Gigabyte Leaks AMD Zen 4 Details: 5nm, AVX-512, 96 Cores, 12-Channel DDR5

Gigabyte isn’t having a great month. The company was hit by an attack that stole 112GB of data off its servers last week, including (apparently) data on AMD’s upcoming Genoa servers, based on its Zen 4 architecture.

Zen 4 isn’t expected to launch in the near future. The next server CPU expected from AMD is Milan-X, which will combine 64MB of vertically mounted L3 cache per chiplet with an existing Zen 3 core, or up to 512MB of additional L3 in total. Zen 4, when it arrives next year, will bring a series of fundamental changes (and, presumably, will keep its larger L3 cache).

The leaked documents, if accurate, describe a Zen 4 Genoa server CPU with up to 96 cores and 12 channels of DDR5 memory rated at DDR5-5200. This works out to half a terabyte of memory bandwidth in total. While the number of memory channels per chiplet remains the same (one), the massive L3 cache increase expected from Milan-X, combined with Genoa’s 1.62x increase in memory bandwidth, will keep the core fed like never before. If AMD builds a top-end workstation chip out of this core, we’d expect it to be less memory-bandwidth-bound than the Threadripper 3990X for these two reasons alone.

The leak also shows that TDPs on Genoa servers will be headed upwards, from the current 280W to 320-400W for the top products. The 64-core 400W TDP variant shown above will presumably hit higher frequencies than the 96-core variant. Note that there’s an even higher 700W target that the CPU is allowed to hit for 1ms or less.

As for the AVX-512 claim, there’s a specific document mentioning AVX-512 support in upcoming AMD processors.

Despite the shift to 5nm, adding features like AVX-512 means the actual chiplet sizes haven’t changed much. The increase in power consumption from Milan to Genoa is probably related to a number of factors, including increased L3 cache (we’re assuming this feature carries over from Milan-X), increased interconnect power to keep all 96 cores fed, and the power requirements of AVX-512. Intel embraced AVX-512 years ago, despite the fact that it carries a heavy power penalty and requires a CPU to run at lower frequencies to stay within thermal limits. AMD may be picking up the feature now that the impact is smaller.

These leaks are broadly in line with what we expected, but they confirm that AMD is going to start nudging core counts upwards again. After introducing 16 cores on the desktop and 64 cores in workstations and servers, AMD has given the market a few years to digest the improvements. We’ve seen a sharp increase in the number of gamers with six-core and eight-core chips over the past 12 months, suggesting that gamers are indeed taking advantage of the increased core counts now on offer from AMD and Intel. It is not a given that AMD intends to increase consumer desktop core counts, but bumping up to 96 cores at the top of the stack gives the company some breathing room.

As always, rumors and leaks should be taken with a grain of salt. Hat-tip to ComputerBase.de for the information.

Continue reading

Report: Intel Will Outsource Core i3 Production to TSMC’s 5nm Node

This would be the first time the chip giant has built one of its Core CPUs on a different company's process node. It's not an exaggeration to call this a pivotal moment in Intel's history.

Report: Google Tensor Chip Will Use 5nm Process, Samsung to Manufacture

This chip will come in the new Pixel 6 and 6 Pro later this year, but Google didn't offer any real hardware details on Tensor, except to say it's based on technology from its Tensor AI server chips. A new report includes a few tantalizing details, though.

EUV Integration at 5nm Still Risky, With Major Problems to Solve

TSMC, GlobalFoundries, Samsung, and Intel are all integrating EUV at future process nodes, even though major research firms are warning about high defect densities and other problems. Why the rush?

Samsung, ARM Expand Collaboration in HPC at 7nm, 5nm

Samsung is expanding its partnership with ARM at 7nm and 5nm and preparing to build machines for HPC and server markets.