AMD Radeon RX 640 OEM

The AMD Radeon RX 640 OEM is a Polaris 23 part built on GCN 4.0 and fabricated by GlobalFoundries on a 14 nm process. The die measures 103 mm² and holds 2,200 million transistors, a density of 21.4 million per square millimeter. Compute resources include 640 shading units, 40 texture mapping units, and 16 ROPs, producing 1.658 TFLOPS of FP32 and the same 1.658 TFLOPS for FP16 at a 1:1 ratio. Memory is 4 GB of GDDR5 on a 128-bit bus, clocked at 1500 MHz (6 Gbps effective), for 96.00 GB/s of bandwidth. The card occupies a single slot, draws 50 W, requires no auxiliary power connectors, and pairs with a 250 W suggested PSU. It was released on 2020-04-08 and is now end-of-life, sitting at the 50th percentile of all GPUs in the database.

How It Compares

The nearestRivals array in the database is empty, so there are no named competitors with scores or deltaPct values to cite. The only positioning data available is the percentileVsAllGpus figure of 50, which places the RX 640 OEM exactly at the median of the entire GPU population tracked by the database. Half of all GPUs score higher, half score lower. That mid-pack standing is consistent with the internal specifications: 1.658 TFLOPS of FP32 compute and 96.00 GB/s of memory bandwidth are modest by modern standards but not bottom-tier. The 4 GB GDDR5 frame buffer on a 128-bit bus is typical of an entry-level part from the era of its 2020-04-08 release.

Without rival entries, the comparison must be drawn from the spec sheet rather than measured deltas. The pixel rate of 20.72 GPixel/s and texture rate of 51.80 GTexel/s define the fill-rate ceiling, and these figures align with a card that would slot between low-end integrated graphics and mid-range discrete GPUs in the aggregate database distribution. The absence of rival data means no percentage deltas can be reported; the percentile is the sole comparative metric. The 50th percentile ranking is a neutral verdict — this card is neither a standout performer nor a laggard in the database's full GPU population.

Ray Tracing and Feature Set

The RX 640 OEM has no RT cores and no tensor cores listed in the database. Its GCN 4.0 architecture predates dedicated ray tracing hardware, so hardware-accelerated ray tracing is not part of the feature set. API support is DirectX 12 (12_0), OpenGL 4.6, and Vulkan 1.3. The DirectX 12 feature level 12_0 covers the base feature set but without the DXR ray tracing tier that requires dedicated hardware. Vulkan 1.3 support is current and allows modern compute and graphics workloads, but again without ray tracing extensions that rely on RT cores.

FP16 compute runs at 1.658 TFLOPS, identical to FP32 at a 1:1 ratio, which means there is no half-precision acceleration advantage — a notable distinction from architectures that double FP16 throughput. The display outputs are one DisplayPort 1.4a and two mini-DisplayPort 1.4a ports, supporting modern monitors through the DisplayPort 1.4a standard. The pixel rate of 20.72 GPixel/s and texture rate of 51.80 GTexel/s are the throughput ceilings for rasterization. With 640 shading units and 16 ROPs, the card is built for conventional raster workloads, not for compute-heavy or ray-traced rendering. The 4 GB GDDR5 memory with 96.00 GB/s bandwidth further limits advanced features like high-resolution texture packs or ray-traced assets. In short, the feature set is firmly rooted in the GCN 4.0 era, with no path to hardware ray tracing.

Power and Cooling

The RX 640 OEM is a 50 W part, which is the TDP listed in the database. It requires no auxiliary power connectors — the card draws all its power from the PCIe slot. The suggested power supply is 250 W, a very modest requirement that makes the card compatible with a wide range of pre-built systems and small form factor cases. The cooler is single-slot, meaning it takes up only one expansion slot. The 14 nm process from GlobalFoundries and the 103 mm² die with 2,200 million transistors contribute to the low power draw. The transistor density of 21.4 million per square millimeter is a reflection of the mature 14 nm node.

With no power connectors, installation is straightforward: the card plugs into a PCIe 3.0 x8 slot and provides display output via one DisplayPort 1.4a and two mini-DisplayPort 1.4a ports. The 250 W PSU recommendation leaves headroom for a typical CPU and drives in a basic system, but the card's own consumption is only 50 W. The single-slot cooler must dissipate that 50 W, which is well within the capability of a passive or low-profile active cooler. The end-of-life production status means availability is limited to existing stock, but the power characteristics remain relevant for legacy system upgrades.

FAQ

Q: Does the Radeon RX 640 OEM support hardware ray tracing?

A: No. The database lists no RT cores for this card, and its GCN 4.0 architecture does not include dedicated ray tracing hardware. It supports DirectX 12 (12_0), OpenGL 4.6, and Vulkan 1.3, but ray tracing acceleration is not part of the feature set.

Q: What power supply is recommended for this card?

A: The suggested PSU is 250 W. The card itself has a TDP of 50 W and requires no auxiliary power connectors, so it draws power entirely from the PCIe slot.

Q: How much memory does the RX 640 OEM have, and what is its bandwidth?

A: It has 4 GB of GDDR5 on a 128-bit bus, running at 1500 MHz (6 Gbps effective), providing 96.00 GB/s of bandwidth.

Q: What display outputs are available?

A: The card has one DisplayPort 1.4a and two mini-DisplayPort 1.4a outputs.

Q: Is FP16 performance faster than FP32 on this card?

A: No. FP16 runs at 1.658 TFLOPS, which is a 1:1 ratio with FP32 (also 1.658 TFLOPS). There is no half-precision speedup.

Q: When was this card released, and is it still in production?

A: It was released on 2020-04-08 and is now end-of-life. Its predecessor is Polaris and its successor is Vega.

Who Should Consider It

The 50th percentile ranking and the spec sheet point to a card for 1080p gaming at modest settings. The 1.658 TFLOPS FP32 compute and 20.72 GPixel/s pixel rate are sufficient for older titles, e-sports, and less demanding games at 1080p with reduced detail. The 4 GB GDDR5 frame buffer and 96.00 GB/s bandwidth mean texture-heavy modern games will exceed the memory capacity at high settings, so users should plan for medium or low presets. The 50 W TDP and 250 W suggested PSU make this a candidate for office desktops, HTPCs, and small form factor systems where power and space are limited.

The single-slot cooler and lack of power connectors simplify installation in pre-built machines. The PCIe 3.0 x8 interface is compatible with virtually all motherboards from its release era. Users who need multiple display outputs will find one DisplayPort 1.4a and two mini-DisplayPort 1.4a ports adequate for multi-monitor productivity setups. The 1:1 FP16 ratio means no compute advantage for machine learning or scientific workloads that rely on half precision. For 1440p or high-refresh gaming, the 1.658 TFLOPS and 96.00 GB/s bandwidth are insufficient; this card is not aimed at those use cases. The end-of-life status suggests it is a replacement part for existing systems rather than a new-build recommendation.

Benchmark Performance

The database records an average benchmark score of 0 for the RX 640 OEM, and the benchmark array is empty, so no measured scores exist to analyze. The only quantitative performance indicator is the percentileVsAllGpus value of 50, which places the card at the median of all GPUs in the database. Because nearestRivals is empty, no deltaPct values can be computed, and no rival names or scores can be cited. The performance envelope must therefore be inferred from the specification data.

FP32 compute is 1.658 TFLOPS, texture rate is 51.80 GTexel/s, and pixel rate is 20.72 GPixel/s. These figures are internally consistent: with 640 shading units at the listed clocks, the compute and fill rates align with a mid-pack part. The memory subsystem — 4 GB GDDR5 at 96.00 GB/s over a 128-bit bus — is the likely bottleneck in memory-intensive workloads, as the bandwidth is low relative to the compute throughput. The 1:1 FP16 ratio (1.658 TFLOPS) means mixed-precision tasks gain nothing over FP32.

The 50th percentile ranking suggests that in the aggregate database, this card outperforms half of all tracked GPUs and underperforms the other half. Without benchmark scores, the percentile is the sole comparative metric, and it indicates a balanced, middle-of-the-road part. The 2,200 million transistors on a 103 mm² die at 14 nm and the 50 W TDP confirm that this is a low-power design, and the 250 W PSU recommendation reflects the modest system-level requirements. In the absence of measured data, the spec-derived rates — 1.658 TFLOPS, 51.80 GTexel/s, 20.72 GPixel/s, 96.00 GB/s — are the best available proxies for real-world performance.