Nvidia: its Grace processor would be 2x faster than Intel Ice Lake chips


At this year’s GTC, NVIDIA showcased a host of impressive new computing technologies, including its new Grace “Superchip” CPU, which promises to be twice as fast as Intel’s best chips.

Credit: Nvidia

At GTC 2022, Nvidia presented its new Grace Superchip CPU, aimed at AI infrastructures and high-performance computing. The new data center CPU 144-core Arm Neoverse consists of two 72-core CPU chips coupled by NVLink, the company’s high-speed, low-latency chip-to-chip (C2C) connection. This is similar technology to what Apple uses for its recent M1 Ultra chip found on board the Mac Studio.

According to data from Nvidia, its chip would be really overpowered. NVIDIA mentions that the system offers 50% better performance than dual x86 CPUs found in current DGX A100 systems. NVIDIA here refers to AMD’s ‘Rome’ EPYC 7742 CPUs which have 128 cores and 256 threads in total. However, it must be remembered that these are based on the Zen 2 architecture, not AMD’s latest Zen 3 architecture.

Read also : Intel surpasses Nvidia and AMD, its ARC GPUs are the 1st to support the AV1 format 100%

Nvidia’s chip outperforms Intel’s best chips

In addition to beating AMD, Tom’s Hardware found that Nvidia also benchmarked its new processor against Intel’s best chips from the Ice Lake generation. In a benchmark, Nvidia claims that Grace is 2x faster and 2.3x more power efficient than Intel’s current generation Ice Lake in a commonly used weather research and forecasting (WRF) model.

To achieve such a level of performance, Grace would have a unique and efficient memory subsystem composed of LPDDR5x memory with Error Correction Code (ECC). According to NVIDIA, its memory offers 1 TB/s bandwidthalmost double the bandwidth of traditional DDR5 designs, and would consume a 500W maximum. All this comes in addition to an impressive cache of 396 MB. Finally, Nvidia announced that its superchip can be configured in servers as standalone systems with a single processor or as GPU-accelerated servers with one, two, four or eight Grace Hopper GPUs.

Source: Tom’s Hardware



Source link -101