The AMD Execution Thread [2018]

And third-party IP connected to the semi-custom IO die for a (slightly more than) semi-custom MCM.

 

Delet this pls.

 

Pretty much everything besides stacked memory is possible.
FPGA chiplets with soft IFOPs?
I'd expect something like that for MS.

 

Milan has been described as being part of a chain of compatibility from Naples to Rome to Milan. While there can be features or speeds that can improve, the socket and platform compatibility would seem to rule out significant changes to the quantity of external connections for Milan.

 

Agreed, but without knowing the intended workload it's difficult to judge. The concern I had was what portion of HPC CPU loads rely heavily on vectorised FP64 that won't overlap with GPU acceleration. Nice to have sure, but so is the added memory and IO scaling.

Both, but my point was AMD could create a new custom IO die for embedded or a different socket entirely as mentioned above with the older process and less complicated circuitry. Even then varying just the cache size might be a possibility. Or updating to different memory controllers.

Agreed, as my speculation likely would mean a different socket with the same 7nm processors. That said they could make adjustments to the IO die in process and configuration while maintaining the socket. Not saying they will, just that the possibility might exist for some corner cases if the IO die is relatively affordable. Even a move to a 7nm, or whatever process is suitable, in a year or so to improve cache could work. Probably not Milan, but could be interesting. With the design, is Milan the entire package, IO, or compute cores?

 

It seems from recent history that the name give is based on the overall configuration of chips, socket, and package since names change as the products vary along those axes. As such, if there's something with a different socket or package with different components, I would first expect it to have a different name.

 

AMD promises that OEMs will issue more driver updates for Ryzen Mobile

https://www.guru3d.com/news-story/a...sue-more-driver-updates-for-ryzen-mobile.html

 

That's crazy. That sounds like the mid-2000s notebook driver situation. Or like the awful driver updates of the AMD Hybrid Crossfire crap. Steer clear people.

 

It's not just AMD, I still can't just go and download IGP drivers from Intel for my Lenovo ThinkPad and install them without doing it manually via device manager. This thing doesn't even have any hybrid/switching/whatever graphics, just the Intel IGP

edit: and for clarification, this is not mid-2000's laptop, i5-6200U with HD Graphics 520

 

Yeah that's true. I've seen that with Intel as well. The difference is playing the latest games is not much of a priority with Intel though heh.

When I said mid-2000s I was referring to how you always had to wait for the notebook manufacturer to provide drivers, or use 3rd party hacks to install the latest ATI/NV drivers.

 

In the aftermath of the collapse of mining, NVIDIA gained a 10% market share on rival AMD in dGPUs.



https://www.jonpeddie.com/store/add-in-board-report

 

It's interesting how conflicting the reports are within same company
https://www.jonpeddie.com/press-rel...tch-reports-gpu-shipments-up-in-q318-from-las

Of course this is IGPs and laptops included, but according to this while both AMD and NVIDIA lost share to Intel, NVIDIA lost more

 

Lost more to it's own previous percentage, If NVIDIA sells 100k GPUs, and it lost 15% it still sells 85K. AMD can sell 25K and lose only 10%, but NVIDIA still sells way more even though it loses more.

 

You got it wrong, I commented about market share changes, which are percentage points, not percentage of share. AMD lost ~0.6 percentage points (it should actually be 0.53), NVIDIA lost 0.97 percentage points while Intel gained 1.5 percentage points

 

They are talking percentages (%), not percentage points.

It's the same principle, NVIDIA and AMD lost percentage in a different pool (dGPUs + iGPUs). When it comes to dGPUs alone, NVIDIA's loss of 1% is way less than AMD's loss of 0.5%, because NVIDIA sells more GPUs and controls more market share.

It can also be explained by AMD not really shipping any dGPUs this quarter because of their excess inventory.

 

No they're not, they're percentage points as clearly shown by the fact that AMD + NVIDIA lost share equals Intel gained share (for whatever reason they're only showing 2 decimal points on NVIDIA, don't know why), just like every quarter before.
I'm not familiar enough with the english/american traditions on how to mark percentage points in text, but these clearly are percentage points and not percentage change from their previous market shares.

edit:
For claritys sake:
Q2
AMD’s overall market share decreased -0.2% quarter-to-quarter, Intel’s total market share increased 1.6% from last quarter, and Nvidia’s decreased -1.4%.

In Q1 they still used the handy graph instead


From those numbers, the shares should now be (rounded) AMD 14.1%, Intel 69.7%, NVIDIA 16.1

Shipment changes are in percents, market share in percentage points

(and for what it's worth, if we'd read it literally like it's written, "AMD’s market share from last quarter decreased -0.6%, Intel’s increased 1.5%, and Nvidia's market share decreased -0.97%.", both AMD and NVIDIA should have actually gained share since if you decrease something by negative amount you're actually adding more)

 

Actually for Q2 AMD's overall market share decreased by 2 percentage points and Nvidia decreased by 1 percentage point.
The chart @DavidGraham quoted is just discrete gpu marketshare.

 

I have no idea where WCCF got their graph or data, but JPR clearly states about overall marketshares this:

That graph is not part of the public market watch report at least.
This is for all graphics, not just discrete.

 

Well, at least for the latest results which is Q3'18 JPR states for discrete graphics as is indicated in their table.

Edit:
Maybe we're talking the same thing ... the latest JPR report is only showing discrete market share numbers.

 

So does my PC with an onboard intel and discrete nvidia card count for both or just the nvidia?