It looks like Xilinx’s Versal ACAP chips are going to complete Samsung’s 5G base station offerings. SemiAccurate thinks the only loser from this is Intel but they made their own bed, then set it aflame.
Samsung has been cleaning up on the 5G base station market going from near zero design wins to >30% of the market a year ago, and vastly higher from there. Why? They have the best 5G offering in the west, assuming the politically based Huawei lockout continues. Who loses from this? Ericsson, Nokia, ZTE, all the western carriers, 5G users, and Intel. Of that list, there is one name responsible for the mess and the related squelching of commentary, and there are three winners, Xilinx, Marvell/Cavium, and Samsung.
Lets start out with today’s news before we get in to the commentary. Xilinx and Samsung are teaming up to use the very slick Versal ACAP platform, an FPGA with a lot of added hard functions and programmable fabrics. It is a really slick design and several of the variants are aimed directly at the 5G base station market.
One part missing
If you go back and look at SemiAccurate’s coverage of Marvell/Cavium’s latest Octeon and Fusion offerings, the very last slide is about a 5G base station, also pictured above. Although Marvell would not talk directly about their customers, SemiAccurate has known for years that Samsung was the big fish using their chips, and using them heavily. On the left of the diagram there are lines that go to the “RU” or Radio Units that use unspecified hardware. Can you guess where this is going?
If you said that the unspecified RU hardware is in fact a Xilinx Versal, specifically one of the SKUs meant for 5G operation, you win… nothing, it was too easy. There seems to only be one platform out there that can do the heavy computation needed for beamforming in realtime at power levels needed for energy and thermally constrained locations, that is Versal.
Xilinx is saying the platform is shipping to key customers now and will be widely available in Q4. To SemiAccurate this means that Samsung is buying up lots of 5G Versals and that rate won’t abate until later in the year. If you know how redundant and repairable FPGAs and Versal-like architectures are, you will quickly come to the conclusion that there are no production or yield issues. Better yet the software for it can be updated on the fly, that is the beauty of FPGAs, anything you miss at launch is quite doable later on with minimal effort.
If you are using 5G in the west, you are probably using the Xilinx/Marvell/Samsung combo if you are getting decent throughput. If your 5G service is as slow as or slower than your previous LTE coverage, you probable have the Intel ‘solution’. Before you think this is because the much touted 10nm Snow Ridge platform isn’t up to performance levels that Intel claims, you would be wrong. Not wrong that it is up to claimed performance, that part is still an open question because it isn’t shipping. No, the 10nm 5G wonder-chip that was promised literally years ago is still AWOL. It has been teased, hinted at, and everything else but shipping in real, promised quantities it is not. The multi-year ‘stopgap’ solution is barely functional but in many cases used for reasons other than performance.
This has lead Samsung, and by proxy Marvell and Xilinx, to go from roughly zero design wins to 30+% of the market in Q1 of 2019 to far more than that now. Intel keeps promising parts and then delaying and delaying and delaying. How bad is it? See this and this.
That was two years ago and similar things are happening at Nokia and ZTE who also were promised Snow Ridge years before that 2018 debacle. Intel has yet to deliver and likely will never be able to deliver anything close to the promised functionality in volumes that matter to the customers and their carrier partners. Xilinx, Samsung, and Marvell are cleaning up in the mean time.
So if you wanted to take the worst possible view of the Xilinx Versal ACAP parts, you could say they have no competition so they sell. If you were realistic you could say they delivered a very capable, versatile part that is both aimed directly at the 5G radio head market and is future-proof with it’s configurability. The competition continues to have curious excuses but no silicon, and their customers take the blame from carriers but don’t comment publicly. Kind of makes you wonder why, doesn’t it?S|A
Latest posts by Charlie Demerjian (see all)
- Intel’s Alder Lake is needless complexity - Oct 27, 2021
- Apple does something really interesting with the M1 SoC - Oct 27, 2021
- AMD’s Bergamo is a very interesting CPU - Oct 26, 2021
- Rambus releases 2nd Gen DDR5 RCDs - Oct 13, 2021
- What’s going on with Intel datacenter GPUs? - Oct 6, 2021