Monday, November 18th 2019
VIA CenTaur Develops a Multi-core x86 Processor for Enterprise with in-built AI Hardware
Tasting Intel's blood in the water with AMD's return to competitiveness, dormant x86 licensee VIA wants to take another swing at the market, this time with a multi-core processor targeted at enterprises and possibly workstations, developed by its subsidiary CenTaur. The company appears to want to cash in on the AI boom, and could develop turnkey facial-recognition CCTV solutions with the chip. CenTaur is ready with a working prototype. It features eight 64-bit x86 CPU cores, and an on-die "AI co-processor" named NCORE. A ringbus connects the eight CPU cores and the NCORE with the processor's other components. The processor features 16 MB of shared L3 cache, a quad-channel DDR4-3200 memory interface, and a 44-lane PCI-Express gen 3.0 root-complex, along with a fully integrated southbridge, making it an SoC. It also appears to be multi-socket capable, although VIA didn't detail the interconnect in use.
The NCORE is a PCI-mapped device to the software, which provides functions such as DNN building and training acceleration. From the looks of it, there's more to NCORE than simply a fixed-function hardware that multiplies matrices. Its developers state that the device accelerates AI at a rate of "20 trillion AI operations/sec with 20 terabytes/sec memory bandwidth." The CPU cores on the processor tick at 2.50 GHz, and while VIA hasn't made any IPC claims, it has mentioned support for the cutting-edge AVX-512 instruction-set, something even "Zen 2" lacks, which possibly indicates a powerful FPU. The silicon measures 195 mm², and has been built on 16 nm FinFET node at TSMC. VIA will demonstrate the unnamed processor and its testbed at ISC East 2019, held on November 20 and 21.
The full technology announcement slide-deck follows.
The NCORE is a PCI-mapped device to the software, which provides functions such as DNN building and training acceleration. From the looks of it, there's more to NCORE than simply a fixed-function hardware that multiplies matrices. Its developers state that the device accelerates AI at a rate of "20 trillion AI operations/sec with 20 terabytes/sec memory bandwidth." The CPU cores on the processor tick at 2.50 GHz, and while VIA hasn't made any IPC claims, it has mentioned support for the cutting-edge AVX-512 instruction-set, something even "Zen 2" lacks, which possibly indicates a powerful FPU. The silicon measures 195 mm², and has been built on 16 nm FinFET node at TSMC. VIA will demonstrate the unnamed processor and its testbed at ISC East 2019, held on November 20 and 21.
20 Comments on VIA CenTaur Develops a Multi-core x86 Processor for Enterprise with in-built AI Hardware
But the selling point of the NCORE would be that those 8 other cores could be freed up to do other things, and yet, there isn't a whole lot of information on how well those other 8 cores can do general purpose compute, which seems kind of important given the stance they've taken here. And what's "high" IPC relative to? 2.5GHz is a rather middling max clock.
I guess this throws more a wrench into AMD's works, since AVX-512 is a leg up, but Intel is the big player here, and they have those instructions too with a lot more cores for general compute. But those Xeon Silvers and Bronzes look ripe for the picking with half the AVX-512 hardware, if these really run 2.5GHz all-core.
@TheLostSwede piss poor marketing slides, I suspect. From the NCORE block diagram slide, it appears to imply that the NCORE is multi core and it runs at 2.5GHz, but another slide infers that the entire platform (cores + NCORE) runs at 2.5GHz.
It looks like it's only the NCORE that runs at 2.5GHz, which technically isn't one core in the traditional sense.
Also keep in mind that this is only a reference design so far, not MP, so they might still be able to make it run faster.
From a chat with one of their guys a while ago, they're apparently getting similar IPC to Ryzen, so it should still be pretty decent in terms of performance.
I guess there's a reason why this is an SoC and doesn't need a chipset, as VIA was never very good at supporting Centaur...
This thing is about a year late, but I believe this should be a lot more competitive, if not quite up there with AMD and Intel in terms of overall CPU performance, but as I said, the Centaur people I've spoken to (that I've known for well over a decade) told me that the IPC should be very similar to Ryzen.
From Wikipedia:
PCIe 3.0
SoC
Edit:
I'm wrong. This has a GPU instead of the NCORE, a dual channel memory interface and is probably still using Nano cores.
source: browser.geekbench.com/v5/cpu/526995
EDIT: disregard the above; I'm a brainlet
browser.geekbench.com/v5/cpu/compare/526995?baseline=562975
VIA implemented SSE4.1 even before AMD could. Now it's going to beat AMD to AVX-512.
Thanks for correcting me! :toast:
[MEDIA=instagram]B5GCSZkhrL2[/MEDIA]
Still going by the "just faster than Goldmont Plus performance", it seems that CPU is still using a core descendant from the Isaiah core.
They did make some pretty cool chipsets though at times. K7 with 4 slots of dual channel ram and sata built in... mind bending... lol
*still hates Nforce with a passion*