Intel Data Center GPU Flex 170

The Intel Data Center GPU Flex 170 is a dual-slot, 150 W accelerator built on the Xe-HPG architecture, specifically the DG2-512 die fabricated on TSMC's 6 nm process. Released on August 23, 2022, and now marked as end-of-life, it targets data center workloads rather than consumer gaming. Its successor is the H3C Graphics, and its position in the overall GPU landscape is squarely average, sitting at the 50th percentile among all GPUs. The data provided, however, lacks direct benchmark scores and nearest rival comparisons, meaning its performance must be interpreted through its raw specifications, which are substantial for a 150 W part.

Benchmark Performance

The Flex 170's compute potential is defined by its 4096 shading units, 256 texture mapping units, and 128 ROPs. These translate to a peak FP32 throughput of 16.79 TFLOPS, a figure that places it in the upper-mid range for data center accelerators of its generation. In a 2:1 ratio, the FP16 performance doubles to 33.59 TFLOPS, which is a critical metric for AI inference and training workloads where reduced precision is common. The texture fill rate is 524.8 GTexel/s, and the pixel fill rate is 262.4 GPixel/s, indicating a strong balance between geometry processing and fragment shading.

Without nearest rival data or a specific average benchmark score, a relative performance assessment is indirect. The 50th percentile ranking among all GPUs suggests it outperforms half of the hardware in the database, but this is a broad measure. From the architecture, one can infer that its raw compute is roughly in line with a mid-range desktop GPU of the same era, but its data center focus means drivers and software optimizations are tailored for virtualized environments and media processing rather than gaming. The data shows a clear capability for parallel workloads, but the lack of a benchmark score means the practical performance in gaming or rendering tasks is not quantified here. The FP32 rate is the most telling figure, suggesting that in pure compute tasks, it will be competitive with other 150 W-class accelerators, though its end-of-life status implies newer parts have superseded it.

Ray Tracing and Feature Set

The Flex 170 includes 32 dedicated ray tracing cores, aligning it with Intel's Xe-HPG architecture's goal of bringing hardware-accelerated ray tracing to its GPUs. This is a modest count compared to consumer flagship parts from the same era, but it is a functional implementation. The API support is comprehensive: it supports DirectX 12 Ultimate with feature level 12_2, which ensures compatibility with the latest DXR-based ray tracing titles and features like mesh shaders and variable rate shading. OpenGL 4.6 and Vulkan 1.4 are also supported, making it viable for Linux-based workloads and cross-platform applications.

The ray tracing performance is not directly benchmarked in the provided data, but the presence of 32 RT cores suggests it can handle ray-traced effects at lower resolutions or with moderate settings. In a data center context, this is less about gaming and more about rendering tasks, such as offscreen ray-traced scene generation or VDI (virtual desktop infrastructure) for professional users who need hardware RT support. The lack of tensor cores in the specification is notable; for AI workloads, the Flex 170 relies on its FP16 throughput rather than dedicated matrix math units. This means it will be less efficient at certain deep learning operations compared to NVIDIA parts with tensor cores, but the 2:1 FP16 rate is still useful for inference. The display outputs—1x HDMI 2.1 and 3x DisplayPort 2.0—indicate it can drive multiple high-resolution displays, which is atypical for a data center card and suggests a niche for workstation or cloud gaming use cases.

Who Should Consider It

Given the 16 GB GDDR6 memory and 512.0 GB/s bandwidth, the Flex 170 is best suited for workloads that require large frame buffers but not extreme memory bandwidth. The 256-bit bus width and 16 Gbps effective memory clock yield that 512.0 GB/s figure, which is sufficient for 1440p or 4K rendering with high-resolution textures. The FP32 compute is strong for its power class, making it a candidate for batch rendering, virtualized workstations, and media transcoding where the 150 W TDP is an advantage for density.

For gaming-specific scenarios, the data is ambiguous. The 50th percentile ranking and lack of benchmarks mean it is not clearly a high-end gaming card. However, the 4096 shading units and 16.79 TFLOPS FP32 rate suggest it can handle modern titles at 1080p or 1440p with high settings, though ray tracing will be a performance hit given the modest 32 RT cores. The 16 GB VRAM is a future-proofing asset for texture-heavy games, but the card's data center heritage means driver overhead and latency may not be optimized for interactive gaming. This is better suited for cloud gaming instances or professional 3D modeling, where the dual-slot form factor and 1x 8-pin power connector make it easy to deploy in servers. The 50th percentile position indicates it is not a top-tier performer, so users expecting flagship frame rates should look elsewhere; it is a workhorse for compute throughput rather than latency-sensitive tasks.

How It Compares

The FACT PACK lists no nearest rivals, which is a significant gap. Without delta percentages or scores for competing products, a direct comparison cannot be made from the provided data. The only comparative metric is the 50th percentile ranking among all GPUs, which offers a general sense of standing. In the absence of specific rivals, the comparison must be qualitative: versus NVIDIA's data center offerings of the same generation, the Flex 170 lacks tensor cores, which would handicap it in AI inference where FP16 with tensor acceleration is common. Against AMD's CDNA-based accelerators, the Flex 170's 16 GB VRAM and 512 GB/s bandwidth are modest, but its 150 W TDP is significantly lower, enabling higher server density. The architecture's Xe-HPG roots also mean it shares features with Intel's consumer Arc GPUs, so driver maturity is a known variable. The data shows an end-of-life product with no successor information beyond the H3C Graphics, which suggests it has been superseded, but the lack of rival data prevents a definitive positioning statement.

Power and Cooling

The Flex 170 has a thermal design power (TDP) of 150 W, which is remarkably low for a GPU with 4096 shading units and 16 GB of VRAM. This power efficiency is a key selling point for data center deployments where power density and cooling are critical constraints. The suggested power supply unit (PSU) for a system containing this card is 450 W, which accounts for the rest of the system's draw. The power connector requirement is a single 8-pin, which is standard and widely compatible with existing server PSUs. The dual-slot form factor means it occupies two PCIe slots for cooling, but the low TDP allows for a simpler heatsink design, likely a passive or low-speed fan solution, though the details are not specified. The PCIe 4.0 x16 bus interface ensures adequate bandwidth for data transfer to and from the host system, which is essential for compute workloads. The low power draw also means it can be paired with high-core-count CPUs without exceeding typical server power budgets, making it an attractive option for dense, multi-GPU nodes.

FAQ

Q: What is the memory configuration of the Intel Data Center GPU Flex 170?

A: It features 16 GB of GDDR6 memory on a 256-bit bus, yielding a bandwidth of 512.0 GB/s. The memory clock is 2000 MHz with a 16 Gbps effective rate.

Q: Does the Flex 170 support hardware ray tracing?

A: Yes, it includes 32 dedicated ray tracing cores, and its API support for DirectX 12 Ultimate (feature level 12_2), Vulkan 1.4, and OpenGL 4.6 enables ray-traced workloads in compatible applications.

Q: What is the power requirement for this card?

A: The TDP is 150 W, and the system requires a suggested PSU of 450 W. It uses a single 8-pin power connector.

Q: What is the compute performance in FP32 and FP16?

A: The FP32 throughput is 16.79 TFLOPS, and the FP16 throughput is 33.59 TFLOPS, achieved via a 2:1 ratio.

Q: Is this card suitable for gaming?

A: The data does not provide gaming benchmarks. With 4096 shading units, it has the raw compute for 1080p or 1440p gaming, but its data center focus and 50th percentile ranking suggest it is not optimized for interactive frame rates. It is better suited for rendering or compute tasks.

Q: What is the production status and release date?

A: The production status is end-of-life, and it was released on August 23, 2022. Its successor is the H3C Graphics.

Memory Subsystem

The memory subsystem is a critical component of the Flex 170's performance profile. It is equipped with 16 GB of GDDR6 memory, which is a substantial amount for a 150 W card, allowing for large datasets and high-resolution textures to reside on the GPU without spilling to system memory. The memory operates at 2000 MHz with a 16 Gbps effective data rate, and it is connected via a 256-bit bus. This configuration produces a peak bandwidth of 512.0 GB/s.

For high-resolution workloads, such as 4K rendering or large language model inference, the 16 GB capacity is more important than the bandwidth. 512.0 GB/s is adequate for most tasks, but it is not exceptional; modern high-end GPUs often exceed 1 TB/s. However, the low TDP means this bandwidth is achieved at a high efficiency per watt. The 256-bit bus width is a balanced choice, providing enough parallelism to feed the 4096 shading units without excessive power draw. In practice, the memory subsystem can handle 1440p and 4K textures with ease, and the 16 GB capacity ensures that multi-frame rendering or batch processing will not be bottlenecked by memory capacity. The lack of tensor cores means that some AI workloads may be memory-bound rather than compute-bound, but the 512.0 GB/s rate is sufficient for inference tasks that are not excessively data-hungry. The data shows a well-rounded memory design that prioritizes capacity and efficiency over raw speed.