I'd love to hear what you think! Please drop me a line and let me know what you like and what could be better. 🙏

Intel Arc Pro B65 32GB Specs, Benchmarks & Pricing

The Intel Arc Pro B65 is a professional workstation GPU announced on March 25, 2026 alongside the Arc Pro B70, with AIB partner availability in mid-April 2026. Built on the Xe2-HPG (Battlemage) architecture using TSMC's 5nm process, featuring 20 Xe2 cores, 160 XMX (Xe Matrix eXtension) engines for AI acceleration, and 20 ray tracing units. The B65 is distinguished from the Arc Pro B60 primarily by its expanded 32GB GDDR6 memory capacity on a wider 256-bit memory interface running at 19 Gbps, delivering 608 GB/s of memory bandwidth. This makes it particularly well-suited for large-scale AI inference workloads and professional applications requiring extensive VRAM. The GPU delivers 12.28 TFLOPS of FP32 compute performance and 197 TOPS of INT8 AI inference performance through its 160 XMX engines. The XMX engines support precision formats including INT4, INT8, FP16, and BF16, providing exceptional flexibility for AI inference tasks. With a TBP of 200W, the Arc Pro B65 shares its power envelope with the Arc Pro B60. The GPU uses a cut-down BMG-G31 die (27.7 billion transistors on a 368mm2 die, manufactured on TSMC N5), the same larger die used by the Arc Pro B70 but with 20 of 32 Xe2 cores active. This distinguishes the B65 from the Arc Pro B60 and Arc B580, which use the smaller BMG-G21 die. The B65 supports PCIe 5.0 x16 interface and features DisplayPort 2.1 outputs with support for up to four displays at resolutions up to 7680x4320. The card has been certified for professional applications including Autodesk Maya, SolidWorks, Adobe Premiere Pro, Blender, and DaVinci Resolve. The GPU also supports SR-IOV, Linux containers, and parallel multi-GPU operation. Intel's expanded VRAM strategy positions the Arc Pro B65 as an inference workstation GPU targeting LLM inference tasks where 32GB VRAM enables loading larger AI models without quantization or offloading.

Release Date: April 15, 2026
MSRP: $799 USD
GPU Architecture: Xe2-HPG (Battlemage)
Hardware-Accelerated GEMM Operations:
FP64 FP32 TF32 BF16 FP16 FP8 FP4 INT8 INT4
CUDA Compute Capability : n/a

Strengths

Excellent memory capacity (top 21% of GPUs)

Considerations

Lower FP32 compute performance compared to other GPUs
Lower FP16 compute performance compared to other GPUs

Specifications for Intel Arc Pro B65

Specification	Performance Ranking
FP32 TFLOPs	16th @ 12.28 TFLOPs
FP16 TFLOPs	13th @ 24.58 TFLOPs
Tensor Core Count	39th @ 160 Cores (Entry Tier)(Entry)
Memory Capacity (GB)	79th @ 32 GB (Top Tier)(Top)
Memory Bandwidth (GB/s)	52nd @ 608 GB/s (Mid Tier)(Mid)
Int8 TOPs	35th @ 197 TOPs (Entry Tier)(Entry)
FP8 TFLOPs	Data not available
FP4 TFLOPs	Data not available

Real-time Intel Arc Pro B65 GPU Prices

We're tracking 3 of the Intel Arc Pro B65 GPUs currently available for sale. The lowest average price is $1,670 (what's this?)

Buy Now

Compare Price/Performance to other GPUs

We track real-time prices of other GPUs too so that you can compare the price/performance of the Intel Arc Pro B65 GPU to other GPUs.

Compare GPU Price/Performance

Compare Intel Arc Pro B65 to Another GPU

Compare the Intel Arc Pro B65 directly to another GPU to see specs, benchmarks, and prices side-by-side.

Compare GPUs Side-by-Side

Price History

This chart tracks Intel Arc Pro B65 prices over the past 6 months based on eBay listings. The yellow line shows the average of the three lowest-priced listings each day—a useful indicator for finding deals.

Trend: Over this period, the lowest average price has been up 55.0%. The current lowest average is $1,549, compared to a typical $1,460 over the period. The lowest average price is calculated as the average of the three lowest-priced listings each day.

gpupoet.com

Product Identifiers

Manufacturer Part Numbers (1)

Spec Code: 245796

References

Notes

FP32 TFLOPS of 12.28 sourced from Intel official product specifications. The Arc Pro B65 uses a cut-down BMG-G31 die with 20 active Xe2 cores (of 32 total) at 2400 MHz boost clock, delivering the same compute throughput as the B60 despite using a different die.
FP16 TFLOPS of 24.58 based on 2:1 ratio of FP16 to FP32 performance, consistent with Xe2 architecture and identical to the Arc Pro B60. Intel does not publish FP16 TFLOPS directly for professional workstation GPUs. Value follows the same methodology documented for the Arc Pro B60.
INT8 TOPS of 197 represents dense (non-sparse) INT8 AI performance sourced from Intel official product specifications. Intel Arc XMX engines do not implement NVIDIA-style 2:4 structured sparsity, so published values are inherently dense and used directly for cross-vendor comparison. The B65 uses the same compute configuration as the B60 (20 Xe2 cores, 160 XMX engines at 2400 MHz), resulting in the same INT8 TOPS.
Memory bandwidth of 608 GB/s calculated as 256-bit interface x 19 Gbps / 8 bits/byte = 608 GB/s. The B65 uses a wider 256-bit memory bus compared to the B60s 192-bit bus (456 GB/s), providing significantly higher bandwidth while maintaining the same 19 Gbps GDDR6 memory speed.
Memory capacity of 32GB GDDR6 is the primary differentiator from the Arc Pro B60 (24GB on 192-bit bus). The wider 256-bit bus enables both higher bandwidth and larger VRAM capacity.
TBP (Total Board Power) is 200W sourced from Intel official specifications. Despite using the larger BMG-G31 die (vs B60 on BMG-G21), the TBP matches the B60 because only 20 of 32 Xe2 cores are active.
XMX (Xe Matrix eXtension) engine count of 160 based on 20 Xe2 cores x 8 XMX engines per core, confirmed by architecture documentation. These are Intel dedicated AI accelerators, equivalent to tensor cores in NVIDIA GPUs.
FP64 double precision performance runs on the general-purpose shader units (XVEs) per Xe2 architecture documentation, though Intel does not officially publish FP64 TFLOPS for this GPU. It is not included in supportedHardwareOperations because that field is scoped to the XMX matrix/tensor (GEMM) path only, not general XVE ALU precisions (see Phase 5 audit note below).
Estimated MSRP of $799 USD based on market positioning between the B60 ($599) and B70 ($949-$950). Intel has not published an official MSRP for the B65. The B65 premium over the B60 reflects the additional VRAM (32GB vs 24GB) and memory bandwidth (608 vs 456 GB/s) from the wider 256-bit memory interface.
Announced March 25, 2026 alongside the Arc Pro B70. The B65 is available through AIB partners only (Intel is not producing a reference design), with availability estimated mid-April 2026. Source: ServeTheHome, Tom's Hardware, and HotHardware launch coverage.
The B65 uses the BMG-G31 die (27.7 billion transistors, 368mm2, TSMC N5), the same die as the Arc Pro B70, but with 20 of 32 Xe2 cores active. This is distinct from the Arc Pro B60 and Arc B580 which use the smaller BMG-G21 die (19.6 billion transistors, 272mm2). The B65 features PCIe 5.0 x16 (vs B60's PCIe 5.0 x8) and a 256-bit/32GB GDDR6 memory configuration.
The GPU supports DirectX 12 Ultimate, Vulkan 1.3, OpenGL 4.6, and OpenCL 3.0 APIs.
Intel Project Battlematrix multi-GPU strategy applies to the B65, enabling configurations with multiple B65 GPUs providing very large combined VRAM pools for LLM inference workloads.
No third-party AIB products documented as confirmed product data for B65 AIB cards was not available in public sources at time of research. AIB partners confirmed for the B60 (ASRock, Maxsun) may also offer B65 variants.
Phase 5 precision audit (2026-09-02): supportedHardwareOperations corrected to FP16, BF16, INT8, INT4 (removed FP32, FP64, and TF32) per the canonical Xe2-HPG (Battlemage) supportedHardwareOperations definition in gpu-researcher.md. TF32 was a data-entry inconsistency copied from the same erroneous assumption applied to the B60/B70 files: it is not documented for the Xe2/Battlemage generation in Intel's official B60 or B70 datasheets, and appears to conflate Xe2-HPG XMX with the separate Xe-HPC (Ponte Vecchio) architecture, which is where Intel's generic oneAPI documentation actually applies TF32 support. FP32 and FP64 were removed because this field is scoped to hardware-accelerated XMX matrix/tensor (GEMM) paths only; FP32/FP64 run on the general-purpose Xe Vector Engines (XVEs) per Intel VTune Profiler documentation, which defines a dedicated XMX instruction-count metric scoped to FP16/BF16/INT8/INT4/INT2 only. The B65 shares the same 160 XMX engine / 20 active Xe2-HPG core configuration as the B60, so the same precision set applies.
fp8TFLOPS and fp4TFLOPS are intentionally omitted (not added to unverifiedSpecs) because this is a confirmed hardware negative, not a research gap: Intel's official Arc Pro B70 datasheet (same BMG-G31 die and Xe2-HPG generation as the B65) and the chipsandcheese Battlemage architecture analysis document no FP8 or FP4 XMX data type support anywhere in the Xe2/Xe2-HPG generation.
gpuArchitecture changed from "Xe2" to "Xe2-HPG (Battlemage)" during the 2026-09-02 audit to normalize naming across all six Arc B/Pro B-series files. Intel's official Arc Pro B65 datasheet (datasheet-b65-gpu.pdf) directly confirms "Xe-cores: 20 Xe 2-HPG" in its Specifications table, alongside matching 197 AI TOPS Int8 Dense, 608 GB/s, 32GB, 256-bit, 160 XMX engines, and 200W TBP figures already in this file - this is a direct confirmation, not an inference from the B60/B70 datasheets. Intel's ARK page for the B65 uses the terser "Microarchitecture: Xe2" label, matching the same ARK-vs-datasheet format difference seen across the whole B-series family (B570, B50 ARK pages also omit "HPG"); no genuine hardware-based Pro-vs-consumer naming split exists.