AMD has held an event called Next Horizon. It was devoted to future advanced products of the company, in particular, the world's first graphics processor, made by 7-nm process technology.
It is not hard to guess that we are talking about a new version of the Vega graphics processor, which figured in the leaks under the code name Vega 20. The graphics processor "Vega 7nm", as AMD itself calls it, has an area of only 331 mm2, but at the same time it placed 13.23 billion transistors due to the use of TSMC 7-nm FinFET process technology. For comparison, the 16-nm graphics processor NVIDIA Pascal GP104 with an area of 314 mm2 It has only 7.2 billion transistors.
Compared with the original 14-nm version of the Vega, the new 7-nm version doubled the density of transistors. With the same power consumption, the graphics processor is capable of providing 25% better performance. And at the same frequency, the power consumption of the new product is 50% lower.
AMD introduced the Radeon Instinct MI50 and MI60 computing accelerators on a new graphics processor. Between themselves, these accelerators differ only in the configuration of the GPU, as well as the amount of video memory. The junior Radeon Instinct MI50 is built on a stripped-down version of the “Vega 7nm” with 3840 stream processors and a clock frequency of up to 1746 MHz. In turn, the Radeon Instinct MI60 offers a full-fledged Vega 7nm with 4096 stream processors and frequencies up to 1800 MHz. The power consumption of both accelerators is 300 watts.
Both updates use HBM2 memory with ECC error correction support. The junior model received 16 GB, while the older AMD allocated 32 GB of memory at once. In both cases, memory with a frequency of 1000 MHz was used, and a 4096-bit wide bus was used to connect it. As a result, the memory bandwidth reaches 1 TB / s.
As for performance, AMD calls the Radeon Instinct MI60 the fastest accelerator for double precision calculations (FP64). Productivity in such operations reaches 7.4 Tflops. In single (FP32) and half (FP16) accuracy operations, the performance indicators are 14.7 and 29.5 teraflops, respectively. Finally, in operations like INT8, performance reaches 59 TOPS. The Radeon Instinct MI50 has slightly lower performance because of the lower number of stream processors and frequency.
The performance of the new Radeon Instinct MI60 accelerator is comparable to the performance of the NVIDIA Tesla V100 accelerator based on the Volta GPU. And in terms of theoretical performance, expressed in teraflops, AMD's new product outperforms its competitor a bit, and in training the neural network in ResNet-50, there is an advantage in the direction of the NVIDIA accelerator.
Another important feature of the new Radeon Instinct accelerators is that they received support for the PCI Express 4.0 x16 interface. It is able to provide bandwidth between a CPU and a GPU up to 64 GB / s in both directions at the same time. Of course, there is also backward compatibility with the PCI Express 3.0 x16 interface.
Also note that for communication between GPUs, when they work in bundles, the Infinity Fabric bus is used with a bandwidth of 100 GB / s per channel. The new video cards have two such channels, that is, the total bandwidth is 200 GB / s. For comparison, the NVLink interface offers a bandwidth of 50 GB / s per port and up to 200 GB / s per GPU.
The new Radeon Instinct accelerators can be used not only for calculations and tasks related to artificial intelligence, but also they support third-generation hardware virtualization, which does not require a software component, and therefore avoids the cost of software for virtualization. One accelerator supports up to 16 virtual machines, or one virtual machine can include up to 8 Radeon Instinct accelerators.
AMD has announced a new and improved open source software platform for developing ROCm 2.0. It was also noted that the ROCm platform runs on NVIDIA GPUs and Intel CPUs, not just AMD products.
AMD Radeon Instinct MI60 computing accelerator will go on sale in the current fourth quarter of 2018. In turn, the younger accelerator Radeon Instinct MI50 will have to wait until the first quarter of 2019. Unfortunately, the cost of new products, AMD does not specify.
If you notice an error, select it with the mouse and press CTRL + ENTER.