The AMD Execution Thread [2020]

I skimmed through the article and missed the date as well, it says mid 2021 deployment so yeah definitely not Zen 4. It's not necessarily Milan though, could be the "custom" CPU they are designing for Frontier.

Edit: From the Anandtech article - "Forrest clarified that this custom CPU is not in the same way that Intel defines custom – i.e. it’s not simply the same silicon with adjustments in core counts / frequency / cache. The CPU for Frontier will be a fully custom design, built with CPU-to-GPU IF links in mind, without any excess. When asked if IF is going to be a connectivity in other platforms, Forrest would only confirm that it’s the connectivity for Frontier. Though, for what it's worth, AMD announced back at the Frontier unveil that the CPU would ultimately become available for other enterprise customers as well."

 

https://www.amd.com/en/press-releas...ing-the-industry-s-high-performance-computing

AMD acquire Xilink

https://ir.amd.com/news-events/pres...-quarter-2020-financial-results?sf239269791=1

third quater financial result

eraning slide
https://d1io3yog0oux5.cloudfront.net/_1c911660cd1a84222da249749d89af54/amd/db/778/6608/file/AMD+Q3'20+Earnings+Slides.pdf

 

Paid in full as AMD shares. Funny, someone suggesting the same a year ago would have been laughed off the planet

 

All stock was the only way it made sense for AMD and I have to say it's a good use of their stock value at the moment.

Earnings were solid as well though I expected the projection for Q4 to be even higher. Perhaps Milan revenue will show up more by Q1 2021.

 

So how does this work? How much % do Xilink shareholders now own of AMD?

 

Combined value is estimated 135billion dollars. Stock transactions is 35billion. That would give xilinx owners 34.7% of amd. (35/135) But I wonder how things would change if amd stock gains or looses. Would xiling still get 35bilion? What if amd value doubles or halves before transaction goes through?

 

Your math is off quite a lot there, 35/135 is 25.9%, not 34.7%.

 

From the press release - "Under the terms of the agreement, Xilinx stockholders will receive a fixed exchange ratio of 1.7234 shares of AMD common stock for each share of Xilinx common stock they hold at the closing of the transaction. Based on the exchange ratio, this represents approximately $143 per share of Xilinx common stock2. Post-closing, current AMD stockholders will own approximately 74 percent of the combined company on a fully diluted basis, while Xilinx stockholders will own approximately 26 percent. The transaction is intended to qualify as a tax-free reorganization for U.S. federal income tax purposes."

 

Another SYCL implementation bites the dust, I guess they might soon be working on the ROCm stack ...

 

This is a big news. Also intresting Nvidia ist Emulating GPUs on Cadence which use Xilinx FPGa
https://www.cadence.com/ko_KR/home/...-x1--the-first-scalable--data-center-opt.html

 

Not really. Having models of a given chip design on FPGAs first is standard practice in the industry.

 

You are right. But think about GPU size and which size FPGa is needed to simulate them. I think Xilinx and Intel hava a (Duo)Monopol on such big FPGas.

 

A full upper end GPU will never fit on a single FPGA anyway. Of course, it is still helpful to use larger FPGA for this purpose
Also, for a given design there won't be many FPGAs used anyway. (a lot) Less than 100 would be my estimate. So the sales volume this generates is quite small.

 

In the Video the NVIDIA guy is saying that it’s not so difficult to design the circuits it is more an issue of time consumption. When you now take in consideration that AMD has now the first and fastest access to the newst simulation technology, this will give AMD a huge advantage in reducing development and production time. Also they get these technic at a really low coast.

Because they geht the fpgas now cheep they will get more and can now simulate more, which also saves time and rise quality and innovations dramatically.

 

You're implying that AMD seeks vertical integration (buying up their suppliers, e.g. what broke Bell), instead of horizontal service diversification. That seems unlikely. Sure AMD might have it easier to access Xilinx stuff, but this convenience must surely be a drop in the ocean relative to what Xilinx does for other customers; basically AMDs so far interaction with Xilinx is extremely small in comparison to Xilinx market. I would assume the vertical part played zero role in the decision.

 

Your design can be split among many FPGA's, you don't need just one big one.

 

Cadence Design Systems is the provider for the prototyping and design suite, which happens to use Xylinx chips as a component. Having just the chips leaves out much of what goes into the overall product, sort of how AMD makes x86 chips but that doesn't mean it has advance access to Amazon cloud servers.

 

@3dilettante at a high volume product you are righ, but we are talking about a key product which don't have a big marked share. Maybe in a few months we will see special contracts between cendence and Xilinx which gives AMD maybe a 1 year lead to order the newest technology.

 

The time consumption issue can be mitigated by using faster FPGAs. And FPGA cost isn't particularly a large issue for a company like AMD though, and since they're in house, the price is irrelevant. Getting them cheap would be taking money from one pocket and putting it in the other.

Cadence has a number of other customers who wouldn't take kindly to that so why would they agree to it?

There are plenty of other reasons for the acquisition:-

1. Excellent use of their current stock price. All stock deal so no debt and immediately accretive to earnings
2. Broad synergies across design, infrastructure, software, sales and marketing, etc
3. Integration with data center products to compete with Intel/Nvidia
4. More leverage and better pricing with TSMC as they become the largest customer

 

That would potentially require Cadence to delay its product pipeline for a year, or AMD would be willing to use products a year away from being considered ready.
I'm not sure what would make Cadence do this, unless AMD threatened to cut them off. Other than that kind of threat, I'm not sure what would have stopped AMD from negotiating an early-access agreement before now.