Intel Xeon Phi 7120P

Manufactured by Intel, the Xeon Phi 7120P is a Knights Corner x100 accelerator built on a 22 nm process and using Intel as the foundry. It packages 5,000 million transistors onto a 720 mm² die, a transistor density of 6.9M/mm². Released on 2013-06-16, the part is end-of-life, and its launch MSRP was 4,129 USD. The board connects through PCIe 3.0 x16 and has no display outputs. Its base clock is 1238 MHz, rising to 1333 MHz boost, while the 16 GB GDDR5 memory runs at 1375 MHz with 5.5 Gbps effective speed. The record shows an empty benchmarks array and an empty nearestRivals array, with only a 50th percentile placement among all GPUs.

Memory Subsystem

The 7120P has 16 GB of GDDR5 memory on a 512-bit bus. The effective memory speed is 5.5 Gbps, and the resulting bandwidth is 352.0 GB/s. A 512-bit bus is the structure that lets the memory controller pull enough data from the GDDR5 chips to reach that aggregate figure; the 1375 MHz memory clock alone would not explain it. Wide buses exist to move large blocks of data in parallel, and 352.0 GB/s is the listed result.

High-resolution graphics typically place heavy demands on memory bandwidth. This product, however, has no display outputs, zero ROPs, and a pixel rate of 0 MPixel/s. There is no pixel output stage to send frames to a display, so the memory subsystem cannot be evaluated in terms of high-resolution gaming or framebuffer settings. It is better characterized as a large, fast pool for data that must be accessible to the compute units.

The 1375 MHz memory clock is the base operating frequency reported for the memory, and the 5.5 Gbps effective rate describes per-pin transfer speed. With 16 GB of capacity, the accelerator can hold substantial datasets on-board, avoiding repeated transfers over the PCIe 3.0 x16 host link. That host link is the only listed connection to the rest of the system, since no display outputs exist. For any workload where moving large amounts of data in and out of a 352.0 GB/s memory pool matters, this configuration is directed at that kind of operation rather than at producing images.

Ray Tracing and Feature Set

The database lists no ray tracing cores and no tensor cores. It also lists no DirectX, OpenGL, or Vulkan support. Those null fields define the product’s feature set as compute-oriented rather than graphics-oriented.

The execution resources present include 976 shading units, 32 texture mapping units, and no ROPs. That configuration yields 42.66 GTexel/s of texture rate and 2.602 TFLOPS of FP32 throughput. The pixel rate is 0 MPixel/s, consistent with the zero ROP count and the lack of display outputs. No dedicated ray tracing hardware is listed, so any ray traversal would have to rely on general compute units. No tensor cores are listed, so no dedicated matrix acceleration block is present in the data. No FP16 figure is listed either, leaving half-precision throughput undefined.

The API fields are null for DirectX, OpenGL, and Vulkan, so no graphics API version claims can be made from this record. The listed host connection is PCIe 3.0 x16, and the absence of display outputs closes off traditional graphics output. In practical terms, a system integrator sees an accelerator with a wide memory interface and substantial FP32 / texture resources, but without the usual graphics features expected from a consumer GPU.

Benchmark Performance

The benchmarks array is empty and the average benchmark score is 0. Because of that, there are no measured scores to analyze. The nearestRivals array is empty, so the database provides no competitor names, scores, or deltaPct values for direct comparisons.

The one relative field is percentileVsAllGpus, set to 50. That places the product at the middle of the global GPU distribution in this database. It is a ranking point, not a benchmark score. With an average score of 0 and no benchmark rows, the percentile cannot be tied to a specific rival. The data offers no basis for statements such as “ahead by X percent” or “behind by Y percent.”

The raw capability limits are reported in the specification fields: 2.602 TFLOPS FP32, 42.66 GTexel/s, and 352.0 GB/s of bandwidth. These are not benchmark outputs. The base clock of 1238 MHz and boost clock of 1333 MHz bound the execution rate, but no measured workload results are present. Without actual scores, a performance ranking cannot be derived from benchmark data, only from the hardware’s listed peak figures.

How It Compares

nearestRivals is empty. There are no rival entries to describe, so no one-paragraph-per-rival comparisons can be written. The predecessor and successor fields name Knights Ferry and H3C Graphics, but those are not listed as nearest rivals, and no scores are attached to them in this record.

The only comparative position in the data is the 50th percentile among all GPUs. That is a middle-of-the-pack placement. Without rival scores or deltaPct values, the 7120P cannot be said to be ahead of or behind any specific product in this database. The absence of nearestRivals is itself a finding: the database does not classify any GPU as close enough to this part for direct comparison.

Who Should Consider It

Because it has no display outputs, a 0 MPixel/s pixel rate, and zero ROPs, the 7120P should not be considered for driving a monitor or for high-resolution gaming settings. It is an accelerator with 16 GB of GDDR5 memory and 352.0 GB/s of bandwidth, so compute workloads that need a large on-board working set are the relevant use case.

The compute resources are 976 shading units and 32 texture mapping units, producing 2.602 TFLOPS of FP32 throughput and 42.66 GTexel/s texture rate. A PCIe 3.0 x16 connection ties it to the host. Since no DirectX, OpenGL, or Vulkan support is listed, a standard graphics API path is not indicated by the record. The intended use is better understood as host-driven compute over the PCIe bus rather than game-engine rendering.

The product is end-of-life, so availability is limited by that status. The 50th percentile global placement indicates neither a top-tier nor bottom-tier classification in the database. Resolution and settings recommendations do not apply, because the part cannot output pixels. Instead, the practical considerations are memory capacity, bandwidth, FP32 throughput, and the absence of display hardware.

Power and Cooling

The TDP is 300 W, and the suggested power supply is 700 W. The board is dual-slot in width and 248 mm / 9.8 inches long. No power connector requirements are listed in the data.

A 300 W TDP is the thermal target the cooling solution must handle, and the dual-slot design is the listed physical cooling form factor. The 248 mm length places a minimum on chassis clearance. The bus interface is PCIe 3.0 x16, so the accelerator fits into a standard x16 slot. The database gives no information about additional power connector counts, so this analysis cannot specify them. The 700 W suggested PSU is the only system-level power guidance supplied.

The combination of a 300 W TDP, dual-slot width, and 248 mm / 9.8 inch length gives an integrator a clear spatial footprint. The absence of display outputs means there is no monitor-side power draw to consider. Power and cooling guidance in the data is limited to those figures, but they are sufficient to plan around the physical and electrical requirements.