The chip will combine many CPUs and graphics processors in a single package. It will deliver multiple teraflops of performance, and will be targeted at high-end graphics, supercomputing and data modeling.
The chip doesn’t have a name yet, and could be released in the next two years, said Forrest Norrod, Senior Vice President at AMD, during an investor meeting on May 6th.
There is no chip that mixes CPUs and GPUs, and AMD wants to introduce the concept to high-performance computing, Norrod said.
For decades, AMD and Intel have been locked in a leapfrogging race, one-upping each other by increasing their chips’ clock speeds. At the height of the chip wars in 2011, AMD claimed a Guinness World Record for achieving the “highest frequency of a computer processor” with its eight-core FX chip, which ran at a clock speed of 8.4GHz. The silliness of the chip battles dissipated soon after, though chip makers still love to parade high-performance chips as a sign of technological progress.
Intel’s Core i7-5960X Extreme Edition is considered the world’s fastest CPU. Intel this week introduced an 18-core Xeon chip that could outperform the Core i7 chip, but benchmarks are not yet available. GPUs typically add more teraflops of performance.
In recent years, AMD has been silent about high-performance chips as it has focused on power efficiency. But the new mega-chip will be designed for speed and for delivering performance that could make it possible to do high-end graphics, calculations and simulations on servers and supercomputers.
The chip is targeted at servers and workstations. It may be possible to adapt a workstation version of the chip for a gaming rig, but it could be more suited for processing large data sets associated with servers or supercomputers. The chip will support a programming model so high-performance tasks can be effectively executed over multiple processing resources on the chip.
GPUs are being used in some of the world’s fastest computers for complex math and scientific calculations, and AMD is trying to get its FirePro GPUs into more supercomputers.
AMD could also tweak the chip for other machines that require lots of computing performance. AMD has already done something similar for the Sony and Microsoft gaming consoles, where it put together a CPU and graphics cores on a big custom chip.
The Zen CPU core in the new chip will be at least 40 percent faster than the current CPU core code-named Excavator. Based on information shared by AMD at the meeting, the chips could also support DDR4 memory, and the GPUs a new form of high-bandwidth memory.
The chip seems to have the big memory throughput needed to feed data to a large numbers of CPU and GPU cores, said Jim McGregor, principal analyst at Tirias Research.
But when doing a massive amount of computing, it’s challenging to put parts like a CPU and GPU in one package, McGregor said.
There could be problems related to balancing power consumption and performance as GPUs are power hogs that generate a lot of heat. That could severely limit the server design, which would hinge on components to help cool down the high-performance chip.
But AMD may already have orders to make such chips, and perhaps sees a long-term opportunity, McGregor said.
AMD’s rivals are taking a different approach. Instead of integrating a GPU, Nvidia’s high-performance Tesla GPU is offered as a co-processor alongside CPUs. Nvidia toyed with the idea of integrating a Tesla-like GPU in its Tegra chips, but dropped the idea. Intel is pairing a bunch of vector processing cores alongside generic low-power x86 chips in its Knights Landing chip, which can deliver over 3 teraflops of performance.