NVIDIA Tesla K20s

The NVIDIA Tesla K20s is a Kepler-generation compute accelerator built around the GK110 chip, fabricated by TSMC on a 28 nm process. The die contains 7,080 million transistors across 561 mm², for a transistor density of 12.6M per mm². The compute configuration is 2,496 shading units, 208 texture mapping units, and 40 ROPs. Memory is 5 GB of GDDR5 on a 320-bit bus, providing 208.0 GB/s of bandwidth. The FP32 compute rating is 2.870 TFLOPS, and the card carries a 225 W TDP with a dual-slot cooler and a suggested 550 W power supply. It was released on 2013-02-17 with a launch MSRP of 3,199 USD, and is now end-of-life, sitting between Tesla Fermi (predecessor) and Tesla Maxwell (successor). There are no display outputs, marking this as a dedicated compute device rather than a graphics card.

Benchmark Performance

The database records no benchmark scores for the Tesla K20s; the benchmarks array is empty and the average benchmark score is zero. That absence is a meaningful signal: this card was never part of the gaming or workstation benchmark loops that populate most GPU entries. The only positioning metric available is the 50th percentile against all GPUs, which places it at the exact median of the tracked field, neither a performance outlier nor a laggard.

Without recorded scores, the compute and memory specifications become the primary performance indicators. The FP32 throughput of 2.870 TFLOPS is the headline number; it is the raw single-precision compute the chip can sustain and the figure most relevant to scientific and simulation workloads that rely on dense floating-point math. The texture rate of 119.6 GTexel/s and pixel rate of 29.90 GPixel/s describe the rasterization throughput of the 208 TMUs and 40 ROPs. These are secondary for a compute card, but they characterize the silicon's capability beyond the FP32 path.

Memory bandwidth of 208.0 GB/s is the other key performance limiter. In compute tasks, data movement often dominates execution time, and this figure determines how quickly the 5 GB frame buffer can be fed. The 320-bit bus width is wide for the era, and the 1300 MHz memory clock (5.2 Gbps effective) yields that 208.0 GB/s figure. The 50th percentile ranking suggests the K20s lands in the middle of the GPU population in aggregate, but the empty benchmark array means the database cannot confirm this with direct score deltas against any named competitor.

Who Should Consider It

The Tesla K20s has no display outputs, so it is not a graphics card in the conventional sense. It cannot drive a monitor, and it lacks the modern feature set for gaming. This is a compute accelerator for servers and workstations. The 5 GB of GDDR5 memory and 208.0 GB/s bandwidth define the envelope for workloads: data sets that fit within 5 GB can be processed locally, and the bandwidth determines how fast they can be streamed through the 2,496 shading units.

For resolution and settings guidance, the usual question for graphics cards, the K20s is not applicable, because it renders nothing to a screen. Instead, the relevant considerations are workload size and precision. The 2.870 TFLOPS of FP32 compute suits single-precision scientific codes, molecular dynamics, and similar HPC tasks. The absence of FP16 support (the field is null) means half-precision deep learning workloads are not a target. The 225 W TDP and 550 W suggested PSU indicate a system builder should plan for a power supply with reasonable headroom; the dual-slot cooler and 267 mm (10.5 inches) length dictate chassis clearance.

The PCIe 2.0 x16 interface is an older bus standard; in a modern system, this will be the transfer bottleneck for host-device communication, though compute kernels themselves run on the card. Buyers looking at this card today are likely dealing with legacy HPC clusters or specific Kepler-era software stacks, where the card's compute profile and 5 GB capacity are still serviceable.

Ray Tracing and Feature Set

There are no ray tracing cores and no tensor cores on the Tesla K20s. The chip is pure Kepler compute: 2,496 shading units, 208 TMUs, and 40 ROPs arranged around the GK110 die. Ray tracing acceleration, as it exists in modern GPUs, is entirely absent. Tensor cores for AI and deep learning matrix math are also missing; the card's FP32 path is the only compute path, and FP16 is not supported (null in the specification).

The API support reflects the Kepler generation. DirectX 12 is listed at the 11_0 feature level, meaning the card can run DirectX 12 applications that target the 11_0 feature set but not the full DirectX 12 feature levels. OpenGL 4.6 and Vulkan 1.2.175 are supported, which keeps the card usable for compute and rendering contexts that rely on those APIs. For compute workloads, the relevant interfaces are the general compute paths exposed through the graphics APIs; the pack does not list dedicated compute frameworks.

The feature set is minimal by modern standards. No display outputs, no RT cores, no tensor cores. The card is a single-purpose compute device, and its feature list is entirely about feeding the FP32 pipeline.

FAQ

Q: What architecture is the Tesla K20s based on?

A: It uses the Kepler architecture, specifically the GK110 chip, manufactured on TSMC's 28 nm process.

Q: Does the Tesla K20s support ray tracing or tensor cores?

A: No. The card has no RT cores and no tensor cores; it is a pure FP32 compute device with 2.870 TFLOPS of single-precision throughput.

Q: How much memory does the K20s have and what is its bandwidth?

A: It has 5 GB of GDDR5 memory on a 320-bit bus, with 208.0 GB/s of bandwidth, running at 1300 MHz (5.2 Gbps effective).

Q: Can the Tesla K20s output video to a display?

A: No. It has no display outputs, it is a compute accelerator, not a graphics card.

Q: What is the launch MSRP of the Tesla K20s?

A: The launch MSRP was 3,199 USD.

Q: Is the Tesla K20s still in production?

A: No. It is end-of-life, released on 2013-02-17, with Tesla Fermi as its predecessor and Tesla Maxwell as its successor.

How It Compares

The database lists no nearest rivals for the Tesla K20s, so direct score deltas against competing cards cannot be drawn from the record. The only comparative anchor is the 50th percentile against all GPUs, which places it at the median of the tracked population. The empty benchmark array means there is no average score to contrast with other entries.

Against its own product lineage, the K20s sits between Tesla Fermi (predecessor) and Tesla Maxwell (successor). The Fermi generation was the prior compute architecture; Maxwell followed as the successor. Without benchmark data, the comparison is qualitative: the K20s represents the Kepler step in that progression, and its 28 nm process and 7,080 million transistor count define its place in that history.

The card's position in the broader GPU market is that of a specialized compute part. Its 50th percentile ranking reflects that it is neither a top-tier performer nor a weak one, it is exactly mid-pack in the database's tracking. But the lack of display outputs and the absence of RT and tensor cores mean it competes in a different arena than consumer graphics cards, and the database's benchmark suite does not capture its workload.

Memory Subsystem

The memory subsystem is built around 5 GB of GDDR5 on a 320-bit bus. The memory clock is 1300 MHz, which translates to 5.2 Gbps effective and yields a bandwidth of 208.0 GB/s. This is the data pipe for the 2,496 shading units and the 2.870 TFLOPS FP32 pipeline.

For compute workloads, the 5 GB capacity is the working-set limit. Data sets larger than 5 GB must be tiled or streamed through the PCIe 2.0 x16 interface, which is a slower path than the on-card memory. The 208.0 GB/s bandwidth is the sustained rate at which data can be read from or written to the frame buffer; compute kernels that are memory-bound will see performance scale with this number, while compute-bound kernels will scale with the FP32 throughput.

The 320-bit bus width is the structural enabler of that bandwidth, a narrower bus would need a much higher clock to achieve the same figure. The pixel rate of 29.90 GPixel/s and texture rate of 119.6 GTexel/s are downstream of the memory subsystem and the ROP and TMU counts; for a compute card, these are less critical than the raw bandwidth and capacity, but they round out the memory interface's capabilities. At high resolutions or with large data sets, the 5 GB capacity and 208.0 GB/s bandwidth define the practical ceiling; exceeding either forces data movement over the PCIe bus, which is the slowest part of the system.