Intel Stratix chip performs over 10 trillion calculations per second

Greg S

Posts: 1,607   +442

Compute solutions are becoming more situational for high performance applications. Quantum, neuromorphic, and parallel processing architectures are pushing boundaries of what is possible. Intel has shared details on its latest field programmable gate array capable of performing more than 10 trillion floating point operations per second.

Intel's new Stratix 10 FPGA is currently the fastest available chip of its type. Targeting applications such as speech-reconition, 5G networking, artificial intelligence, search engines, and high-performance computing, Stratix 10 offers extreme levels of performance in a small package.

In order to achieve such performance, over 30 billion 14nm transistors are packed into each chip. For comparison, 8 core Ryzen CPUs have around 4.8 billion transistors produced at 14nm. Intel has not disclosed transistor counts for their latest CPUs, but states that this is more than three times the amount found "in the chips that run today's fastest laptops and desktops."

The Stratix 10 family of chips comes varying package sizes depending on how many logic elements and other features are available. The smallest chip measures 35mm x 35mm while the largest scales up to 55mm x 55mm. Even though the largest chip is slightly bigger than most general purpose CPUs, it is still a difficult task to fit all of those transistors on the die.

To be able to handle over 10 trillion operations per second takes a lot of bandwidth. Each Stratix 10 has up to 144 full duplex transceiver lanes with serial rates up to 58 Gbps. Intel states that the equivalent of 420 Blu-ray Discs can be pushed through the chip every second.

General purpose CPUs are unable to be significantly changed after they are manufactured. FPGAs are able to be completely reprogrammed at any point in time to perform a new specific task. Should network operators or AI developers need to run new software, its a simple matter of uploading a new program to the chip.

Permalink to story.



Posts: 5,416   +3,487
This all sounds great from a marketing standpoint, but having programmed relatively simple FPGAs in the past, how well this will pan out in practice is going to greatly depend on the CAD software that programs this chip, and how well that software is able assemble those 30-billion transistors into logical units that can do real work.

As I see it, programming the structure of the logical units on this chip might be considered akin to laying out a CPU, only more complex due to the fact that there are significantly more transistors in this chip than in modern CPUs. It might require a team of experts many months to come up with the right design even with the proper software. That you can download an new program to the chip and instantly change its function from one task to another says nothing about how much time it will take to come up with that new design to download to the chip.

In other words, even if it might be tasked to mine cryptocurrency, it might take months to come up with a suitable software design.

So, all I have to say is buyer beware of the marketing spiel - it is likely significantly more complex than the marketing spiel is attempting to lead potential buyers to believe.
Although the programming will require a serious commitment by companies wishing to utilize this tech for their programs, which will require various incentives (number of computers with this chip, most common computer activities for those running systems with this ship type, overall demand, etc.), eventually the tech will likely become prominent as the prices go down. FPGAs have the advantage of flexibility, which will likely increase the appeal of this CPU architecture over time. Several of the products I use have FPGA type processors in them and the advantages have been wonderful. E.G. I am a professional mixing / mastering engineer, and the PCI cards that connect my primary AD/DA converters are controlled by FPGAs, which allow the manufacturer to fix bugs and add functionality to the system at the hardware level; this in turn has extended the service life of the converters beyond most other contemporary products in the same class.
Wait so what they are saying is that the newest intel chips have around 10 million transistors but is only marginally better than ryzen that is by the article only 4.8 million.