Product Review: Intel Arc Pro B50 GPU for Workstations

THIS IS OUR FIRST TIME REVIEWING AN INTEL DISCREET professional-level GPU—the brand-new Intel Arc PRO B50. Powered by a new Xᵉ2 Architecture with large amounts of GDDR6 memory, the new Pro B50 offers a compelling entry-level or lower mid-tier workstation solution for computer graphics in CAD and 3D industries.

This new Intel GPU also delivers high performance on AI workloads as well as professional workstation 3D computer graphics. As AI increasingly infuses our common workloads, the Xᵉ2 GPUs on the Intel card feature dedicated AI hardware and matrix engines, known as Intel Xᵉ Matrix Extension (XMX). Each Vector Engine is paired with an XMX AI engine capable of 2048 FP16 ops/clock and 4096 INT8 ops/clock.

General Introduction

For our Architosh readers, we have focused on AEC/O application workloads rather than AI performance workloads. Readers are advised to refer to this and this for additional information on the B50’s performance characteristics for AI workloads.

We reviewed our Intel Arc PRO B50 on our trusty BOXX APPEX E3 workstation running the Windows 10 Professional operating system. This is the same exact rig and setup as all our past GPU card reviews, dating back to the AMD Radeon Pro W6600 GPU in the fall of 2021. The Intel GPU is priced at 349.00 USD (suggested retail price).

For purposes of direct comparison, we also tested an NVIDIA RTX A1000 GPU, which sells for between 404 and 508.00 USD (in other words, similarly priced). We will use these prices (using the lower number) for the purposes of our GPU economics charts and calculations (see page 3). In other words, when we compute performance on a per-dollar basis between the Intel Arc PRO B50 and the NVIDIA RTX A1000 reference card, we will use 404 USD for the calculations on the Nvidia side.

GPU Details: Intel Arc PRO B50

The B50 is a small form-factor (SFF) dual-slot GPU card that boasts 16 GB of GDDR6 memory, which is twice the amount provided by the Nvidia RTX A1000 reference GPU. Built on the new Xᵉ2 Architecture, the BMG-G21 GPU provides excellent performance across AAA games, workstation graphics, and running ML/AI workloads. Key details include:

• 128 Xᵉ Vector Engines
• 128 XMX Matric Engines
• 16 Ray Tracing Units
• 170 TOPS (INT8)
• TSMC N5 node (19.6 B / 272mm2)

 

• PCI Express x8 Gen5 (x16 Physical)
• OpenGL 4.6 / Vulkan 1.4 / DirectX 12 Ultimate / OpenCL 3.0
• 16 GB GDDR6
• 224 GB/s memory bandwidth
• 4 Displays (up to 8K120 HDR) (DisplayPort 2.1) (4X 4K120 HDR)
• Up to 7680×4320 @ 60Hz (DisplayPort 2.1)
• 70W max power consumption (no auxiliary power needed)
• SFF, Dual Slot
• VR Ready

The Intel BMG-GPU SoC (System on a Chip) is illustrated in the image at the bottom of the page. Click on the image for a larger view to see details of the chip design.

New Design and Chip Architecture

The Intel Arc PRO B50 is a discrete GPU card built by Intel, and the famed semiconductor company says it has taken feedback from its A-series customers and redesigned the platform and thermal solution, improving efficiency and acoustics. In our own testing, we were impressed with how quiet the unit was under multiple testing scenarios and general usage.

The actual GPU card is striking visually, with Intel’s signature blue being incorporated as accents and on an entire side of the CPU. As for the new compute architecture, the updated second-generation Xᵉ2 Architecture features a wider SIMD16 width for improved efficiency and compatibility, as well as a 33% larger L1 cache with a total capacity of 256 KB. The Ray Tracing Units are also second generation in this B50 GPU, and the benchmarks and visuals during testing attest to the strength of Intel’s GPU solution.

In fact, for AEC workloads, which we focus on here at Architosh, the rendering performance of the Intel Arc PRO B50 makes this GPU a surprisingly good choice for lower-tier workstations. But we are jumping ahead of ourselves, but you will see it in the benchmarks we provide below (we do all of our testing ourselves).

Since benchmark performance is where the rubber meets the road, let’s go to the next page to see how Intel’s latest professional GPU card stacks up against its primary Nvidia competitor and even some previously reviewed GPU cards and SoCs.

next page: Benchmarking and Performance Results

Benchmarking

At Architosh, we select a diverse range of benchmarks, including some that are cross-platform. We are especially interested in having readers understand the trade-offs and benefits of discrete GPUs versus GPUs included on a System on a Chip (SoC) design, and between the Windows and the Mac world of computers. That being said, many of the best benchmarks for showing off the strength of a GPU don’t work across platforms, and we focus on those as well.

Our benchmark suite for Architosh continues to evolve and, for this review, includes some updates. Here are the benchmarks:

• Geekbench 6 GPU  — Comprehensive general GPU benchmark
• BlenderMark — Very good for ray-tracing testing
• CompuBench —  Catmull-Clark and SSS testing
• VRMark (Orange Room) — for testing VR headset capabilities
• SPECviewperf  — for shaded CAD/BIM/Rendering testing

Sadly, the highly regarded Maxon Cinebench does not support the Intel Arc PRO B50. However, the VRMark and Enscape tests inside of SPECviewperf V15 can serve to illustrate rendering capabilities.

For this review, we have migrated from SPECviewperf 2020.v3 to SPECviewperf v.15 for testing. The new v.15 results include new tests, such as the Enscape benchmark. Enscape is a highly popular rendering tool in architecture workflows.

Benchmarks

Geekbench 6 GPU

Geekbench 6 is one of the world’s most touted and trusted benchmark applications for CPUs, SoCs, and now GPUs. As a GPU-focused benchmark, Geekbench 6 focuses on GPU compute performance using workloads more applicable to generalized computing rather than CAD, BIM, and 3D applications. These are still meaningful because the average architect, for example, has meaningful workloads that span the gamut from blur effects in Zoom and Teams calls, working with image editors and Adobe products, and ML (machine learning) and AI workloads, which Geekbench 6 GPU specifically tests.

In Geekbench 6 GPU, the subtests remain consistent, but the APIs utilized by the tests are different. The more important API is Vulkan, as it is newer and a lower-level API designed to replace OpenGL. In contrast, OpenCL is a more mature, higher-level API for general-purpose parallel computing workloads across both CPUs and GPUs. Regardless, what we see is that the Intel GPU is considerably faster than its Nvidia reference GPU.

For the OpenCL API, the Intel Arc PRO B50 is 31% faster, while for Vulkan, the B50 is 32% faster. This delta applies to more generalized GPU workloads involving machine learning (AI), image synthesis and simulation functions, as noted above.

Blendermark GPU

Now, diving into a professional 3D software program like Blender, the Blendermark performance advantage of the Intel B50 over the RTX A1000 reference GPU comes in at just 26%, which we think somewhat belies the Intel GPU’s rendering performance advantages as seen in other rendering benchmarks. (see below).

While the Blendermark benchmark gives you a specific set of tests that applies to that application, the next series of tests gets more specific to a wider range of tools.

CompuBench 

CompuBench helps us see specialized GPU compute performance, particularly for professional and digital content creation (DCC) workflows.

The Catmull-Clark tests in CompuBench evaluate the GPU’s ability to handle surface subdivision, which creates smooth, curved surfaces from a coarse polygon mesh. The algorithm is found in tools like Maya, Blender, and Pixar’s OpenSubdiv libraries. The Intel B50 scored not only much better than the Nvidia RTX A1000 but also the A2000 GPU. On the Subsurface scattering test, it performed 83% faster, while on the Catmull-Clark test, it performed 27% faster.

Subsurface Scattering (SSS) simulates how light penetrates a translucent object, scatters beneath the surface, and then exits at a different point. This effect is critical for rendering realistic materials like human skin, marble, and wax. All of this is applicable to rendering workflows across a range of applications. However, CompuBench has a limitation in that it primarily relies on OpenCL for these tests, whereas newer APIs like Vulkan and CUDA are often used in the latest versions of DCC software tools.

Regardless, CompuBench is meaningful for evaluating cross-platform comparisons of a GPU’s parallel compute performance using an open standard (OpenCL) to test common and meaningful algorithms used in DCC and CAD tools. This way, GPUs from Intel and AMD can be compared to Nvidia, which prioritizes CUDA as an API for its own graphics chips.

VRMark (Orange Room) Test

VRMark is a virtual reality benchmark that allows one to test a GPU’s likely performance in driving VR workflows. The benchmark has many different tests within its suite, but we stick with the Orange Room test.

The Intel Arc PRO B50 delivered a score that was 72% better than the reference RTX A1000 GPU and 46% faster than the RTX A2000 GPU.

SPECviewperf v15 — Creo Composite and Shaded with Edge Subtests

We can see from the Subsurface Scattering test and the VRMark score that the Intel B50 GPU seems to shine on rendering workloads, and this will be consistent with its Enscape subtests in the new SPECviewperf v15 benchmark. Turning to the Creo Composite and subtests below, we now focus on OpenGL workloads so common in the MCAD world and with many 3D tools in AEC like Rhino and SketchUp. These tools are beginning to transition to low-level APIs like Direct3D and Vulkan. Rhino 9 will be primarily based on Direct3D, whereas the current versions, 7 and 8, rely mostly on OpenGL.

In our Creo Composite test, we can see that the Intel Arc PRO B50 outperformed the reference Nvidia RTX A1000 GPU by 39%. This is a significant beat, but when we dig into the many types of subtests, we see illuminating differences. On tests where AA (anti-aliasing) was turned on, the performance gain over the RTX A1000 was less significant. We saw just a 13% better score over the RTX A1000 on the Creo World Car Shaded with Edges and 4xAA test. However, on the Scorpion without the AA subtest, we observed nearly twice the advantage.

The Creo “Shaded with Edges” subtests composite score above mixes a blend between two subtests we have focused on in the past, with several reviews (though in the previous version of SPECviewperf). With this blend, we see the Intel B50 performing 15% faster, which is more modest. Still, one must consider GPU economics evaluations, which we do on the next page. This changes the story significantly

SPECviewperf v15 — Enscape Composite

Finally, we were delighted to see an Enscape set of tests added to SPECviewperf v.15. Enscape is the most dominant of the interactive, real-time rendering tools in the AEC industry. It competes head-on with Twinmotion and others, but is the most popular choice.

The Intel Arc PRO B50 performed 46% faster than the RTX A1000 GPU. Though we didn’t test it, but could have, as we keep most GPUs sent to us for evaluation, the B50 may have even bested the RTX A2000, though this is speculation. We suspect they would be close.

Performance Results

Across our benchmarks, the Intel Arc PRO B50 outperformed the Nvidia RTX A1000 reference GPU by a wide margin, spanning performance deltas between 15% on OpenGL-based Creo composite tests to up to 100% faster on specific Creo OpenGL shaded with edges model viewports without AA. In general, the Intel GPU was significantly faster on rendering workflows like the SPECviewperf v.15 Enscape tests, where it was 46% faster, which is notable.

What is perhaps most significant is that the Intel Arc PRO B50 is very affordably priced. As we noted in our last Nvidia GPU review, Nvidia’s GPUs have tended to become more expensive as GPU consumption in the era of AI has become massive. That’s where economic metrics come into play (next page).

next page: Economic Metrics & Conclusions

Economics Metrics

GPU Economics

Our GPU economics charts are partial and selective, but hopefully informative. This kind of metric allows you to compare the value proposition of different chips, especially in isolation, to a specific application or type of workflow. However, one must keep in mind that benchmark scores alone do not tell the whole story of “value” or “performance,” as overall system performance impacts GPU performance. Furthermore, workstation-class GPUs offer other tangible and intangible benefits, including the certification of professional applications (which the Intel B50 has).

Since we have numerous benchmarks to test, which ones do you select for the performance-per-dollar calculation and comparison? For this review—and for the AEC audience on Architosh—we have selected three:

• (1) Creo (OpenGL-based) Composite Test 
• (2) Geekbench GPU (composite OpenCL and Vulkan) Tests
• (3) SPECviewperf v.15 Enscape Sub-Test

These three provide an optimal balance between general computing, ML/AI, OpenGL-based rendered viewport performance common to BIM and 3D modeling, and rendering performance (especially interactive).

About Computation

Unlike our last GPU review article, as noted in the GPU Economics section on page one, we are comparing GPUs on two different metrics that are really just inverses of each other. In this first Geekbench 6 GPU Composite chart, we are looking at the Compute Unit (CU) achieved per USD. For every dollar you spend on either of these GPUs, what CU do you achieve on the Geekbench 6 GPU benchmark composite scores? The calculation looks like this:

 

Compute Unit (CU) per USD =  benchmark score / cost of the GPU

 

In looking at the “value” delivered by the Intel Arc PRO B50, for overall GPU performance across a variety of workloads, including machine learning (ML) for AI, we can see that the B50 delivers outstanding value against its nearest Nvidia RTX rival, the A1000 GPU. In the chart below, longer is better, and one USD earns you the following score value on the composite scores of the Geekbench 6 GPU benchmark. On a performance-per-USD basis, the Intel GPU is 48% better across general GPU compute.

For the next two GPU economics charts, we focus on OpenGL performance (or similar lower-level API shaded with edges graphics with and without AA (anti-aliasing) and interactive and photorealistic rendering graphics. In modern BIM and 3D modeling workflows, from Rhino and SketchUp to all the major BIM authoring tools, this combination of graphics rendering is predominant.

In the next two charts, we look at the cost in USD per CU achieved. This inverts the formula above and is identical to the GPU econ charts in our prior reviews, such as here. The formula is:

 

Cost in USD per Compute Unit (CU) = cost of GPU / benchmark score

 

On a cost-value basis, the Intel Arc PRO B50 delivers stunning value compared to the RTX A1000. For 62 cents on the dollar, the Intel B50 delivers superior performance. Recall that the Creo Composite performance is 39% better than the Nvidia RTX A1000. When we consider the cost difference in conjunction with this OpenGL performance, the Intel GPU becomes very compelling.

And this compelling value is also delivered when we examine rendering workflows. Recall that the Intel Arc PRO B50 even outperforms the big sister of the A1000, the Nvidia RTX A2000, on the CompuBench SubD (Catmull-Clark) and SSC (sub-surface scattering).

And it overtakes the RTX A1000 by a bigger margin than the RTX A1000 GPU overtakes the Apple M3 SoC, or by more than 35%.

The Value Proposition

The value of the Intel Arc PRO B50 for MCAD and OpenGL-based workflows is quite notable compared to the competition referenced in this review, specifically the Nvidia RTX A1000. For example, CDW sells the RTX A1000 for 507.99 USD, while Walmart has it for 421 USD, and PNY has an RTX A1000 that sells for 403.99 USD. As such, we have used a price of 404.00 USD for our GPU economics charts.

As we can see from our GPU economics section, Intel is delivering superb value in this entry-level workstation-class GPU.

Conclusions

Increasingly in AEC workflows, the role of interactive, real-time rendering is significant. Design workflows today are highly 3D-based, and having a 3D or BIM model early in the design process provides not only interactive rendering options but also access to AI-generated imagery that is highly specific to your design process. We were delighted in this review to have access to SPECviewperf v.15’s new Enscape composite tests. Although it was unfortunate that we could not run Cinebench on this GPU card due to compatibility issues, the Blendermark and VRMark performance results also strongly support the Intel Arc PRO B50’s 3D rendering capabilities. One must also remember that this GPU comes with 16 GB of onboard memory.

The one weakness we seemed to detect in the Intel Arc PRO B50 is the performance of OpenGL workflows with AA turned on. That is, anti-aliasing seemed to have a larger impact on the SPECviewperf v.15 Creo suite of tests than we would have preferred to see, though it was still minor in the grand scheme of things. Although we are not primarily focused on MCAD tests for this review, the B50 appears to be a smart investment for economy-oriented MCAD workstations.

We used an “average low price” for the Nvidia RTX A1000 in our GPU econ charts, but the results would not be far off if we equated the two at the “same price.” As for our benchmark tests and those of Intel, they are close on several items, including Enscape and Creo SPECviewperf tests. Besides this GPU’s excellent performance for the price, other strong aspects of the Intel Arc PRO B50 include its 16 GB of GDDR6 memory. This was twice the memory of the Nvidia RTX A1000 (Ampere generation) reference GPU, and becomes meaningful for detailed ray tracing rendering jobs and managing larger textured architectural scenes.

As AI infuses more of our common workloads, the Xᵉ2 GPUs on the Intel card have dedicated AI hardware and matrix engines called Intel Xᵉ Matrix Extension or XMX, and they appear to perform very well on AI workloads like inference. This is another big bonus for this specific Intel GPU.

Recommendations

As much as we love how Nvidia has pushed the limits of GPU performance over the recent years, the CAD market needs competition in the GPU space. It is wonderful to see Intel’s new, professionally oriented Arc Pro-B-Series graphics cards truly deliver on performance-per-dollar metrics over the competition. We tested the B50, but it has a bigger sibling in the B60 with up to 24GB of memory and 20 Ray Tracing Units, not 16 like the B50. We have to give the Intel Arc PRO B50 a big thumbs up. This makes a superb lower-to-mid level GPU for CAD professionals. You can learn more here at Intel.

---

 

Pros: This GPU option offers powerful performance per dollar, and even if you could find the Nvidia RTX A1000 at the same or slightly lower price, the Intel Arc PRO B50 has significant performance advantages over it in areas like rendering (charts 3 [Blender], 4, 5, and 9 [Enscape]). It is even faster than the big sister to the RTX A1000, the RTX A2000, in the VRMark (Virtual Reality) test, and in CompuBench rendering performance and subsurface division performance. A significant benefit of this GPU is its 16 GB of GDDR6 memory, as well as its thermal design and power consumption, which require only 70 watts. It powers itself via the PCIe slot. We are also impressed with its OpenGL performance, although a bit less than in rendering workflows. Finally, we can’t deny that Intel’s Xᵉ2 Architecture with its AI hardware and matrix engines delivers compelling ML/AI workload performance over rival GPUs. The Intel Arc Pro B-Series GPUs have ISV certifications, more info here about that. 

Cons: The one noticeable performance letdown with the Intel Arc PRO B50 occurred during the SPECviewperf v.15 Creo Composite benchmarks. Overall, the OpenGL performance of the B50 is solid but not as convincingly better over its Nvidia rival, and when AA is turned on, the gap closes a wee bit further. We can see this in SPECviewperf v.15 “Shaded w/ Edges” Creo subtests (Chart 8), where the World Car Shaded with Edges, 4xAA performance is just 13% faster, whereas the composite, which includes a subtest with AA turned off, averages at 15% faster. The only other con is that Cinebench is not compatible with this GPU, so it prevents us from using one of the benchmark mainstays for GPU testing. Perhaps Intel can work on that issue. 

Advice: For CAD industry professionals with visualization workloads, the Intel Arc PRO B50 offers a very compelling option, especially when configuring a workstation on a tighter budget or when prioritizing CPU performance over GPU performance, which is what you would do, for example, if your main application and workflow is Autodesk Revit at over 70% of your daily workflow. For large architecture firms outfitting dozens of workstations for workflows that mainly consist of Rhino, SketchUp, Revit, and AutoCAD, this is the GPU I would be considering while leveraging and prioritizing a more expensive GPU. 

Cost: 349 USD MSRP

 

Volume of New Content = 5 — Intel’s Arc PRO B50 offers very good performance for the dollar CAD and 3D rendering graphics performance, offering a compelling mixture of shaded with edges (OpenGL, etc), viewport rendering, and advanced photorealistic 3D rendering and animation capabilities.

Quality of Execution = 4.5 — The Intel GPU card is extremely well-made and executed in an attractive and quiet design. The installation of both hardware and software was simple and well-planned. The Intel graphics driver software was also quite strong, though we spent no time in the article discussing it.

Underlying Technologies = 4.5 — The new Intel chip is built on a TSMC 4nm “N5 node, packs 19.6B transistors, and comes in an all-new card design with improved thermal design and quieter fan performance. The Xᵉ2 Architecture delivers performance across the board with particular strength in rendering and DCC tools like Maya, Blender, Enscape in AEC. It also boasts impressive ML/AI performance.

Future Proofing = 4.5 — The GPU is on the leading edge of graphics technologies, including GDDR6 memory, PCIe Gen 4 or Gen5x8 interface, fast memory bandwidth, and state-of-the-art ML-optimized AI engines. It also ships with 16 GB of memory and excellent video encoding/decoding technology.