The AMD Execution Thread [2007 - 2017]

Status
Not open for further replies.
What struck me about trying to read that article is that I simply couldn't bring myself to finish it.

That's partly the fault of writers (it's way too long), but also because the subject matter is so unexciting.

It's the problem of all AMD APUs: they can't compete against Intel in terms of pure CPU performance, so they are left to compete at a lower level, but that's a level where GPU performance don't matter much anyway, so it's very logical for OEMs to go for cheap.

I don't see how there's a way out of that.
 
Yeah you clearly didn't read the article.
I read the first 5, then some benchmarks (that were mostly without Intel context), then a few with Intel context, then a few pages looking for something of a conclusion (without much success.)

But it's all just boring stuff about a boring product for a very boring commodity market.

You clearly think otherwise. What did I miss?
 
Umm, they say HP Elitebook accepts dual channel and only comes with one 4GB stick, but didn't test it with 8GB?
Anyway, too bad. If single channel is enough for you then low Broadwell/Skylake's GPU is likely enough for you too.
 
Umm, they say HP Elitebook accepts dual channel and only comes with one 4GB stick, but didn't test it with 8GB?

Because the point was not to evaluate the potential of the platforms on the market right now (Ian actually stated that this is coming in a future article) , but rather the off-the-shelf user experience that current Carrizo solutions offer.
Which makes sense because the amount of people who know they'll get a decent chunk of additional performance if they just slap in another RAM module is probably negligible.
 
Eh well almost anything out there is satisfactory for the majority of computer users. I know people who are perfectly fine with the worst TN panels I've ever seen. In fact I know of one Llano and one Richland owner rocking those. Lots of people are satisfied with a phone or tablet that a Core 2 setup runs circles around.

I haven't a clue how to build the cheap notebook that appeals to uninformed buyers. Color scheme? Speaker brand name? They don't really care about Intel or AMD stickers in my experience.
 
Last edited:
Umm, they say HP Elitebook accepts dual channel and only comes with one 4GB stick, but didn't test it with 8GB?
Anyway, too bad. If single channel is enough for you then low Broadwell/Skylake's GPU is likely enough for you too.

If you want to see performance results with single channel vs dual channel, see the following:
http://browser.primatelabs.com/geekbench3/compare/4619972?baseline=5160760

I ran that test yesterday on my HP 725 G3. I have dual channel enabled obviously. I do think it's weird that Ananadtech didn't put in a single dual channel config just to make the article feel more complete.
 
AMD Zen will get 8 channel DDR4 support and SMT says CERN employee

An interesting snippet of news just before the weekend starts, AMD ZEN based processors will get up-to 8 channels of DDR4 support as well as SMT. The info was spilled by CERN engineer Liviu Valsan who in a recent presentation on datacenter hardware trends shared a thing or two about AMD's upcoming Zen processor architecture.
....
The Zen family processors for consumers will feature a unified AM4 socket with its GPU-equipped "Bristol Ridge" APU counterparts, and feature DDR4 support and a 95W TDP. While newer roadmaps don't confirm the TDP for desktop products, they suggest a range for low-power mobile products with up to two Zen cores from 5 to 15W and 15 to 35W for performance-oriented mobile products with up to four Zen cores.

index.php

http://www.guru3d.com/news-story/amd-zen-will-get-8-channel-ddr4-support-and-smt.html
 
The presenter hints at the 32-core 8-channel variant is an MCM or similar pairing of two 16-core quad channel units, using the inter-die communication bus to tie the halves together.
There could be something lost in translation on how exactly it was to be implemented, as the wording sounded almost like it was still one chip. He might have meant one package, although there is precedent from Intel where the first P4 dual cores were just single cores that were cut out in pairs from the wafer, then tied together via bus lines within the package. It's not necessarily a native solution, but could still be considered one chip/die.
 
AMD already implemented MCM with Opteron 6300 series -- two Bulldozers for total of 16 cores and four DDR3 channels in a single package.
 
The presenter hints at the 32-core 8-channel variant is an MCM or similar pairing of two 16-core quad channel units, using the inter-die communication bus to tie the halves together.
There could be something lost in translation on how exactly it was to be implemented, as the wording sounded almost like it was still one chip. He might have meant one package, although there is precedent from Intel where the first P4 dual cores were just single cores that were cut out in pairs from the wafer, then tied together via bus lines within the package. It's not necessarily a native solution, but could still be considered one chip/die.

At the moment it is still on 1 socket ( whatever it is MCM or not ), it can be considered as a single chip.. ( its like that the OS will see it anyway ).
 
It seems to me it is both : it works as if it were an MCM, but it is a single die chip.
There is an actual precedent, consoles can be seen to have two quad-core Jaguar CPU rather than an eight-core Jaguar CPU, which doesn't exists.

Making a straight 16-core CPU with cache coherency and the cores speaking on equal terms so to speak, that is likely hard but easy enough to get going. 32-core CPU is much harder in terms of internal bandwith, data flow. Perhaps AMD would be able to manage it if they were doing an architecture more similar to Xeon Phi, but that's not the goal here. So Zen goes to 16 cores, but not 32.

To get 32 cores, paste two 16 cores CPU next to each other on the same die. Probably with very fast on-die NUMA links. Classically, you use an MCM or separate sockets in that situation, but I suppose that here you have enough transistor budget to get away with that. Benefits are you don't need expensive packaging, and perhaps higher bandwith or lower latency between the two CPUs.
And the OS really needs to be aware of it (e.g. good old NUMA awareness). Some tasks/processes/threads should be scheduled to run on the same CPU together, as appropriate.
For per-socket licensing it's fair game that it's considered as a single socket.
 
AMD continues to lose GPU market share to Nvidia.

The latest JPR Report: http://www.jonpeddie.com/publications/market_watch

shows that AMD only has 11.5% market share (down 2.7% Y-to-Y) whereas Nvidia now has 15.7% (up 1.6% Y-to-Y).

Nvidia now has 4.2% more market share than AMD even though Nvidia only sells discrete GPUs whereas AMD sells both discrete GPUs along with APUs and Chipsets with integrated graphics.

EDIT: This report seems to be for 2015Q3 not 2015Q4. It would be nice if JPR put a date on these reports.
 
Last edited:
AMD continues to lose GPU market share to Nvidia.

The latest JPR Report: http://www.jonpeddie.com/publications/market_watch

shows that AMD only has 11.5% market share (down 2.7% Y-to-Y) whereas Nvidia now has 15.7% (up 1.6% Y-to-Y).

Nvidia now has 4.2% more market share than AMD even though Nvidia only sells discrete GPUs whereas AMD sells both discrete GPUs along with APUs and Chipsets with integrated graphics.

EDIT: This report seems to be for 2015Q3 not 2015Q4. It would be nice if JPR put a date on these reports.
That's curious, since AMD actually gained back a bit of their share in discrete cards during Q3, and IIRC they gained on mobile solutions too?
 
That's curious, since AMD actually gained back a bit of their share in discrete cards during Q3, and IIRC they gained on mobile solutions too?

Both AMD and Nvidia took market share from Intel who went from 75.2% to 72.8%.

Nvidia just happened to take more share than AMD did and also Nvidia's share gain was in high cost units thus more revenue for them.
 
Status
Not open for further replies.
Back
Top