Intel Arc Pro B70 Review

Table of Contents

One of the key factors for AI-focused GPUs is memory capacity. In order to run models well — if at all — they must be stored in memory local to the GPU compute die. And given that GPUs are far more performant for most AI workloads than are CPUs, that means it must be stored in VRAM. To that end, we have seen increasingly large memory buffers on midrange professional GPUs targeted towards AI use cases. For example, last year AMD released the AMD Radeon™ AI PRO R9700, which repurposed the 9070 XT into an AI card by doubling the VRAM to 32 GB. At the higher end, NVIDIA released a 72 GB version of its NVIDIA RTX PRO™ 5000 Blackwell.

Now, Intel has thrown its hat into the ring with the launch of the Intel® Arc™ Pro B70. However, unlike AMD and NVIDIA, Intel is not merely retrofitting an existing GPU with additional VRAM. Instead, the B70 is a whole new entry into the product stack, featuring twice the Xe cores and VRAM of the Intel Arc Pro B50, bringing it to 32 GB of memory.

In our previous testing of the Arc Pro B50, we found substantial improvements over their last-gen Alchemist cards, such as the A50. Intel’s second-generation Xe2 architecture brought along a wide variety of enhancements, from SIMD16 execution to next-gen XMX engines, which dramatically improved both raw performance and efficiency. Additionally, over the last few years, Intel has worked hard to improve support, compatibility, and optimizations within its driver stack — with the latter often substantially improving the performance of already-released hardware.

Just like the currently-existing B50, the B70 has dual media engines. Each has an encoder and decoder for up to two 8K 10-bit workloads. In terms of codec support, Intel’s media engine has acceleration for HEVC 8-, 10-, and 12-bit (decode only) 4:2:0, 4:2:2, and 4:2:0, as well as AV1, among other codecs.

As this is a professional GPU, the Arc Pro B70 offers a number of features not typically seen on consumer parts. This includes ECC memory as well as certified drivers that are tested to offer stability and reliability of performance and results across a variety of products from professional independent software vendors such as Adobe, Autodesk, and Dassault Systèmes.

As mentioned earlier, the Intel Arc Pro B70 boasts a much larger GPU die than the B50, alongside a variety of other changes to the graphics card’s specifications. In the table below, we have listed the most relevant specs for comparable Intel, AMD, and NVIDIA GPUs:

On paper, the B70 is basically just two B50s stapled together: It offers twice the Xe-cores and twice the VRAM… at 2.7x the cost. It seems there is an additional cost for putting that much VRAM on a single card, but it is not large enough to be a deal breaker. There are some other differences as well: Intel doubled the memory bus width to 256-bit and more than doubled the total memory bandwidth from 224 to 608 GB/s. Clock speeds also increased, with the “graphics” clock improving from 1700 MHz to 2280 MHz (the graphics clock is a “typical” frequency achieved in workloads), and the boost clock increased from 2600 MHz to 2800 MHz. As would be expected, these increases all come with commensurate power needs, so the TDP more than tripled, from 70 W to 230 W. As a result of all these changes, Intel claims that the expected FP32 performance is 22.94 TFLOPS and the peak INT8 is 367 TOPS.

There aren’t any great direct competitors for the B70. Intel seems to have purposefully undercut the most likely, AMD’s R9700, by about 30%. Both the R9700 and B70 feature 32 GB of VRAM and similar memory bandwidths, but on paper, the R9700 seems much more competent, with twice the FP32 performance and 5 times the matrix performance — though there is no real consensus on how the latter is calculated. When comparing it to NVIDIA, the B70 is $200 more expensive than the 2000 Blackwell, while offering twice the VRAM and similar FP32 performance. The 4000 Blackwell is the next step up, but costs $550 more than the B70 — and, to match the B70’s 32 GB of VRAM with an NVIDIA GPU, we have to go all the way up to the 2.5x more expensive 4500 Blackwell.

For this review, we will primarily be focusing on a comparison with the R9700, followed by the 2000 Blackwell. We will also note comparisons to the B50 when we want to show how the architecture scales.

As is typical with our GPU reviews, we used an AMD Ryzen™ 9 9950X3D and ASUS ProArt X670E-Creator-based test bed. We have found that, in most workflows we test with, the 9950X3D offers the best all-around performance, and so should minimize potential CPU-side bottlenecks. We used the latest available GPU drivers for this testing, and Windows settings like VBS were kept enabled. We tested using a standard set of professional productivity benchmarks for this review: most of the Puget Bench suite, some offline and real-time renderers like Unreal Engine, Blender, and Unigine, an AI GPU benchmark in MLPerf, and most of our standard Engineering test suite.

The first benchmark we tested with was Lightroom Classic. Overall (Chart #1), we found that the B70 showed 12% uplifts over the B50, though it still ended up slower than any of the other GPUs by a small margin (3% or less). Most of Lightroom is not GPU-accelerated, however, so we didn’t expect to see a huge impact from the GPU on overall performance. AI effects (Chart #2) are often GPU-accelerated, though, so we saw more differentiation there. The B70 maintained its spot second from the bottom, though in those tests it was 27% faster than the B50. The 2000 Blackwell was 7% faster than it, while the R9700 was 26%. We also examined Export times (Chart #3) as those tend to be partially GPU-accelerated. While there was some impact on performance, and the B70 outperformed the R9700, the differences were, overall, small enough not to make a huge impact.

While we wouldn’t recommend a B70 for Lightroom Classic — a 2000 Blackwell is probably the best of those tested, unless using a large amount of AI effects — we wouldn’t normally recommend a professional-class GPU for Lightroom Classic at all.

The next benchmark we tested with was Puget Bench for Premiere. Professional GPUs, while offering advantages in some workflows, typically offer less value for most media and entertainment applications than consumer cards do. Starting with the Overall score (Chart #1), we found that the B70 was 56% faster than the B50 and 8% faster than the 2000 Blackwell. In turn, it was 16% slower than the R9700 and 24% slower than the 4000 Blackwell. Although it had nominally superior media engines for accelerating most LongGOP work, the overall LongGOP score (Chart #2) was dragged down by subpar performance (compared to the R9700 and 4000 Blackwell) in the “real world” tests, placing it solidly in the middle of the pack. The B70 struggled in the RAW media tests (Chart #4), matching only the R9700 but falling well behind NVIDIA’s cards. In GPU Effects (Chart #5), the B70 was similarly middle of the pack, showing more than twice the performance of the B50 but trailing behind the 4000 Blackwell and R9700. Overall, the B70 offers a fine value relative to other professional GPUs for Premiere, but not anything outstanding.

In Adobe After Effects, only one area is properly GPU-accelerated: 3D. Other workflows are almost wholly CPU-based, and so do not scale with GPU. Due to this, we will only examine the 3D score (Chart #3). Our testing found that 3D workflows in After Effects are still highly optimized for NVIDIA GPUs: the B70, though 17% faster than the B50 and 29% faster than the R9700, was 47% slower than the 2000 Blackwell and offered one-third the performance of the 4000 Blackwell. We would recommend end users stick to consumer NVIDIA GPUs for After Effects.