AMD’s third shoe drops at CES 2020: 7nm Zen 2 mobile CPUs
arstechnica.com
arstechnica.com
> Most recently, due to various marketing and discrepancies among foundries, the number itself has lost the exact meaning it once held. Recent technology nodes such as 22 nm, 16 nm, 14 nm, and 10 nm refer purely to a specific generation of chips made in a particular technology. It does not correspond to any gate length or half pitch. Nevertheless, the name convention has stuck and it's what the leading foundries call their nodes. Since around 2017 node names have been entirely overtaken by marketing with some leading-edge foundries using node names ambiguously to represent slightly modified processes. Additionally, the size, density, and performance of the transistors among foundries no longer matches between foundries. For example, Intel's 10 nm is comparable to foundries 7 nm while Intel's 7 nm is comparable to foundries 5 nm.
https://en.wikichip.org/wiki/technology_node
So 7nm does not mean the chip has 7nm features, it means that the marketing team called it "7nm".
Everyone can check the frequency of the CPU so the one on the box mostly matches the one you'll get but no consumer can put the silicone under the electron microscope and measure all the transistor features and feel cheated that it doesn't match the advertised size.
Outside techies, most consumers don't know wtf nm is and how that correlates with performance, maybe some might even think that more nm is better, but on the other hand, most have a vague idea that more MHz is better.
That's one of the reasons why consumer grade tech has its specs totally different emphasized than the professional stuff. If you look at the ryzen box you won't find many specs but you'll find a VR logo. :)
Once, clock speed 'mattered' and was used in marketing to indicate a CPU that did work faster. There was a period of time where Hz was still used by marketing, yet it had stop reflecting how much work the CPU actually performed in a given amount of time. The public caught on that Hz mattered less.
Now, they're hoping nm matters in marketing (no as 'this is a faster chip' but 'this chip uses 'better' technology than TheOtherGuys/ThePreviousGeneration/WhatEver) for at least a few moments before the marketing team has to come up with some other differentiator.
"Do not compare CPUs by process technology or frequency. If you need to compare, use benchmarks that show performance and efficiency and even better, do it on your own application because the results vary between applications."
Sort of. If you look at EC2 instance types [1], about 1/2 of them have a specific CPU related to them. Same thing with Microsoft's Azure instances [2].
There are a limited number of cpus available - they typically are xenons or AMD equivalents, so if you want data on consumer CPUS, the best you can do is try your app out in the cloud and then look at published benchmarks and try to estimate what your expected performance should be.
___
1. https://aws.amazon.com/ec2/instance-types/
2. https://azure.microsoft.com/en-us/pricing/details/virtual-ma...
The correlation of bare-metal cores vs cloud vCPUs is very much nonlinear.
So: is there any analogous new convention for naming process nodes, now that you can’t just refer to them by their size? If there’s a dropdown select box on some form internal to e.g. Apple, where an engineer specifies what process-node of what fab they’d like to use to print a new chip—and said list has had some thought put into making it intuitively unambiguous—then what would the items in that list look like?
AMD Threadripper 3990X : $3990
You also had to think about DX vs SX, if you got L2 cache, 5x86 is really a 486, what is an Overdrive chip?, why does a PR166 CPU run at 100MHz, what is a Celeron, and why is 300A model so much faster than the original 300? Etc etc etc
Obviously the names don't carry much weight, and speed can be checked directly instead of inferred from size, but I'm still curious about the actual state of the production issues.
https://www.tomshardware.com/news/amd-mini-pc-intel-nuc-ecos...
https://www.tomshardware.com/news/amd-mini-pc-intel-nuc-ecos...
Quite apart from that, the NUCs are not exactly cheap. Gaming-ready ones can be $1000+, for which you can build a very decently specced MiniITX rig around a real desktop graphics card.
I get the impression that the GPU in this computer is much more intended for CAD, video processing, light AI work, etc than the kind of 100% duty cycle loading that running a game demands.
Yes - most people don't have a lot of space for gadgets and things. Look at the advertising for Google's Stadia - it's partly that you don't need a console taking up space in your living room.
Most people never upgrade so don't care about that.
It's not quite that small but you might like this: https://github.com/phkahler/mellori_ITX
It's Also available on thingiverse: https://www.thingiverse.com/thing:3262778
Uses an ITX board but is 190x195x60mm. I'm waiting for the next gen APUs to pick a new board and do another iteration. Where are the 4000 series APUs?
The Velcase 3 has the interesting advantage that it can just use straight-up discrete GPUs (mITX style shorty cards) so you can just use a normal 3700X and say a 2070 or similar.
Yes, the lack of a socketed version of the new APUs was kind of disappointing, as was the lack of a B550 launch to support PCIe 4.0 on a cheaper board without a chipset fan.
Hopefully these follow sometime in the future but we may have to wait for computex.
I can see that. The original Mellori has been running over a year with no issues. The plastic is mostly 3mm thick, and the cooling solution covers most of the inside surface. It's my development machine, so it doesn't work as hard as gaming would push it. Compiling code will max all 8 threads for a minute or two, but the fan doesn't even spin up all the way.
>> the lack of a B550 launch to support PCIe 4.0 on a cheaper board without a chipset fan
Yeah, I was looking at the Aorus x570. It looks like the fan is attached to a NVME cooler. I could probably remove that and put a passive heat sink on. The case has good airflow right across that area when there's no GFX card in the slot - air has to go around the edges of the board and out the bottom. With an APU the only use for PCIe 4.0 is for the M.2 slots which are pretty fast even with 3.0. I'm also considering the Aorus B450I as a much cheaper alternative.
I've also considered having it printed from metal.
(Does anyone else fondly remember the short moment in time when all laptops and all projectors had 1024x768 as their native resolution and you could expect presentations to just work?)
The Asus Zephyrus G14 is AMD-only and is their 14" flagship which they are calling the most powerful 14" laptop. Complete with your choice of either a 1080p 120hz panel or a WQHD (probably 60hz) one. Both of which are Pantone validated. And 32GB RAM.
Or the other time where Nvidia didn’t ship their graphics chip to them with support for the old school Cinema Display. So they sat in warehouses for 18months while Nvidia produced the part and only then, Apple could start integrating.
Then Nvidia tried to sue Samsung+Qualcomm when the iPhone came out (because apple used GPUs from these companies) claiming patent infringement.
Then in 2016 Apple said “no” to Nvidia on putting their GPUs into their laptops due to power consumption to performance concerns. And apparently that was not a healthy conversation.
Nvidia is a bully, and Apple is a big player. I will not shed a tear over 10% performance in predominantly game workloads.
Intel on the other hand has an edge with thunderbolt right now, but that’s an open standard and Apple could produce an AMD machine with thunderbolt now. I would suspect that they’ll live with the lesser performance until ARM is possible.
Apple has always been using PowerVR for iPhone since day 1. Even the recent so called self designed GPU are pretty much 80% PowerVR with all PowerVR proprietary and patented features.
[0]: https://www.anandtech.com/show/15318/amds-64core-threadrippe...
We happen to be seeing a bump right now due to AMD migrating across two process nodes at once (and Intel having fallen a little behind), but that's just an instantaneous thing. The days of the 486 are still gone.
[1] My own dates. I'm bounding this roughly by the point where EDA tools caught up with process improvements (allowing straightforward die shrinks and scaling of logic) and the end of "free" frequency scaling due to thermal and power limits. Someone else probably has a different definition. Fight me.
I don't particularly remember feeling romantic about processor naming schemes; I do remember certain processors for being workhorses. However, I would agree that today, one needs to acquire an informed view to discern processor design and performance, variously across different architectures and manufacturers, over a period of time.
https://www.anandtech.com/show/14043/upgrading-from-an-intel...
If you feel lost within the Intel naming convention, at least they provide a very handy guide.
https://www.intel.com/content/www/us/en/processors/processor...
And the introduction of the 488sx shattered the dream of everyone finally having floating point hardware by default.
...wow
It's confusing and dishonest. Multiple times I've jumped at almost getting a Ryzen 3000 laptop only to remember that it's not really Ryzen 3000.
Most people who buy these laptops know nothing about Zen2, Zen1, Zen+.
I don't think you can go wrong focusing on gaming. This is my opinion of how Microsoft succeeded so well. The network effects push all of computing forward.
I feel like this just boils down to what we've known for a while: Intel is catering to an enterprise crowd while AMD is catering to the average PC enthusiast.
And of course you still need that local client to play it, which might as well be an AMD-powered ultrabook. At which point Stadia just resulted in AMD selling up to 3 products instead of just 1 (GPU & CPU in the server + APU in the laptop)
Stadia is for the person who has 1hr every few days to play a game casually on the TV or 20 minutes during lunch.
It caters perfectly for the crowd of people who do not want to be locked to a specific device (gaming pc) or update their console/games.
The high static time cost of gaming in 2020 is what stadia mitigates.
And anyway. Stadia is powered by AMD underneath.
Probably good only for Google employees who are forced to use Chromebooks.
Let's put it this way: My taste for products that I can't not use is rather different from the majority of society.
As long as you can wait and upgrade when that major leap happens you're good for a few years.
That's why we follow tech news, leaks and rumours, to estimate when the next major leap will be and not be suckered into buying hardware that will be too quickly obsolete.
That's why HW manufacturers have NDAs in place. To keep the consumers buying the current stock instead of holding out for the next gen.
Example: PCIe 4.0 has insane bandwidth allowances. I don't see any SSDs ever going beyond PCIe 4.0 bandwidth. Other iterative improvements like a few dozen MHz more in CPUs, or a few dozen more cores on the GPU, or a few more hundred MHz in RAM... they are usable without their potential being wasted in the new chipsets -- although I am not an expert, it does seem that way after several reads of the their parameters.
And even if next-gen stuff with bigger improvements is coming, I don't feel we can go much higher before these things hit big diminishing returns in noticeable performance in all but very specialised programs. Currently I am using a workstation with 10-core 4.0GHz CPU / 64GB 2666MHz DDR4 ECC / NVMe SSD at ~2.8 GB/s. I seriously cannot make any programming software even utilise that SSD to full capacity -- only Postgres manages to saturate it to 50% of its potential when doing `pg_restore`.
Granted I haven't run any deep learning stuff and I don't intend to. I am focusing on everyday and professional-but-not-strongly-specialised software. And there, I feel, stopping at motherboards with PCIe 4.0 and their appropriate AMD CPUs and uber-fast SSDs and the fastest RAM you can find today... will be more than enough for like 5 years.
All IMO of course.
Even post-mitigations Sandy Bridge is still "good enough" for most people IMO, even though current chips are much nicer.
Zen 4 will be PCI-E 5.0 and DDR5. Unless Zen 4 is a gigantic leap forward in performance and not an iteration. The cost of Jumping in those tech will be quite expensive.
So Zen 3, PCI-E 4.0, DDR4 will be like the sweet spot for desktop for a while.
( I do wish I am wrong and they drive down the cost of PCI-E 5.0 and DDR5 faster. But history has shown those tends to take 2 - 3 years to pick up the pace )
Hell, even the current high-end PCIe 3.0 setups like mine seem to be quite future-proof.
It's an amazing machine in every way.
a coworker of mine has a generation 2 ryzen which runs games like factorio perfectly fine - big setups (according to him) are pretty smooth. no GPU. he says he hasn't yet found any games that can't run fairly well on it yet.
it's funny because desktops are now basically the thin clients we thought we'd all be using in a futuristic tech setup, with the actual work being done on "the cloud". your desktop isn't fast enough to do any real "work" other than some unit tests, compiling, etc.
I think my problem with Nvidia is they are ( for now ) hard to loathe. They were like the old Intel where their constant innovation just baffles me. And there is nothing but respect. ( Apart from their part with Apple's dispute ) But I have been an ATI Fans since early days. Still has an ATI Mach and Rage somewhere in my room.
That said, I have seen plenty of ARM powered Chromebooks and I don't think any of them had more than four cpus or were known for high performance. It's not impossible to make a high performance ARM chromebook, but it's not the market people are building for. High performance x86 Chromebooks happen because once you've built for a dual core mainstream x86 laptop processor, you can also easily solder in a higher performance processor with the same footprint.
So instead you'd need to ask how does Zen 2 compare to Cortex A76 or something like that. Which we don't know since we don't have the 15W Zen 2 mobile parts to benchmark yet.