Simply over a month in the past, Tachyum, a small tech startup, introduced a brand new processor household, Prodigy, which it calls the world’s first common processor platform. Past that new label although are quite a lot of hard-to-believe claims that the corporate has put ahead together with the truth that it delivers a 10x efficiency enchancment on typical processors.
We caught up with Tachyum CEO Dr. Radoslav ‘Rado’ Danilak by way of e-mail, to seek out out extra in regards to the enterprise and the place precisely does these seemingly outlandish claims come from.
1. How is it totally different from the Heterogeneous System Structure (HSA) Basis?
Tachyum’s Prodigy processor is a brand new and progressive processor structure, developed utilizing a and software program codesign from day-1. It has a single programming mannequin, a single instruction stream, absolutely coherent reminiscence, and absolutely coherent inter-core communication. We have now additionally added knowledge parallelism to our processor programming mannequin in an effort to higher deal with sure AI purposes.
2. Your press launch mentions an AI chip, GPU and CPU as being a part of that household. Are you able to inform us extra?
Tachyum’s Prodigy processor is a single unified structure, which displays out of order efficiency, with processor (learn: transistor counts / core dimension) just like easy in-order execution machines. We have now achieved this by offloading to our compiler, duties historically carried out in CPU . The ensuing IPC, clock pace and energy discount enhancements not solely provide a compelling worth proposition in our core market, Hyperscale Information Facilities, however in addition they allow Tachyum’s Prodigy processor to exceed NVIDIA Volta efficiency on Neural Nets.
Tachyum has created a brand new processor structure which not solely outperforms the competitors in Information Middle workloads, but additionally outperforms the competitors in all AI disciplines. Prodigy demonstrates an order of magnitude higher efficiency on Symbolic AI, Bio AI, and Common AI (attributable to their management intensive nature) than present AI-centric chips. Tachyum has NOT mixed a CPU, GPU, and AI chip.
Tachyum has developed an progressive processor structure that gives a disruptive worth proposition throughout a number of software domains. We have now additionally included within the Prodigy structure sure architectural enhancements to reinforce its efficiency on AI workloads, akin to compressed Eight-bit floating-point coefficients and matrix multiply-add operations.
three. Your organization is promising to durably disrupt the compute market with some, frankly extraordinary claims. How did you obtain a lot by spending (in relative phrases) so little in R&D in comparison with the likes of Samsung, Nvidia or Intel.
The Prodigy structure is the results of many years of expertise that I developed designing processors (e.g. Ps 2, Tesla), flash reminiscence controllers (Sandforce), and flash based mostly methods (Skyera). A number of years of self-funded R&D preceded Tachyum’s emergence from stealth mode. I’ve all the time been desirous about fixing “system physics” challenges, akin to reliability points in twin degree cell flash reminiscence, as I did at Sandforce. Prodigy is one other instance of that. With the last decade lengthy stagnation of processor clock pace, due largely to sluggish wires relative to transistor switching pace, and matched with CPU architectures which had been designed when wires had been infinitely quick in comparison with transistors, a recent take a look at an optimum 21st century processor structure was warranted. We began from a clear sheet of paper with a design philosophy of lowering the variety of sluggish wires on a chip, and lowering the common size of present wires. The result’s breakthrough efficiency and low energy consumption.
four. Why can Tachyum succeed the place bigger organizations have failed?
I must say it is because of my technical instincts born from laborious gained expertise, together with studying from others’ errors, and disciplining myself to work solely on vital challenges. Even in giant firms akin to Intel and NVIDIA, the true innovation normally springs from a small group of innovators. At SandForce which I based, my rivals had been Intel, Samsung, Toshiba, Sandisk, Micron, Western Digital, LSI, Seagate, and plenty of others with 1000s [of] engineers, and with lower than 100 staff we gained.
Having the intuition to go in proper route is vital, hiring greatest crew, studying from errors of rivals, and dealing solely on vital stuff. An instance is the Intel Itanium that failed on compilers, so we developed compilers first and construct structure round compilers. Even at nVidia, key innovation was completed with a pair [of] teams with ten engineers, not 1000’s. When you’ve got crew of “gods”, [head] depend just isn’t that crucial and may be crammed with contractors.
5. You say that constructing a candidate for the human mind mission will take lower than three years with about 250,000 of your chips, is it protected to imagine that this processor has a peak efficiency of four teraflops?
Tachyum’s 64 core Prodigy processor, will generate ~128TFLOPS. We declare that in 2020, with quantity manufacturing of Prodigy underway, system integrators will be capable to assemble ~250,000 Prodigy processors right into a community able to operating human mind sized neural nets. The processing density of Prodigy, mixed with its disruptive low energy consumption, allow these methods to be constructed starting in 2020.
6. The administration of Tachyum was concerned within the sale of Skyera to WDC. Does storage and compute due to this fact share the identical basic scale points?
Tachyum and Skyera are two utterly totally different firms, besides that some founders of Tachyum had been at Skyera. We make no claims about comparable scaling points between the reminiscence and processor domains.
7. We spoke so much about however what the position of software program in your plans?
Our sensible compiler is important to the Prodigy answer, dealing with many processor duties historically dealt with in . We have now GCC at the moment, and can present LLVM subsequent 12 months in addition to Java JIT. Linux and FreeBSD shall be natively supported. We’ll work carefully with software builders to insure they will absolutely exploit the efficiency traits of Prodigy for each knowledge heart apps, in addition to AI purposes in all domains.
Eight. What’s your small business mannequin? Do you propose to license the IP to 3rd events (just like Rambus and Arm) or do you propose to tackle the entire market?
Tachyum is a semiconductor firm. We promote chips to finish prospects, ODM’s and OEM’s. Tachyum just isn’t an IP supplier.