards deliberately. If your stack runs one process against one device, a second card buys you little.

This is the single most important thing to understand about the dual B70 build. Capacity that no framework can address is not capacity. Inference servers that support tensor or pipeline parallelism can use both devices, and Intel's oneAPI tooling exists for exactly that. A single GPU application, a game, or a tool with hardcoded single device behavior will ignore the second card entirely.

ECC memory is the other half of the pitch. Error correcting memory is a workstation and datacenter feature, not a gaming one. VideoCardz framed the B70 line as aimed at local AI inference and multi GPU workloads, and ECC is part of why. Long inference runs on consumer GDDR without ECC can corrupt results quietly, which is not an acceptable failure mode for professional work.

VideoCardz also noted that these dense dual GPU builds fit the B70 positioning rather than gaming desktops. That distinction matters because expectations differ completely. Nobody buys a workstation for frame rates, and the hardware is not marketed that way.

Arc Pro B70 specs that matter

The B70 uses Intel's larger BMG-G31 Battlemage GPU, per Intel's specifications relayed by VideoCardz. That chip carries 32 Xe cores and 256 XMX engines, with up to 367 INT8 TOPS of integer throughput. Intel lists memory bandwidth at 608 GB/s and reference board power at 230W, while noting that partner implementations can sit anywhere in a 160W to 290W range.

The wide partner power range deserves attention. A 230W reference design is not what ends up in your chassis, and a partner card running near 290W changes the power budget of a dual GPU machine by more than a hundred watts. Budget from the actual card you buy, not from Intel's reference number [NEEDS VERIFICATION for GUNNIR's specific board power figure for the cards used in these workstations].

SpecIntel reference for Arc Pro B70Why it matters here
GPUBMG-G31 BattlemageThe larger Battlemage die, not the cut down one
Xe cores32Compute throughput driver
XMX engines256Matrix throughput for AI workloads
INT8 TOPSUp to 367Local inference headline number
Memory32GB GDDR6 ECCECC matters for long professional runs
Memory bandwidth608 GB/sPer card, so 1216 GB/s aggregate across two
Reference board power230WReference design, not every partner card
Partner power range160W to 290WPlans your PSU and cooling headroom
Multi GPU supportVia oneAPISoftware must shard work across devices

Aggregated bandwidth is a legitimate benefit of two cards when the workload splits, because 608 GB/s per device adds to 1216 GB/s when both devices stream in parallel. It is not an advantage for a single device workload, and memory latency across PCIe does not make the two cards behave like one wider memory interface.

GUNNIR's own B70 TF 32G and the MXM modules

GUNNIR is not only a systems integrator at this event. Per VideoCardz, the company already sells its own Arc Pro B70 TF 32G with a blower style cooler aimed at continuous workstation use. GUNNIR's listing carries the same 32 Xe cores, 256 XMX engines, 32GB of ECC memory and 367 TOPS as Intel's reference figures.

A blower cooler on a workstation card is a deliberate choice, not nostalgia. Blower coolers push air through the card and exhaust it out the back of the chassis, which is what you want in a dense multi card server style build where each card sits in a restricted slot. It is louder and less thermally efficient than an open air cooler in a normal desktop, and it is the right trade when you are stacking two or more accelerators [NEEDS VERIFICATION for GUNNIR's published TF 32G noise and thermal ratings].

GUNNIR also displayed MXM format Arc Pro modules based on the B50 and B65, per the VideoCardz report. The stated purpose is compact workstation systems where a normal PCIe card would be too large. MXM is an older module standard, but its footprint advantage is real in small chassis, and it lets a system builder pick a mid tier accelerator where a full height card will not physically fit.

Putting those two product lines next to each other shows the strategy. GUNNIR covers the full size dual B70 workstation, the smaller high density workstation, the single slot blower card, and the MXM module route. That is a vendor treating local AI as a product line rather than a single launch SKU.

Five steps before you buy local AI hardware

If you are weighing hardware for local AI work rather than gaming, these five steps in order will save you more money than any discount. Each one is a decision that is much cheaper to get right before the purchase than after.

  1. Decide if one card is enough. Start with your actual model size and batch size, not with a memory number. A single 32GB B70 card may already fit your workload, in which case a second card adds cost, power and cooling load for no gain. Add the second card only when you have measured that a model will not load, or that throughput is limited by per device memory. Buying 64GB because it sounds better is the most common expensive mistake in this market.
  2. Check that your framework supports multi GPU. This is the step people skip, and it invalidates everything above it. Verify that your inference server or training stack explicitly supports tensor parallelism, pipeline parallelism or device mapping across two Intel GPUs through oneAPI. If your tool only ever binds one device, a second B70 will sit idle and you will have paid for a heatsink. Test the two device path before committing to the dual system rather than after.
  3. Budget ECC memory and board power in the price. ECC memory is part of why this hardware exists, so confirm your software and drivers actually use the correction, and confirm the platform supports ECC across the whole board if system memory matters to you. On power, remember that Intel's 230W figure is a reference board and partner cards run from 160W to 290W. Two cards plus a 200HX class CPU means planning a power supply well above what a gaming desktop would need [NEEDS VERIFICATION for GUNNIR's recommended PSU wattage on the dual B70 system].
  4. Plan liquid cooling headroom before the chassis. Two 230W class accelerators in one box generate more heat than most single card air cooled designs can move. GUNNIR used liquid cooling in the larger machine and a vapor chamber in the smaller one, per VideoCardz, which is the pattern you should expect to need. If you build your own, confirm radiator size, pump reliability, loop fill and the service access required before the cards go in. Planning cooling after assembly means rebuilding the machine.
  5. Treat it as AI hardware, not a gaming build. VideoCardz explicitly framed these dense dual GPU systems as fitting the B70 positioning rather than gaming desktops, and Intel positions the B70 for local AI inference and multi GPU workloads. That means you should judge these machines on memory capacity, ECC behaviour, thermal stability under sustained load and driver maturity. Gaming benchmarks, RGB lighting expectations and upgrade headroom are the wrong criteria for this category, and buying on them will point you at the wrong product.

None of these five steps needs a purchase to complete. The first two are software checks you can do this week with hardware you may already own. Only after both come back positive does the power and cooling planning become a real budget question.

Pricing and why the price went up

Intel launched the Arc Pro B70 earlier in 2026 with a suggested starting price of $949 for Intel's own card, per VideoCardz by WhyCry. Partner pricing varies by implementation, and the variation is not small. A blower cooled partner card with different power limits is a different product at a different price [NEEDS VERIFICATION for current street pricing on GUNNIR TF 32G and partner B70 boards].

More importantly, VideoCardz noted earlier coverage of 26% to 48% B70 price increases across global retailers. That trend matters for anyone reading this and planning a purchase in the near future. The direction of travel on Arc Pro B70 pricing has been upward, not downward, which is unusual for a newly launched workstation part and reflects demand from inference buyers rather than supply problems.

That price movement also explains the timing of these workstation launches. GUNNIR is not building dense dual accelerator systems to compete with gaming desktops on value. It is filling a capacity segment where buyers need memory that will not error out and cooling that will hold under sustained load, and where the 26% to 48% increase has not discouraged demand. For a buyer, the practical implication is that waiting for discounts is probably a longer wait than usual on this specific product.

FAQ

Is the 64GB in the GUNNIR dual B70 workstation a single graphics card?
No. Per VideoCardz by WhyCry, it is two Intel Arc Pro B70 cards with 32GB of GDDR6 ECC memory each, added together. Software sees two devices, not one 64GB card.

How does software actually use two Arc Pro B70 cards?
Per Intel's multi GPU support relayed by VideoCardz, compute spans both devices through oneAPI. That requires a framework written to shard work across devices, so single device software will only use one card.

What CPU did GUNNIR pair with the Arc Pro B70 cards?
Per VideoCardz, Intel Core Ultra 200HX Arrow Lake-HX processors. The specific SKUs offered in each workstation were not in the coverage we saw [NEEDS VERIFICATION].

How much power does a dual B70 workstation draw?
Intel lists 230W reference board power and notes partner implementations range from 160W to 290W per card, per VideoCardz. Two cards plus a 200HX class CPU puts the system well above normal desktop levels, so size the power supply accordingly.

How much does an Arc Pro B70 cost?
Intel launched the B70 earlier in 2026 at a $949 suggested starting price for its own card, per VideoCardz. Partner cards cost more, and VideoCardz noted earlier coverage of 26% to 48% price increases across global retailers.

Why no buying links or prices for the GUNNIR systems?
GUNNIR did not publish model names, configurations or prices for the two workstations in the coverage relayed by VideoCardz by WhyCry on September 27, 2026. We will not invent a price or a store link for hardware that has not been listed.

Bottom Line

GUNNIR packing two Intel Arc Pro B70 cards into one liquid cooled workstation, per VideoCardz by WhyCry reporting from Intel Connection 2026 in China, is a real signal that the Arc Pro B70 is moving from component to finished system. Sixty four gigabytes of combined ECC VRAM in one chassis solves a real capacity problem for local AI inference buyers, and the smaller vapor cooled machine extends that idea into a tighter form factor.

The catch is that 64GB of combined VRAM only helps if your software can address two devices through oneAPI. Check that first, because a second card in a single device application is a heatsink with a price tag. If your framework does shard work, then the power budget, the liquid cooling headroom and the 26% to 48% price increases all point the same direction, which is that this is professional capacity hardware with professional cooling costs, not a gaming build.

Watch for GUNNIR publishing model names, configurations and pricing when they release the systems. Until then the honest position is that the hardware direction is confirmed and the buying specifics are not.

Sources

All reported facts in this article come from VideoCardz by WhyCry, published September 27, 2026, covering GUNNIR's Intel Connection 2026 workstation demonstrations and relaying GUNNIR, Intel Connection 2026 and IThome. Arc Pro B70 specifications including 32 Xe cores, 256 XMX engines, up to 367 INT8 TOPS, 32GB GDDR6 ECC, 608 GB/s memory bandwidth, 230W reference board power, the 160W to 290W partner range and the $949 suggested starting price are Intel figures as relayed by VideoCardz. GUNNIR product details for the Arc Pro B70 TF 32G blower card and the MXM modules based on the B50 and B65 come from GUNNIR as reported by VideoCardz.

Sources referenced: VideoCardz (WhyCry, September 27, 2026 report on the GUNNIR dual Arc Pro B70 workstations), GUNNIR (Arc Pro B70 TF 32G blower card and MXM Arc Pro modules) and IThome (original coverage relayed by VideoCardz). Intel product specifications were cited as relayed by VideoCardz rather than linked directly.

Every claim marked [NEEDS VERIFICATION] needs confirmation against Intel, GUNNIR or IThome primary material before it is repeated as fact. This article carries no retailer links, no preorder links and no store listings, because no verified pricing for the GUNNIR workstations existed at the time of writing.

Share