✨ From vibe coding to vibe deployment. UBOS MCP turns ideas into infra with one message.

Learn more
Andrii Bidochko
  • Updated: March 26, 2026
  • 17 min read

Intel Announces Arc Pro B70 and B65 GPUs – Full Overview

Home Reviews Forums Downloads Case Mod Gallery Databases Databases… Back VGA Bios Collection GPU Database CPU Database SSD Database Review Database Upcoming Hardware Our Software Our Software… Back GPU-Z RealTemp NVCleanstall TPUCapture MemTest64 More More… Back Articles Old Stuff Computer Trivia TPU Live RAM Latency Calculator Contact Us Wednesday, March 25th 2026 Intel Announces Arc Pro B70 and Arc Pro B65 GPUs, Maxes Out Xe2 “Battlemage” Architecture by btarunr Wednesday, 16:21 Discuss (32 Comments) Intel today announced the Arc Pro B70 and Arc Pro B65 graphics cards for advanced AI compute workloads on workstations, and professional visualization. The two primarily target local inferencing, software development, and deployments in multi-GPU configurations for rack scale AI GPU compute acceleration. The Arc B70 GPU in particular stands out, because it is the most powerful discrete GPU based on Intel Xe2 “Battlemage” graphics architecture, with 32 Xe cores, and a 256-bit wide GDDR6 interface.If you recall, the Arc B580 gaming GPU only comes with 24 Xe cores and a 192-bit interface, and we for long wondered if Intel would ever max out the silicon for a more powerful SKU. The Arc B70 is that SKU. The Intel Arc Pro B70 is configured with 32 Xe cores (Xe2-HPG), 256 XMX engines, and 32 Ray Tracing Units. It comes with 32 GB of GDDR6 memory across a 256-bit wide memory interface, with 608 GB/s of bandwidth on tap. The card comes with a PCI-Express 5.0 x16 host interface.The Arc B70 offers a peak throughput of 367 TOPS (INT8). On the graphics side of things, it supports DirectX 12 Ultimate, OpenGL 4.6, Vulkan 1.3. Compute APIs include Intel’s own oneAPI, OpenCL 3.0, and OpenVINO. Its media engine supports AV1, HEVC, VP9, and H.265 hardware-accelerated encode and decode. Display outputs include four DisplayPort 2.1 ports. The card comes with power draw ranging between 160 W to 290 W depending on partner implementation (230 W for the Intel reference card).Intel will provide certified drivers for Windows 11, Windows 10, and Linux. Intel Arc Pro B65 is an interesting variant, it comes with a similar core configuration to the Arc B580 gaming GPU, but with a few key changes. It has 24 Xe cores, 160 XMX cores, and 20 Ray Tracing Units. Where it differs is memory and PCIe. It provides 32 GB of GDDR6 memory across a 192-bit wide memory interface, unlike the Arc Pro B70. The card is probably using asymmetric memory chip configuration to achieve this.Secondly, the card uses the full PCI-Express 5.0 x16 interface available to the GPU, and not PCI-Express 5.0 x8 found on the Arc B580. API support, media acceleration features, and display I/O are same as the B70. The card comes with a 200 W typical board power.Related News Tags: Arc Pro Battlemage Intel Intel Arc Intel Xe Windows 11 Xe2-HPG Sep 9th 2025 Intel Arc B770 “Battlemage” Nears Launch as Intel Prepares Packaging (59) Sep 3rd 2025 Intel Arc Pro B50 GPU Arrives at $349 for Small-Form-Factor Workstations (25) Mar 20th 2026 Intel Offered Hardware Optimization Help to Crimson Desert Developer for Years, Company Says (71) Jan 10th 2026 The 5 Biggest No-Shows of CES 2026 (83) Oct 10th 2025 Intel XeSS 3 Expands Multi-Frame Generation Support to All XeSS 2 Titles (19) Jan 5th 2026 Intel Launches Core Ultra Series 3 “Panther Lake” Mobile Processors on Intel 18A Node (40) Jun 6th 2025 NVIDIA Grabs Market Share, AMD Loses Ground, and Intel Disappears in Latest dGPU Update (312) Dec 6th 2025 Intel Arc “Battlemage” BMG-G31 GPU Emerges in Official Software Support (40) Jan 14th 2026 DDR3 Platforms Make a Comeback Amid Severe Memory Shortages (63) Mar 2nd 2026 Microsoft Shader Execution Reordering Brings 90% Performance Increase on Intel Arc B-Series, 80% on NVIDIA “Blackwell” GPUs (60) Add your own comment 32 Comments on Intel Announces Arc Pro B70 and Arc Pro B65 GPUs, Maxes Out Xe2 “Battlemage” Architecture 1 to 25 of 32Go to Page12PreviousNext #1 Here’s hoping the B70 is less than $1000, but with the DRAM-pocalypse I don’t know if that’ll happen. The fact that it uses GDDR6 helps at least. It’d make sense for this to cost about twice what the B50 cost, which would put this around $700; maybe a little less since it doesn’t have the SFF tax. Of course, that framebuffer is the main reason I (and many others I’m sure) want this GPU. EDIT: Newegg has it listed for $950 and releasing on 4/24. To preorder or not to preorder. #2 B65 seems to be listed as a 256-bit memory bus, so not an async configuration. Maybe the slideware is wrong? www.intel.com/content/www/us/en/products/compare.html?productIds=245797,245796sam_86314Here’s hoping the B70 is less than $1000, but with the DRAM-pocalypse I don’t know if that’ll happen. The fact that it uses GDDR6 helps at least. It’d make sense for this to cost about twice what the B50 cost, which would put this around $700; maybe a little less since it doesn’t have the SFF tax. Of course, that framebuffer is the main reason I (and many others I’m sure) want this GPU.EDIT: Newegg has it listed for $950 and releasing on 4/24. To preorder or not to preorder.B70 is $949. B65 is partner-only. #3 sam_86314Here’s hoping the B70 is less than $1000, but with the DRAM-pocalypse I don’t know if that’ll happen. The fact that it uses GDDR6 helps at least. It’d make sense for this to cost about twice what the B50 cost, which would put this around $700; maybe a little less since it doesn’t have the SFF tax.Of course, that framebuffer is the main reason I (and many others I’m sure) want this GPU. EDIT: Newegg has it listed for $950 and releasing on 4/24. To preorder or not to preorder.Lets go #4 Unless the memory bandwith is doign some serious lifting, the gaming performance of this chip would put it just under the 5070, consistently losing to the 9070. I can kinda see why intel hasnt had any interest in a consumer version of this card. #5 $950 for 32GB VRAM is pretty good.#6 I wonder how great a demand there is for these – well one is already sold I can see 👍 #7 TheinsanegamerNUnless the memory bandwith is doign some serious lifting, the gaming performance of this chip would put it just under the 5070, consistently losing to the 9070. I can kinda see why intel hasnt had any interest in a consumer version of this card.For gaming performance, I’m guessing around the level of a 6800 XT or 9060 XT.As for ML performance (I know, I know), I’ve found the B50 to be about half as fast as my 9070 XT with diffusion models* (which is insane for a card that only uses 50W), so I’d expect the B70 to be comparable to the 9070 XT. For LLMs, GPT OSS 20B runs at around 48 tokens/sec* on the B50 and around 57 tokens/sec* on the 9070 XT. In theory, the B70 should be a lot faster here. *I’ll verify these numbers after work.#8 sam_86314For gaming performance, I’m guessing around the level of a 6800 XT or 9060 XT. As for ML performance (I know, I know), I’ve found the B50 to be about half as fast as my 9070 XT with diffusion models* (which is insane for a card that only uses 50W), so I’d expect the B70 to be comparable to the 9070 XT. For LLMs, GPT OSS 20B runs at around 48 tokens/sec* on the B50 and around 57 tokens/sec* on the 9070 XT. In theory, the B70 should be a lot faster here.*I’ll verify these numbers after work.Based on the performance of the B580, it would be well above the 9060xt, but below the 9070. B580 is ~29.6, B70 has twice the core count and memory bus, so ~59.2 FPS, unless the nearly triple memory bandwidth fixes a previous bottleneck. 9060xt is only 43.8 FPS, and 9070 is 76.5. RTX 5070 is 61.2. This seems pretty consistent in other games I looked at. So it would have been a 5070 competitor. Not bad, with 16GB of VRAM it would have been a better deal.But it would have been power hungry, 290w vs 250w for the 5070. #9 TheinsanegamerNBased on the performance of the B580, it would be well above the 9060xt, but below the 9070. B580 is ~29.6, B70 has twice the core count and memory bus, so ~59.2 FPS, unless the nearly triple memory bandwidth fixes a previous bottleneck. 9060xt is only 43.8 FPS, and 9070 is 76.5. RTX 5070 is 61.2. This seems pretty consistent in other games I looked at. So it would have been a 5070 competitor.Not bad, with 16GB of VRAM it would have been a better deal. But it would have been power hungry, 290w vs 250w for the 5070. The B580 has 20 compute units and the B70 will have 32. That puts the relative performance around 47.4 FPS, or right in line with the 9060 XT. I was comparing it to my B50 since that has 16 compute units and 16GB VRAM, so the B70 is literally twice the GPU. #10 sam_86314The B580 has 20 compute units and the B70 will have 32. That puts the relative performance around 47.4 FPS, or right in line with the 9060 XT. I was comparing it to my B50 since that has 16 compute units and 16GB VRAM, so the B70 is literally twice the GPU.so it is. I thought the B580 was a B50. That makes it even worse. 9060xt performance while pulling more power them a 5070. oof. #11 TheinsanegamerNso it is. I thought the B580 was a B50. That makes it even worse. 9060xt performance while pulling more power them a 5070. oof.As for the power usage, however, the B50 is incredibly efficient, with a power limit of 50W. The relative performance calculation would put the B50 at 23.7 FPS, for 0.47 frames per watt. Compare that to roughly 0.24 frames per watt for the 9060 XT (assuming 180W). I can’t verify those numbers for RE9 since I don’t own it, but I can say that the B50 consistently seems to perform about half as well as my 6800 XT (at 255W) in games.I’m also assuming the power figures on TPU are core power only, not total board power (which is 70W for the B50). I’d imagine the B70 might do well with being power limited and undervolted (assuming the options are even available). Then again, we’ll see how big Battlemage scales compared to Alchemist. #12 2 questions- 1. How many Virtualized Functions are going to be supported? B50 ended-up being only 2 (despite support for 12 at/shortly after launch) 2.Is that an 8-pin EPS on the Intel 1st Party Card? Considering, the 75W PCIe slot power limit and 290W rating, it shouldn’t be a single 6+2 PCIe. #13 TheinsanegamerNUnless the memory bandwith is doign some serious lifting, the gaming performance of this chip would put it just under the 5070, consistently losing to the 9070. I can kinda see why intel hasnt had any interest in a consumer version of this card.It has less memory bandwidth than the 5070, in fact.256-bit vs 192-bit bus, but GDDR6 vs GDDR7. #14 LabRat 8912 questions- 1. How many Virtualized Functions are going to be supported? B50 ended-up being only 2 (despite support for 12 at/shortly after launch) 2. Is that an 8-pin EPS on the Intel 1st Party Card? Considering, the 75W PCIe slot power limit and 290W rating, it shouldn’t be a single 6+2 PCIe.It’s written on one of the slides the Intel branded card is 230 W.P6000 runs on a single 8 pin too, configured for 250 W, should be OK #15 @lambda actually brought up a good point here: [EMBED content=”post-5697823″]https://www.techpowerup.com/forums/threads/intel-arc-pro-b70-shows-up-on-newegg-with-april-release-date-and-949-99-price.347710/post-5697823[/EMBED] R9700 is $1350 on Newegg, theoretically has much better raster performance and raw compute, but might be quite a bit slower with ML. Should we continue this discussion over there?Dr.DroIt’s written on one of the slides the Intel branded card is 230 W. P6000 runs on a single 8 pin too, configured for 250 W, should be OKAs long as we don’t get a repeat of the RX 480 apparently sometimes burning out PCIe slots, due to the 150W TBP but only having a single 6-pin PCIe power. #16 sam_86314As long as we don’t get a repeat of the RX 480 apparently sometimes burning out PCIe slots, due to the 150W TBP but only having a single 6-pin PCIe power.Wasnt that confirmed as either a rumor or only having happened once in a freak accident? #17 TheinsanegamerNWasnt that confirmed as either a rumor or only having happened once in a freak accident?The over-spec current draw off the PCIe slot was a real and known issue. -can’t speak towards any damages it caused but, it was common on the RX 480. www.techpowerup.com/223981/amd-releases-pci-express-power-draw-fix-we-tested-confirmed-works #18 LabRat 8912.Is that an 8-pin EPS on the Intel 1st Party Card? Considering, the 75W PCIe slot power limit and 290W rating, it shouldn’t be a single 6+2 PCIe.Do GPUs always take the maximum allowed power from the slot? Might be problematic in multi-GPU setups. #19 WirkoDo GPUs always take the maximum allowed power from the slot? Might be problematic in multi-GPU setups.AFAIK, it’s presumed that they do. IIRC, in actual testing, it varies.You’re not wrong though- It’s why MSI more-often uses x1/x16(phys) slots, w/ the full x16’s GNDs connected and why many of their boards have a PCIe 8-pin AUX-input for the slots’ +12V/GND rail/plane*. All of their ‘AI certified’ -or whatever- boards *can* safely support multiple GPUs. *On PCIe slots, all +12V and +3.3V pins are in the first block of 22 (11 ea. side) pins. However, there are additional ground conductors surrounding ea.lane’s Rx and Tx conductors #20 WirkoDo GPUs always take the maximum allowed power from the slot? Might be problematic in multi-GPU setups.I follow and observe this a lot in GPU-Z. Assuming GPU-Z is reporting properly, GPUs preferentially draw power from the PSU connections while taking ~20-40W from the PCIe slot. One thing I never did (for some unknown reason?) was track my GTX 1080’s 8-pin vs slot power draw as it topped out at 216W which is only 9W shy of 150+75W max draw from these 2 power sources. Maybe separate power draw wasn’t reported on that GPU or something but currently my 1080 is lacking a cooler for various reasons so I can’t re-test it. #21 sam_86314As long as we don’t get a repeat of the RX 480 apparently sometimes burning out PCIe slots, due to the 150W TBP but only having a single 6-pin PCIe power.Well, I mentioned the P6000 because I have one. The TBP is 250 W, but it runs off a single 8-pin. I suppose this card technically violates the specification even if ever so slightly, but then again I guess Nvidia’s engineering team knew what they were doing when they released an $5,999 workstation product this way. The 230 W reference B70 should be safe, too. #22 This needs to be said if it hasn’t been already but, B770 when?#23 DeltaHawk001This needs to be said if it hasn’t been already but, B770 when?I don’t think we’re getting one, unfortunately. These last two generations, AMD and Intel can only realistically compete at the budget level where consumers are the most price-conscious. At ~$450 and above, the vast majority of consumers are willing to pay the $50-150 premium for Nvidia features to go with their 1440p monitors and better computers.Intel knows this, which is why it’s not even bothering with the consumer market in that segment, instead focusing on workstation cards, where saving $500-1000 over Nvidia can be justified. Especially since AMD and Intel are more competitive on features in the professional segment than the gaming segment.#24 I know how much everyone loves AI around here (/s) I went ahead and ran ML tests (because AI is a stupid buzzword and ML is more descriptive) on my B50 (for extrapolating to B70), my 9070 XT (stand-in for R9700?), and my 6800 XT (to let the B50 show off). TL;DR: The B50 is freaking fast for a 50W card, and if everything scales linearly, the B70 will be faster than AMD’s current-gen best. Spoiler: Tests Systems: – i7 12700, 32GB DDR4-3600, Arc Pro B50 (50W, stock settings) – R7 5800X.64GB DDR4-3600, RX 9070 XT (~240W, -65mV, -250MHz) – R5 2600X, 48GB DDR4-2666, RX 6800 XT (~240W, 1085mV, 2400MHz) SD Next: Built-in benchmark using JuggernautXL – Arc Pro B50 (IPEX, BP16): 5.35 it/s – RX 9070 XT (ROCm 7.2, BP16): 8.76 it/s – RX 6800 XT (ROCm 7.2, FP16): 4.64 it/s CPU might affect the SD Next results, which would put the RX 6800 XT at a disadvantage.LM Studio: GPT OSS 20B, prompt is “explain how a graphics processing unit works”, reasoning set to high – Arc Pro B50 (Vulkan): 44.54 tok/s – RX 9070 XT (ROCm): 55.92 tok/s – RX 6800 XT (ROCm): 45.38 tok/s I checked the difference between Vulkan and ROCm on the 6800XT, and it was pretty much margin of error. Spoiler: Receipts LM Studio: Arc Pro B50 RX 9070 XT RX 6800 XT SD Next: Arc Pro B50 RX 9070 XT RX 6800 XT Intel claims the B70 will be ~2.17x as powerful as the B50 in ML workloads, so in theory, the B70 should get ~11.61 it/s in SD Next and ~96.65 tok/s in GPT OSS 20B. If we go with just a flat 2x over the B50, the numbers become ~10.7 it/s and 89.08 tok/s. Of course, who knows how these workloads would actually scale, and I don’t know if Vulkan uses the ML hardware or just the general-purpose compute. Feel free to point out the flaws in my testing methodology. EDIT: My LM Studio settings were severely handicapping my AMD cards.New LLM results: B50 (Vulkan): 48.07 tok/s 9070 XT (ROCm): 124.49 tok/s 6800 XT (ROCm): 93.55 tok/s B70 theoretical numbers now range from 96.14 tok/s to 104.31 tok/s. #25 “Its media engine supports AV1, HEVC, VP9, and H.265 hardware-accelerated encode and decode.” Excuse me, but HEVC is H.265. You probably mean HEVC and H.264. 1 to 25 of 32Go to Page12PreviousNext Add your own comment Mar 27th, 2026 02:08 +03 change timezone Sign in / Register Latest GPU Drivers NVIDIA GeForce 596.02 Beta AMD Radeon 26.3.1 WHQL Intel Arc & iGPU 101.8626 WHQL New Forum Posts 02:07 by LabRat 891 Overclocking on Intel platform without K Processor and Z Mainboard (7) 02:04 by Darmok N Jalad Windows 11 General Discussion (6803) 02:03 by Ripcord Upgrade to 9070 XT worth it?(100) 01:15 by Veseleil Ghetto Mods (4745) 00:53 by SmookinJoe TPU’s Rosetta Milestones and Daily Pie Thread (2591) 00:26 by FR3D1 Lets talk PC Audio (359) 00:24 by Flogger23m I’m glad if the time between hardware releases increases. Thoughts?(87) 00:23 by _roman_ Free Games Thread (5110) 00:10 by Easy Rhino Crimson Desert (14) 23:58 by remixedcat The Official Linux/Unix Desktop Screenshots Megathread (852) Popular Reviews Mar 23rd, 2026 Intel Core Ultra 7 270K Plus Review – Intel’s Fastest Gaming CPU Mar 23rd, 2026 Death Stranding 2 Handheld Performance Review Mar 24th, 2026 Crimson Desert Handheld Performance Review Mar 23rd, 2026 Intel Core Ultra 5 250K Plus Review – Disrupting AMD’s Entry-Level Mar 26th, 2026 XMG Neo 16 (A25) + Oasis Water Cooler Review – Performs Like an 18-incher Mar 11th, 2026 Upcoming Hardware Launches 2026 (Updated Mar 2026) Mar 25th, 2026 Sharkoon Skiller SGH40W Review – No-Frills Wireless Mar 24th, 2026 Ugreen NASync iDX6011 Pro Review Feb 25th, 2026 Resident Evil Requiem Performance Benchmark Review Mar 20th, 2026 FSP MEGA GM 1000 W Review TPU on YouTube Controversial News Posts NVIDIA Unveils DLSS 5 with Real-Time Neural Rendering (304) Microsoft Promises Slew of Windows 11 Changes in Response to “Microslop” Criticisms (162) NVIDIA Grabs 94% AIB GPU Market Share, AMD Falls to 5% (154) NVIDIA CEO Claims Gamers are “Completely Wrong” About DLSS 5 (154) NVIDIA DLSS 5 Gets 84% Dislikes on YouTube as Backlash Grows (130) NVIDIA Reportedly Preparing 9 GB GeForce RTX 5050 Variant with GDDR7 (126) PlayStation 6 Leak Tips 4K 120 FPS “In Most Games” With 6-12× RT Performance of PS5 (116) NVIDIA Pulls GeForce 595.59 WHQL Game Ready Driver After Widespread Bug Reports (111)


Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.

Sign up for our newsletter

Stay up to date with the roadmap progress, announcements and exclusive discounts feel free to sign up with your email.

Sign In

Register

Reset Password

Please enter your username or email address, you will receive a link to create a new password via email.