Adapteva

Adapteva is a fabless semiconductor company focusing on low power many core microprocessor design. The company was the second company to announce a design with 1,000 specialized processing cores on a single integrated circuit.^[1][2]

Adapteva was founded in 2008 with the goal of bringing a ten times advancement in floating point performance per watt for the mobile device market. Products are based on its Epiphany multi-core MIMD architecture and its Parallella Kickstarter project promoting "a supercomputer for everyone" in September 2012. The company name is a combination of "adapt" and the Hebrew word "Teva" meaning nature.

History

Adapteva was founded in March 2008, by Andreas Olofsson. The company was founded with the goal of bringing a 10× advancement in floating point processing energy efficiency for the mobile device market. In May 2009, Olofsson had a prototype of a new type of massively parallel multi-core computer architecture. The initial prototype was implemented in 65 nm and had 16 independent microprocessor cores. The initial prototypes enabled Adapteva to secure US$1.5 million in series-A funding from BittWare, a company from Concord, New Hampshire, in October 2009.^[3]

Adapteva's first commercial chip product started sampling to customers in early May 2011 and they soon thereafter announced the capability to put up to 4,096 cores on a single chip.

The Epiphany III, was announced in October 2011 using 28 nm and 65 nm manufacturing processes.

Products

Adapteva's main product family is the Epiphany scalable multi-core MIMD architecture. The Epiphany architecture could accommodate chips with up to 4,096 RISC out-of-order microprocessors, all sharing a single 32-bit flat memory space. Each RISC processor in the Epiphany architecture is superscalar with 64× 32-bit unified register file (integer or single precision) microprocessor operating up to 1 GHz and capable of 2 GFLOPS (single precision). Epiphany's RISC processors use a custom instruction set architecture (ISA) optimised for single precision floating point,^[4] but are programmable in high level ANSI C using a standard GNU-GCC tool chain. Each RISC processor (in current implementations; not fixed in the architecture) has 32 KB of local memory. Code (possibly duplicated in each core) and stack space should be in that local memory; in addition (most) temporary data should fit there for full speed. Data can also be used from other processor cores local memory at a speed penalty, or off-chip RAM with much larger speed penalty.

The memory architecture is does not employ explicit hierarchy of hardware caches, similar to the Sony/Toshiba/IBM cell processor, but with the additional benefit of off-chip and inter-core loads and stores being supported ( which simplifies porting software to the architecture). It is a hardware implementation of partitioned global address space.

This eliminated the need for complex cache coherency hardware, which places a practical limit on the number of cores in a traditional multicore system. The design allows the programmer to leverage greater foreknowledge of independent data access patterns to avoid the runtime cost of figuring this out. All processor nodes are connected through a Network On Chip, allowing efficient message passing.^[5]

Scalability

The architecture is designed to scale almost indefinitely, with 4 e-links allowing multiple chips to be combined in a grid topology, allowing for systems with thousands of cores.

Multi-core coprocessors

16-core Adapteva Epiphany chip, E16G301, from Parallella single-board computer

On August 19, 2012, Adapteva posted some specifications and information about Epiphany multi-core coprocessors.^[6]

In September 2012, a 16-core version, the Epiphany-III (E16G301), was produced using 65 nm^[9] (11.5 mm², 500 MHz chip^[10]) and engineering samples of 64-core Epiphany-IV (E64G401) were produced using 28 nm GlobalFoundries process (800 MHz).^[11]

The primary markets for the Epiphany multi-core architecture include:

• Smartphone applications such as real-time facial recognition, speech recognition, translation, and augmented reality.
• Next generation supercomputers requiring drastically better energy efficiency to allow systems to scale to exaflop computing levels.
• Floating point acceleration in embedded systems based on field-programmable gate array architectures.

Parallella project

Parallella single-board computer with 16-core Epiphany chip and Zynq-7010 FPGA

In September 2012, Adapteva started project Parallella on Kickstarter, which was marketed as "A Supercomputer for everyone." Architecture reference manuals for the platform were published as part of the campaign to attract attention to the project.^[12] The US$750,000 funding goal was reached in a month, with a minimum contribution of US$99 entitling backers to obtain one device; although the initial deadline was set for May 2013, the first single-board computers with 16-core Epiphany chip were finally shipped in December 2013.^[13]

Size of board is planned to be 86 mm × 53 mm (3.4 in × 2.1 in).^[14][15][16]

The Kickstarter campaign raised US$898,921.^[17][18] Raising US$3 million goal was unsuccessful, so no 64-core version of Parallella will be mass-produced.^[19] Kickstarter users having donated more than US$750 will get "parallella-64" variant with 64-core coprocessor (made from initial prototype manufacturing with 50 chips yield per wafer).^[20]

Epiphany V

As of 2016, the firm has taped out a 1024 core 64bit variant of their Epiphany architecture.^[21] This features: larger local stores (64kb), 64bit addressing, double precision floating point arithmetic or SIMD single precision, and 64bit integer instructions, implemented in the 16nm process node. This design is backwards compatible with earlier products. There are instruction set enhancements aimed at deep-learning and cryptography applications. It is a DARPA funded project.

Performance

Joel Hruska from Extremetech has the following opinion about the Parallella project: "Adapteva is drastically overselling what the Epiphany IV can actually deliver. 16–64 tiny cores with small amounts of memory, no local caches, and a relatively low clock speed can still be useful in certain workloads, but contributors aren't buying a supercomputer—they're buying the real-world equivalent of a self-sealing stem bolt."^[22]

The criticism that the Epiphany chips cannot provide anywhere near the performance of modern supercomputers is nevertheless correct: actually, Epiphany chips with 16-cores or 64-cores and c. 25 or 100 GFLOPs in single precision, respectively, do not even match the floating-point performance of modern desktop PC processors (Core i7-4770K (Haswell), 4× cores @ 3.5 GHz AVX2: 177 GFLOPS,^[23] double precision) – a fact that is acknowledged by Adapteva.

However, the latest Parallella boards with E16 Epiphany chips^[24] can be compared to many ancient supercomputers in terms of raw performance (just as an example, the Cray 1 – the first supercomputer per se – had a peak performance of 80 MFLOPS at 1976, and its successor the Cray 2 had a peak performance of 1.9 GFLOPS at 1985), and can certainly be used for parallel code development; The architectural similarities to supercomputers (message passing and NUMA) make it a potentially useful development system, compared to traditional SMP machines.

The point being that for a power envelope of 5 W and in terms of GFLOPS/mm² of chip die space, the current E16 Epiphany chips provide vastly more performance than anything else available to date, with an architecture designed to scale, and applicable to more than just embarrassingly parallel GPU tasks. (e.g. it would be capable of running the actor model with many concurrent, fully independent states). It is also suitable for DSP-like tasks where data could be fed directly on chip (from an FPGA or other ASIC) without having to create buffers in temporary memory as for a GPU), making it ideal for robotics & other intelligent sensor applications.

The architecture also allows parallella boards to be combined into a cluster with a fast inter-chip 'eMesh' interconnect, extending the logical grid of cores (creating almost unlimited scaling potential).

The 16-core Parallella has roughly 5.0 GFLOPs/W, and the 64-core Epiphany-IV made with 28 nm estimated as 50 GFLOPs/W (single precision),^[25] and 32-board system based on them has 15 GFLOPS/W.^[26] For comparison, top GPUs from AMD and Nvidia reached 10 GFLOPs/W for single precision in 2009–2011 timeframe.^[27]

See also

• SW26010, a Chinese design featuring a similar architecture, used in their Sunway TaihuLight supercomputer
• Partitioned global address space, a similar memory model found in some HPC applications
• CELL, a multicore processor with scratchpad memory and inter-core DMA
• IBM TrueNorth, a design by IBM with a radically different neuromorphic processor, while having a similar network on chip, extendable with a grid of chips
• Kalray, a company producing a similar network-on-chip RISC processor, aimed at workloads such as convolutional neural networks for vision, again able to extend its grid between multiple chips
• Vision Processing Unit, a class of processor with significant overlapping features
• GRVI Phalanx, an FPGA RISC-V softcore with some similar features
• Asynchronous array of simple processors
• NUMA

References

Further reading

• Linley Gwennap, Adapteva: More Flops, Less Watts. Epiphany Offers Floating-Point Accelerator for Mobile Processors. // Microprocessor Report, June 2011

External links

• Official website
• Parallella Specifications

Single-board computer & Single-board microcontroller
Devices	Arduino Arndale Board Banana Pi BeagleBoard Cotton Candy CHIP Cubieboard Dragonboard Edison Galileo Gumstix Hawkboard IGEPv2 Jetson TK1 Jetson TX1 ODROID OLinuXino PandaBoard Parallella Rascal Raspberry Pi Snowball Tiva-C LaunchPad UDOO
SoCs	ARM Allwinner Ax Exynos i.MX K3V3 MT Nomadik NovaThor OMAP Rockchip Qualcomm Snapdragon Tegra WonderMedia MIPS Jz x86/x86-64 Atom Jaguar-based Puma-based Quark
Software	Apache Hadoop Linaro
Comparison of single-board computers