![Zarathustra[H]](/data/avatars/m/9/9298.jpg?1723407683){kind=avatar}
- Joined
- Oct 29, 2000
- Messages
- 42,445
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Arm Takes Aim at Laptops: Cortex-A78C Processor for PCs Introduced
- Thread starter erek
- Start date
- Joined
- Oct 29, 2000
- Messages
- 42,445
If or when that happens, I am done with tech.
To me there are DIY Desktops and Servers which are the real computers.
Then there is everything else, including pre-built desktops and servers, AIO's, Laptops, tablets, phones, various silly IoT like smart lightbulbs, that stuff can just go die in a fire for all I'm concerned. Don't want one now, don't want one ever, don't care if I ever see one again.
The only thing that interests me in all of tech is customizing my own systems with desktop (or at the very least server) hardware, and having the physical hardware on premises with me, none of that cloud bullshit. Take that away, and I'll just stop caring.
I tend to be more optimistic though. I think desktops will certainly become even more of a niche than they are today, but I don't think they will be going anywere any time soon. There will always be people who need or demand big brute power.
- Joined
- Oct 29, 2000
- Messages
- 42,445
You are missing the point though.
There are almost as many different use cases for computers as there are people who own computers. There is no way even Apple is going to design or include an ASIC for them all.
I see ASIC's as nice little addons for niche use cases, and as boosters to make some tasks people used to not do on laptops because they took too long, now possible to do, but I don't see them ever supplanting the need for a big bad core, because no one will ever be able to ASIC the everything.
You might spend a small amount of time doing ASIC accelerated tasks, but the overwhelming majority of your time you are going to be stuck on the slow mobile trashcores.
I do hope your right. I think where loosing ground in the consumer space pretty quickly. But I do hope your right.
Just because its a ASIC doesn't mean its built for only one purpose. Being built for "this type of work" is the way its going. Apple doesn't tell developers what to do with their neural engine hardware. They won't be able to tell developers what to do with tensors. Just like Nvidia and AMD GPUs are used for math that is best accelerated on their hardware. No GPU company invented bit coin and mining in general. Apples add ons are specific... but there still processors that accel at a specific type of math. General compute isn't going anywhere. I just see its role being reduced greatly. Which means the entire debate of is ARM fast enough is sort of dated. Its smaller... and that is a massive advantage if you need to include other types of cores on a package. Intel will be doing the same with alder lake.... so the future of brute force Intel computing is highly in question now.
"Application Specific Integrated Circuit"
The whole idea of an ASIC is to be tied to a specific task, or set of similar related tasks. If Apple is trying to speed up certain task by building chips or portions of chips to accelerate them, they loose out on flexibility to perform general task on those portions/chips.
Not to say it won't work, when your market is fairly narrow, you can certainly aim to please the majority of your users with task specific hardware.
The whole idea of an ASIC is to be tied to a specific task, or set of similar related tasks. If Apple is trying to speed up certain task by building chips or portions of chips to accelerate them, they loose out on flexibility to perform general task on those portions/chips.
Not to say it won't work, when your market is fairly narrow, you can certainly aim to please the majority of your users with task specific hardware.
I don't think Apple would describe anything in their chips a ASIC accept perhaps for the HEVC hardware. They are ASICs in the same way a GPU is. Its not going to run a OS... but its going to accelerate a specific type of math. No one here can even imagine a machine without a GPU. Who knows what developers will do with Tensor cores on every Mac. No doubt when I bought my first ATI card and saw Quake II on hardware... I didn't think hmmm yes obviously in 20 years these cards will mine money.... nor even imagined them accelerating photoshop filters. Yet here we are. In a decade it may seem just as obvious all the things tensor cores will do on everyday machines.
I think the point being made by other posters is that to achieve the performance numbers that are being touted in this thread, they're including specialized hardware to accelerate those tasks. Sure, there might be tons of uses for tensor cores that we don't see yet. But that might be 5, 10, 15 or even 20 years down the road.
But, if you're apple, and your market is small enough to be niche, then I could see coupling some basic cores and task specific hardware to speed up your common use cases. It's not like most people need a fast laptop anyways, it always amazes me the number of people that will buy laptop, and then stick it on a desk and never move it.
That's fine, but comparing it against GPCPU speeds is BS.
- Joined
- May 7, 2007
- Messages
- 13,443
http://macchiatobin.net/product/macchiatobin-double-shot/
http://macchiatobin.net/compare/
There are quite a few others outside of this one as well.
ARM-powered laptops like the Pinebook/Pro, which are modestly customizable, have been selling out in droves, too.
Here is another:
https://www.gigabyte.com/us/Server-Motherboard/MP30-AR0-rev-11#ov
And another:
https://www.solid-run.com/nxp-lx2160a-family/honeycomb-workstation/
For those who enjoy RISC-V:
https://www.electronicdesign.com/in...nic-design-riscv-lands-on-miniitx-motherboard
Last edited:
Funny how you have missed we only discussed single thread performance in relation to measuring the IPC of a single core, not because only single thread performance is relevant; otherwise the chips compared wouldn't be multicores.
I did NOT wrote that any ARM system is customizable. So your list is so useless as if you list embedded x86 systems.
Here a tower-desktop system
https://www.anandtech.com/show/12571/gigabyte-thunderxstation-cavium-thunderx2-socs
"From hardware standpoint, the ThunderXStation looks rather versatile: it supports various expansion slots, enabling developers to use various add-on cards, accelerators, and storage devices required by their applications."
Maybe you singly discussed single thread performance in relation to measuring the IPC of a single core, but in doing so, you only use spec2006 which I and most of Real World Tech declared mostly irrelevant.
Single core Spec2006 on a ryzen 5950's 64mb L3 cache is as extremely skewed as it is on a M1 with 12mb L2. Unless you're measuring the memory subsystem which is a metric but not really an IPC metric. For when you're out fetching memory, you're not executing instructions.
Even worse is pretending a bunch of single threaded programs executing concurrently constitutes some semblance of multi-thread performance. You're basically pushing the mutex operation to the IO and thusly stalling each program serially.
I have no doubt the new ARM cores will be very competitive, maybe even crown worthy, but there is a long road ahead to get to the point where they can be measured on an even playing field. Even Apple projected years.
Last edited:
That HoneyComb looks very close to what is needed. I looked around and noticed some growing pains with the pcie system, but the fact that it's out there is big.
With the new raspberry pi 4 compute module board having a pcie 1x good work is being done with external devices.
Jeff Geerling pushing an Intel i340 NIC
State of the stack
Last edited:
- Joined
- Oct 29, 2000
- Messages
- 42,445
I thoroughly enjoy the concept of RISC-V. Imagine not needing to pay license to an ARM Holdings whether or not it is an Nvidia subsidiary, or AMD and Intel constantly bickering over the x86 license.
It certainly isn't ready yet, and unless the perennial catch 22 of software/hardware support catches up, it won't be of much use to anyone, but I'm thinking it really is only a one product cycle sprint of deep focus by a team, the likes of which have been led by Jim Keller in the past away from being truly usable.
With an ISA that is truly open, you could have any number of competitors, AMD, Intel, Cyrix/Via, you name it battling and leapfrogging each other without any IP impediments.
The fact that it won't happen is almost a certainty, but if it did, it would be the holy grail of a competitive and thus better advancing and performing CPU market.
Last edited:
- Joined
- Oct 29, 2000
- Messages
- 42,445
There are some obvious flaws with some of these, and they cost a pretty penny for pretty low clocked, low core count chips.
I like the focus on networking, with lots of SFP+ options, but most don't even tell you what chip they use, and the one that does indicates that it is Marvell. You can't trust any network chip not made by Intel, especially in a server.
PCIe expansion is present, but very limited.
I could see some of these be used in router setups or the like, as long as the poor network chips don't hold them back too much, but even the one labeled "Workstation" is not something I would call workstation performance.
2Ghz Quad Core with DDR3 support? We are talking an ultrabook in desktop board clothing here.
The only Workstation I imagine this would be even marginally useful for would be as a development platform for server ARM programs, and even then, you'd likely want to do all of your compiling on something else, as this would be dreadfully slow.
I think people also need to remember that trying to find the perfect benchmark isn't useful to end users. To us it matters what runs our stuff best and yes, that can end up being skewed. If all my programs are x86 only, and have to run on ARM in emulation and are slow because of it, well then I probably don't want ARM. All the arguments of "But if they just recompile and properly optimize they'll be great!" don't matter until it actually happens. Promises of amazing future improvements aren't of any use to people, it is about what gets you good performance for your dollar now. While benchmarks are a little useful for comparisons and are certainly of academic interest, it really isn't what matters to a person buying a system. I don't much care WHY something runs Cubase faster than the other or if Steinberg could fix it if only they weren't such meanies, etc, what I care about is having something that does run it well.
You know, I have to give Broadcom some support here: I think their new chips are fine. We got new Hyper-V servers about a year ago. We got Intel X722 adapters for inter-server communications (VM migrations, etc) but decided to give the 10gig Broadcom LOMs that came with the system a shot for the VMs themselves. They have worked well. VMQs works properly (we don't do SR-IOV), no system load problems, no latency issues, no crashing, etc, etc.
Their adapters seem competent, at this point. I certainly wouldn't go and throw your Intel NICs away, but if you buy a server and Broadcom is the onboard option, well give it a shot, it will probably be fine.
What happens when you effectively kill general purpose cores by adding on tons of ASICs and such only usable for limited tasks and something new comes along which doesn't have a built in ASIC? Do you simply do without being able to crunch that something new until you buy all new hardware? Right now that's not a problem because general purpose cores can handle it even if it's not as fast. Maybe this is the point of pushing to specialized, limited ASICs and such. The moment something new comes out you have to replace a bunch of hardware or you're left behind.
- Joined
- Oct 29, 2000
- Messages
- 42,445
Back in the day my statement used to be "Only trust Intel and Broadcom NetXtreme chips", but I hadn't seen anything from broadcom in a while, only Intel, Marvell and (ugh) Realtek, so I dropped the last part of the statement. Maybe that was premature.
Broadcom doesn't seem to advertise much to any kind of end user space, but they do new networking tech all the time. They have 200gb NICs out now and a whole host of other products. Likewise their switch ASIC division has gone from "who the fuck even uses those" to "find me a switch that doesn't." Cisco, Juniper, Arista, Dell, etc all use Broadcom ASICs in a bunch of their switches. Plus Avago bought Broadcom and then renamed to Broadcom in 2015, and bought Brocade in 2016 so they own a massive amount of the networking chip market. Oh and they suck at naming. Some of the reason you might think nothing is new is because everything gets the same basic name. You see "NetXtreme" and figure it is their old gigabit stuff. Could be, but all their new stuff from 10gb-200gb is called NetXtreme-E. So it is real easy not to notice that they have a bunch of new product lines, unless you look closer. They are the Microsoft of network hardware naming
I still like Intel better and though their products often take a bit longer to support the latest standards (they don't have 200gb stuff yet near as I know) they seem to have better feature support. Their quality RDMA support was a reason we wanted them for VM migrations (seriously, with two 10gbit connections, remote disk ops are basically as fast and low impact as local, you can't even tell). However Broadcom is seems satisfactory these days. Was a time it wasn't so, I used to have big problems if I tried any kind of VM acceleration (like queues) with their NICs, but now they seem to work well.
- Joined
- Oct 29, 2000
- Messages
- 42,445
Broadcom doesn't seem to advertise much to any kind of end user space, but they do new networking tech all the time. They have 200gb NICs out now and a whole host of other products. Likewise their switch ASIC division has gone from "who the fuck even uses those" to "find me a switch that doesn't." Cisco, Juniper, Arista, Dell, etc all use Broadcom ASICs in a bunch of their switches. Plus Avago bought Broadcom and then renamed to Broadcom in 2015, and bought Brocade in 2016 so they own a massive amount of the networking chip market. Oh and they suck at naming. Some of the reason you might think nothing is new is because everything gets the same basic name. You see "NetXtreme" and figure it is their old gigabit stuff. Could be, but all their new stuff from 10gb-200gb is called NetXtreme-E. So it is real easy not to notice that they have a bunch of new product lines, unless you look closer. They are the Microsoft of network hardware naming
I still like Intel better and though their products often take a bit longer to support the latest standards (they don't have 200gb stuff yet near as I know) they seem to have better feature support. Their quality RDMA support was a reason we wanted them for VM migrations (seriously, with two 10gbit connections, remote disk ops are basically as fast and low impact as local, you can't even tell). However Broadcom is seems satisfactory these days. Was a time it wasn't so, I used to have big problems if I tried any kind of VM acceleration (like queues) with their NICs, but now they seem to work well.
I've noticed that too. Intel stuff seems to take longer to get to market, but once it is on the market it is validated to absolute death, actually hits the advertised speeds and is rock stable to the point where it never goes down.
I'm just a hobbyist, but I dabble in enterprise gear (i just buy them years later cheap on eBay) so I have not played with the latest stuff yet. All of my servers recently got an upgrade to Intel x520 10GIG SFP+ NIC's.
D
Deleted member 89018
Guest
Historically, what we've seen when fixed function hardware is available on a platform / within components, over time the general purpose parts of the processor become able to perform what was previously handled by fixed function hardware.
You can take this pretty far back with consumer hardware that a lot of older members here would remember. math co-processors for floating point stuff in the 386 days. 486DX processors had it integrated. sound and network cards that wound up integrated into motherboard chipsets. MPEG2 decoder cards in the early days of DVDs, which eventually became the specialized video playback acceleration logic on modern GPUs even today (h264, h265, etc). Even the whole idea of a video card - you've got it integrated into all of Intel's processors and have for over a decade now (and presumably at some point AMD will have it too). Hardware transform & lighting acceleration in first and second generation GeForce cards gave way to programmable shaders in DirectX8 and the GeForce 3, and by then it was even faster than the previous generation's fixed function hardware. Physx accelerator cards came and went. I'm sure eventually we'll see it with RT hardware capabilities becoming another gpgpu function when they’re fast enough.
What has actually happened is that all of these features that started out as specialized components for accelerating particular tasks wind up having more and more uses, and eventually it all becomes another part of the general purpose capabilities. So if those features are really going to be continually useful, it's just something that the "something new" you mention needs to support, because it's expected functionality. And without it - they'll just use custom ASICs until all of those capabilities can become part of the general purpose architecture.
But, we're already doing that. Want RTX? Can't get it on your 1080, so buy a new card. Want faster RTX? Time to sell that 2070 and get a 3000 series. Want more cores? Replace that Skylake i7 with a Ryzen 5 3600. Need faster I/O? Time to replace your SATA-type m.2 SSD with an NVMe one.
You want HDR and 4k on your wall? Time to replace that 720p plasma you got in 2009.
What you're describing here is why people upgrade or just buy new tech. Motherboards with integrated SoCs or single board computers are being replaced as a single item instead of upgrading 3+ parts (like upgrading your motherboard, processor and RAM all at once because nothing from the Haswell era is compatible with a modern setup). It may look like it's a move towards more disposable computing, but it's actually less parts than it is now.
You don't HAVE to replace hardware and always run the latest and greatest, though. In fact, with stuff advancing as quickly as it is right now - especially with ARM, since thread topic - it's probably easier to just accept that whatever you buy today will be outperformed by what's available in a year (which is something that we, as PC enthusiasts, should be plenty familiar with by now).
Last edited by a moderator:
How realistic is that scenario really. We are not talking about asics designed to run data on one sensor or such things. We are talking about cores analogous to GPU cores. They are semi general compute already. No one worries that their GPUs won't be able to handle software in 4 years... oh ya wait that is how it works. Yes in 5 years you'll need the latest greatest. Its not like that isn't true for general compute. Unless your suggesting your running windows 10 on your 486... or even say your 10 year old core 2 duo. And what are you saying its not productive to mine bit coin on my 2 year old GPU... say it isn't so.
Yes things will be obsoleted... that is hardly untrue of General compute CPUs. Yes you could mine bit coin on a 486 but have fun with that. No chip escapes obsolesces no matter how many ways it can crunch numbers.
You missed is by a bit. It was a Core2Quad, a Q6600 in fact. I was using that as my main system until a bit over a year ago. Win10 and Linux dual boot. With a Radeon HD5770. Could my Radeon run ray tracing? No, but it had no problem running the games I was playing including games much newer than the card. More importantly my Q6600 was even older but it ran everything I wanted and needed. Was is slower than a new CPU? Obviously, but it could still do everything. I had no problem doing my encoding but it took quite a bit longer than my Ryzen 2600x. Or I should say my 2600x is much faster than my Q6600 but both can do the same things but at different speeds. One of those CPUs is much, much older than the other one.
There's also the issue of the efficacy of ASICs. I refuse to use a video card for encoding because I can get better results with my general purpose CPU. The encoders on the video cards while faster are almost always worse quality and/or bigger filesize. I'll stick to slower but better results. With ASICs and the like you don't have that choice and you'll be at the mercy of the ASIC's results.
Those are reasons I don't want ASICs or specific specialized hardware over a good general purpose CPU. The tradeoffs are nowhere near worth it in my opinion.
No matter what you and the RWT folks think, SPEC is a good suite to measure general performance. And the measured IPC gap between Firestorm and Zen3 falls into the expected range from the microarchitectures. Apple Firestorm is about 54% ahead of Zen3, as measured by SPEC2006int, and theoretically Firestorm would be about 57% ahead (~630 vs 256 ROB).
"Limited" is not the same like "you can't even customize the ARM systems", which was the original claim by schmide.
Moreover, I think you missed #92 and this: "From hardware standpoint, the ThunderXStation looks rather versatile: it supports various expansion slots, enabling developers to use various add-on cards, accelerators, and storage devices required by their applications."
- Joined
- Oct 29, 2000
- Messages
- 42,445
Still not my cup of tea, but that looks more like it.
Well, I'm more interested in something like a "how fast can it compile Chrome"-type test.
Oh, and triple-cost ram and storage, of course.
I didn't say I like the current status. I would like the existence of a standard desktop platform and many more devices compatible. It will take time to achieve that.
I was simply debunking the statement that "you can't even customize the ARM systems".
- Joined
- Oct 29, 2000
- Messages
- 42,445
It is fair to state it is currently very limited though.
Also, this ThunderXStstion is two years old, and while Gigabyte has a product page for it, and I have seen a couple of reviews, is it actually for sale anywhere?
Like, if I had my credit card in hand, and today wanted to place an order, is that even a possibility?
Or is it one of those specialty business orders with a PO only shipped directly from Taiwan with a 4 month lead time, type of things?
- Joined
- Oct 29, 2000
- Messages
- 42,445