AMD Tech Day at CES: 2018 Roadmap Revealed, with Ryzen APUs, Zen+ on 12nm, Vega on 7nm
by Ian Cutress on February 1, 2018 8:45 AM ESTRyzen APU Overclocking: A Focus on Memory Support
Also part of the demonstration about the new APUs was overclocking. For a number of years, AMD has hired Sami Makkinen, a well-known former world-record holder in various overclocking categories, to help tune the overclocking on some of its processors. So despite this being an integrated graphics platform, Sami was on hand to show the day-to-day overclocking potential. I was told that liquid nitrogen overclocking might be held at a later date…
Nonetheless, we were shown the system that would be distributed to reviewers for launch day: a mini-ITX motherboard, a Ryzen 5 2400G processor, 2x8 GB of DDR4 memory, and an AMD Wraith cooler. Using the popular benchmarking tool 3DMark Fire Strike, Sami showed a 39% increase over stock performance by overclocking only the memory and the integrated graphics:
- DDR4-2400 + 1250 MHz GPU: 2911 Points - 'stock'
- DDR4-3200 + 1250 MHz GPU: 3322 Points (+14%)
- DDR4-3200 + 1550 MHz GPU: 3596 Points (+24%)
- DDR4-3600 + 1675 MHz GPU: 4048 Points (+39%)
This shows a couple of interesting data points. First, that the DRAM speed directly influences the results a lot: the 14% gain moving to DDR4-3200 and a similar gain again up to DDR-3600 means that we are still dealing with a graphics system that is memory bandwidth limited, even with only 11 compute units. The Infinity Fabric communication pathways are tied to the DRAM speed, so these would also get a speed up – when asked if it would be possible to discern how much of the speedup is due to a pure DRAM improvement and how much from the Infinity Fabric boost, I was told that AMD could do it in house, but it would be impossible for consumers to do.
However, the DRAM speed gains match up with what we have seen with previous generations of AMD APUs, but also it means that Intel’s decision to combine a Vega GPU with high-bandwidth memory is going to be the way forward in this market.
The second point that this data brings is about memory support. One of the major criticisms of AMD’s initial launch of Ryzen was the lack of high-performance memory support, due to a young firmware. AMD has stated that the Ryzen with Vega graphics line of processors have better memory support with newer firmware revisions, capable of driving higher memory speeds. This is, of course, important to APUs. Additional updates will be coming with the new X470 motherboards in Q2.
Facebook
Twitter
RSS
131 Comments
View All Comments
A5 - Monday, January 8, 2018 - link
Anything that is forbes.com/sites/* is not credible. You can literally go sign up for one right now if you want.And Forbes proper isn't credible on anything outside the financial industry.
iwod - Monday, January 8, 2018 - link
I mean Christ, you have financial industry reporter reporting on tech they barely understand and having reader that credit a non technical Journal pieces as creditable source and then attacking a journalist on a technical information site labelling him or her and the site as not creditable and when the journalist has a bloody Oxford PhD in the ElecChem Field.LurkingSince97 - Thursday, January 18, 2018 - link
LOL.GloFo announced LONG AGO that they were skipping the 10nm node and going straight from 14nm to 7nm. They mentioned that there would be some 14nm tweaks.
Other competitors started tweaking their 14nm / 16nm nodes and in some cases re-branded them (tsmc, IIRC rebranded a 16nm improvement as 12nm).
There is more than one feature size measurement in a process, being able to shrink a small subset of these by a little bit will allow for some increased density. Add on some other tweaks and you can lower power too. What do you call a heavily tweaked node that only shrinks a few of the feature sizes but not all? What number is half way between 14 and 10?
It doesn't really matter what they call it. Its ~ 10% faster or 10% lower power, with ~10% to 15% higher density.
The 12nm process is essentially a 14nm+, or ++. Its not a wholly new node. Fact 1: it will be incrementally deployed as a revision to the 14nm process where you start over nearly from scratch and almost the entire manufacturing pipeline is new. Fact 2: Unlike most minor tweaks, it does increase density. There is some justification in giving it a new size due to fact 2, but it is not a brand new process due to fact 1.
A truly new node takes _years_ to develop, not months. This appeared on their roadmap suddenly. Everyone and their brother (other than you) knew that they had a 14nm+ in the works -- AMD's roadmaps had "Zen+" on a GloFo "14nm+" node. Suddenly there was no more "14nm+" on AMD's roadmaps, and it was replaced with "12nm".
Significant tweaks to an existing node however,
LurkingSince97 - Thursday, January 18, 2018 - link
Ugh, I should proof-read:Fact 1: it will be incrementally deployed as a revision to the 14nm process where you start over nearly from scratch and almost the entire manufacturing pipeline is new.
Should say:
Fact 1: it will be incrementally deployed as an upgrade to the 14nm process (many of the manufacturing steps remain identical), while an entirely new node typically implies replacing almost the entire manufacturing pipeline.
Sane Indian - Monday, January 8, 2018 - link
"Acer will use the pre-announced higher-end APUs, the Ryzen 7 2700U and Ryzen 5 2500U, but will also be pairing this with a Radeon RX 560 graphics chip."Does iGPU and dGPU can crossfire or they just independently (switch from iGPU to dGPU and vice versa) depending upon power situation.
Ian Cutress - Monday, January 8, 2018 - link
Literally the rest of the paragraph answers the question.We were told by AMD that the integrated graphics and discrete graphics will be used in a switching context: for video playback, the lower power integrated graphics is used and the discrete is disabled, however the discrete graphics is fired up for gaming work. For compute, or for games that support multi-adaptor DirectX 12 technologies, both the integrated graphics and the discrete graphics should be available, however this is up to the game/software to implement.
Kevin G - Monday, January 8, 2018 - link
AMD executing right now and while there is argument that they've played it safe with Zen so far, they are in a very favorable position in the market. -"While most of the roadmap could have been predicted by those of us embedded in this industry, it was good to see AMD volunteering a lot of information. This can be a bit of a double-edged sword, if a competitor knows what you have planned". AMD knows that Intel's roadmap is in chaos as right now and pushing existing Zen designs to lower power envelopes and/or increasing clock speeds will close the performance gap. If Zen 2 advances in terms of IPC by any significant margin, they could actually pull ahead in 2019 but that entirely depends on Intel's response, in particular Ice Lake.Intel's missteps with 10 nm has left them open as the first Cannon Lake parts were due in late 2016! Meltdown and Spectre meltdown could have been the death blow to Cannon Lake as the incoming lawsuits make it unwise to release products with such significant flaws. Intel's server line up is also in chaos due to Meltdown/Spectre as Cascade Lake was pinned in as a stop-gap 14 nm solution before Cannon Lake-SP in 2019. Cascade lake for release this year should already be sampling but Meltdown/Spectre could force a delay. In particular is that Cascade Lake is supposed to support Optane DIMMs. The 10 nm Knight's Hill HPC processor has been cancelled roughly a month ago and replaced by Knight's Mill as a stop-gap for partners looking into AI. We'll probably find out later today when Intel's CEO takes the stage at CES.
wumpus - Monday, January 8, 2018 - link
"Death blow"? You seem to vastly overestimate the amount of silicon GoFlo can produce. Even if everything goes right for AMD, Intel will be shipping Cannon Lake chips left and right to consumers (the chips in the article appeared mostly for consumers) as GoFlo furiously produces Zepplin (or whatever its followup is called) and AMD ships them as hyperprofitable EPYC chips.That said, Intel certainly takes AMD competition seriously enough to throw together competing products such as the X299, even if such products make the rest of Intel look like they have no idea what they are doing. But still, massive gains for AMD will look like a tiny blip in Intel's revenue.
Orenj - Monday, January 8, 2018 - link
With Zen 2 already complete, I'm wondering if it contains any hardware assists for mitigation of Spectre-type attacks a la Intel's proposed IBRS/STIBP/IBPB?PixyMisa - Monday, January 8, 2018 - link
Unlikely; they've only known about this for a couple of months.