What I have been saying about ARM servers for years. Once a company makes the leap, they are now in a situation where they might consider a different ARM hardware vendor. Going back to a x86 platform is a huge hurdle. I know some people write off the market penetration of ARM in datacetners and servers. I think ARM is over 25% market share in datacenter now. They are shrinking the customer base potential for the x86 companies.
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
-
Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.
NVIDIA "Vera" CPU Benchmarked: Beating Intel Xeon and AMD EPYC
- Thread starter erek
- Start date
If anyone can push the transition it is Nvidia. They are heavily investing into partners to push it. The problem is it won't be cheap.
- Joined
- Jul 13, 2006
- Messages
- 2,192
I would LOVE to have a highly performant arm cpu as my main linux pc.
Most stuff runs just fine on Windows Arm. As for motherboards, I would kind of expect them to come with soldered CPUs and the equivalent of custom UEFI at first, because every vendor does their own thing outside the CPU cores, so standardization would be sold until and unless it's hashed out.
There is arm device tree and other standard outthere now:
https://www.arm.com/architecture/system-architectures/systemready-compliance-program
That why standard Linux for arm work for the new Nvidia Vera cpu system, Nvidia usually use UEFI like x86 cpus:
https://docs.nvidia.com/dgx/dgx-spark/uefi-settings.html
Ampere and Nvidia are both fully compliant to be systemready I think, so you can install the same vanilla ISO of an ubuntu or others and simply install it on them or windows for arm.
Many arm motherboard have regular socket for ram and cpu, for example ASrock Ampere altra micro-atx server board:
Nvidia grace cpu in the DGX station will be soldered too, the amount of CPU-GPU bandwith they have with NV-link could be hard to achieve with sockets, quite the direct and super fast-wide connexion and it is not like cpu options exist for that socket to ever change it for something else, but standard uefi and any armlinux-windows should work, they probably have some special version for the drivers and handling the unified memory system and NVLInk stuff the best too, at least they work on the spark with a very similar SOC.
https://www.phoronix.com/news/NVIDIA-RTX-Spark
NVIDIA's Computex keynote around the RTX Spark was heavily promoting RTX Spark with Microsoft Windows 11. But, fear not, the NVIDIA RTX Spark will still work great on Linux.
DukenukemX
[H]F Junkie
- Joined
- Jan 30, 2005
- Messages
- 11,195
Every CPU has been extended. CPU's age if they don't.
One can say the same thing about ARM. ARM isn't exactly new either. The 3DO has an ARM CPU in it.
AVX was mainly created for servers. It's one of the reasons why AVX-512 was left in limbo as Linus Torvalds had not good things to say about it. Intel wanted it for servers while desktop users had no use for it. Now we're finding uses for it, and it's coming back. SVE is also hardly implemented in ARM chips, but x86's AVX-512 has a good head start over SVE2.
x86's future depends on what AMD and Intel will do. I doubt x86 is going anywhere in 10 years, let alone 20 years. As long as AMD and Intel keep advancing x86 then there's no reason to believe ARM will take over the desktop market. As much as people want to believe Panther Lake was a massive leap forward, it really wasn't. Panther Lake is very much behind AMD's Zen5, let alone Apple's M4's and M5's. I don't expect Nova lake to be all that different. The real x86 leader is likely to be AMD's Zen6, which is said to be using TSMC 2nm. That is going to be a massive game changer for x86. We might see AMD beat Apple to using TSMC's latest manufacturing. It's stupid that AMD has been using TSMC 4nm for such a long time.
I'm not suggesting ARM is perfect. Just that it is 100% a viable solution, and as capable as x86 today. Frankly more capable in the future. ARM doesn't have just two companies and a Chinese knockoff company working on it. ARM does require a license... but seeing as x86 isn't being licensed to anyone new. x86 days are numbered. ARM may be imperfect, but any new player willing to spend the money required to get in the game can. Intel allowed ARM to happen, they could have killed it in the crib anytime they wanted. They instead tried to protect their fiefdom. Now 25% of datacenters are no longer using x86... and as we have talked about in this thread. The chances those outfits EVER go back to x86 are pretty much zero. The market has gotten smaller for x86 and that is mostly on Intel. Its hard to blame AMD, mostly as they have been irreverent to the conversation as well. They got their house in order and started making inroads in server parts, and within a few years Intels choices are going to bite them as well.
All of the new server ARM chips support SVE2... 128 bit hardware, software works 128-2048 bit. Performance is much the same as AVX... with the advantage of being easier to work with on the software side, and being future proof in a way AVX just isn't. I mean software work needs to be done regardless. But yes for inference work which is what Nvidia is pushing SVE for... it works quite well. (and of course being NV they are sure to mention their SVE2 implementations supports FP8 lol AI AI AI) The 128 bit hardware sounds like a limitation but the way it has been implemented on those chips it doesn't seem to be. It is every bit as effective as AVX-512.
The desktop market is irrelevant.
Frankly in a decade there very well may not be a desktop market anymore.
If you haven't noticed every move for the past 5-8 years or so has been to end the entire idea of desktop computing.
x86 imo is doomed. AMD also won't go down with the ship. Its not exactly a well kept secret that they have both lightweight and NON lightweight versions of soundwave developed.
I'm sure the next gen that is already well tapped and in production is safe. But retail PC sales are going to continue shrinking. Unless the AI thing completely explodes not just deflate a little, I don't see much changing either. Desktop compute will continue to decline to a point where it just isn't important. We have a future of cheap APUs, over priced APUs... and probably not much else consumer grade to look forward to. Not to be a Debbie downer.
Last edited:
Maybe but now all modern arm phone cpu (mandatory for those) and AWS Graviton, google server cpu, AmpereOne, Snapdragon X2 and Nvidia N1X has both (SVE and SVE2), Apple going for SME-SME2 (or arm going into Apple AMX and calling it SME...) only do throw a bit of a wrench here but they have such a big and wide NEON units they can do arm to arm translation into those, which probably defeat the efficacy-performance gain wanted anyway.
But those devices tend to already media engines, npu and a gpu on the soc doing what people would do with SVE2... so it could stay a niche cpu usage, specially with Apple large influence. Like wide AVX has been on x86 side of things for personal computers.
Except at least 3 moves industry made or was pushed into that can promote large stationary box with big PSU and cooling capacity, discrete GPU power usage and size has been going up not down, the explosion of popularity of PC gaming worldwide (with the * that chinese market is almost all of that growth and gaming laptop heavy apparently) and recently local AI network machines and how much power you can want in them and for those you have no need for it to move-or for it to have a monitor, making mac-mini type box and the upcoming bigger one from Nvidia, not sure Nvidia would have launch big desktop workstation in 2026 without that last one.
That said, desktop cpu more and more are a mix of made for laptop first tech cpu (next Intel desktop CPU will have an lower power islands core for an extreme example of that) boosted up or past down server core (zen5 cores did seem to be that a lot), that is true.
Last edited:
And what if AMD and Intel decide that licensing the instruction set is in their best long-term interest? And if that new licensee is innovative enough to propose new instructions?
On what basis do you say this? IF X86 processors were improved enough, no doubt some datacenters would choose x86 in their upgrade/refresh cycle.
How can you be so confident? Will we all be using server CPUs in our desktop systems?
Who is behind these moves? People with the best interest of desktop users? Or greedy SOBs who want to squeeze more $$$ out of us every month.
Perhaps you are shillling for ARM?Not to be a Debbie downer.
Which implies that there is a low switching cost for datacenters from ARM-based processors back to x86, really x64
- Joined
- Aug 18, 2006
- Messages
- 11,470
Y'all are cute. You think that this will actually be sold to the public with Jensen out there like "you'll get to RENT your own compute power"?
- Joined
- Jul 13, 2006
- Messages
- 2,192
Well they're still targeting edge computing. I do use my jetson as a desktop for somethings
My bet is it's all solder in junk if x86 dies
Intel has long ago shut the door on licensing the ISA. At this point as they have allowed the competition to grow, they would have to agree to terms they will simply NEVER agree to. Having 2 players in any license talk is a huge issue as well. Getting both Intel and AMD to agree on what markets they are going to allow others to compete with them directly? Ya good luck. They would both want compensation, and both would want to protect their core markets. Why would say Samsung license x86 and pay 40-50% in royalties to satisfy both players... and or agree to a whole bunch of fine print. They can just deal with ARM. Or go open source with RiscV, or heck even Power which IBM open sourced in 2017.
I say this because software is a larger expense for most companies then hardware is. The biggest reasons companies didn't jump on ARM years ago, when x86 server parts were sucking even more power. Was simple their software ran on x86. In the last decade the tools to create ARM software have greatly expanded. Still most companies with larger server installs that swapped to ARM spent a LOT of money converting their software to run on ARM. The tools have made it easier but its still a lot of work even with the best tools. Maybe AI makes it easier to swap back but it would still be a mine field of issues, and costs.
When a company swaps from x86->ARM hardware is only half the cost. The other half is switching/adapting/converting software stacks to ARM. Once a company has spent 10s of millions of dollars converting all their software to ARM. That means they have made the decision to LEAVE x86. To go back to x86 would mean reversing all of that and likely incurring all those costs again. Never mind that that added expense would have to come with an admission that a mistake (a costly one) was made.
It isn't IMPOSSIBLE to swap back, it is just very difficult.
Imagine yourself a Intel sales exec. You go into an office where a company 2 years ago replaced 100million in Intel servers with AMD servers. Your swap back pitch can 100% be based on price alone, or availability of new tech. Whatever the company can really just flip your hardware on and turn off their AMD hardware. All the same compiled software on this fictional data centers storage would just be hot swapped from one system to the other.
Now Imagine your that same Intel Sales exec. You go into an office where the company 2 years ago replaced 100million in Intel servers with ARM servers. You can try to under cut ARM hardware costs... you can try. You can try to pitch the latest greatest. BUT your customer is now going to have to factor in an crazy cost/time cost of recompiling everything back to run on x86. Also your $ pitch is going to be laughed out of the room... cause lets face it 2 years ago you already lost on price. Your also a Intel sales rep, you can't really sell under cost. As someone who worked in sales for a lot of years. I gotta say I'm not even taking that meeting. To close that sale I would have to give away all the profit and probably make ZERO for my efforts. lol
We are already there. How many people use Google docs? Microsoft also wants you to use online office products. Adobe isn't running photoshop AI on your desktop, they are running it on their servers. You don't need a crazy high end desktop to use photoshop AI.... it runs the same on a potato. You can open those google doc files on your phone.
Arguing desktop computers are not dying. In a world where EVERYONE you know is doing 95% of their computing on their phone, means your just not paying attention.
The average person doesn't even own a desktop anymore. Laptop ownership is also down. For many people a phone is all they need. All heavy compute stuff they might do is done yes on a server somewhere. The low power ARM chip in their hand is good enough. In the laptop market... Apple isn't even pretending anymore they are now selling the Neo a Phone chip powered laptop. And you know what ya its good enough for 90% of computer users.
As to who is behind the moves. Yes the entire industry is. They all would prefer you own nothing. Pay for subscriptions. AI for a company like Adobe was the ultimate tie in. They moved to subscriptions but the ultimate is relying on them to do the heavy processing externally. Gaming industry has been trying as well... they keep trying different payment models that result in you owning nothing and paying and paying and paying. There has been multiple game streaming plays. They aren't going to give up on that idea either. In fact game streaming is going to get very attractive in the next few years. Already is really. Expect the average gamer to start just going with it over the next couple years. I mean if a decent mid range gaming PC starts costing 3k and bragging rights territory doesn't kick in till you have spent 5k. The streaming services start looking more and more like a very viable alternative.
Am I shilling for ARM. No I'm just being truthful. CPU compute is too big a business for it to be controlled by TWO companies. Of course Apple dumped Intel. No company likes to be beholden to a single source for anything let alone the core (pun intended) of your product. I remember before they switch and the rumors Apple was talking to AMD. I am sure Apple realized if they engaged AMD for some products Intel would retaliate with pricing and maybe even availability. They had no choice really. ARM isn't the perfect option for the industry either, its just a less shit one compared to x86. Apple with their forever low cost ARM license is set. ARM terms for other players are not as ideal. It isn't a bed of roses, still other companies have no choice. I mean does Nvidia have any choice to go x86? Of course they don't... and Jensen DID try to get a license. Don't forget their was a time when Jensen made x86 chipsets. He was angling hard for a CPU license, Intel told him to fuck off and he hasn't forgotten. I would love to one day see the no doubt very smug messages he must have sent to Intel the last few years as he bought up chunks of Intel and got Intel to agree to making Intel+NV graphics parts.
I mean that is another little caveat for the future of x86 now that Nvidia owns 5% or so of Intel. Essentially their largest rival now owns a say in their future plans.
Last edited:
- Joined
- Jul 13, 2006
- Messages
- 2,192
One thing is for sure. An open ISA fight is exactly the competition we need. The intel vs amd fight is incremental jumps. And arm ISA with Nvidia et al going to x86 market share will force innovation on all sides, and im for it.
Last edited:
They will move it more and more on customer machine to save money, the version they made for Nvida machine will use local power quite a bit would be my guess, the era of cheap subsidized AI is almost over.
https://www.msn.com/en-us/news/tech...napshotKey=GM04E3337C-snapshot-12&uxmode=ruby
Premiere takes advantage of RTX Spark's unified memory, Blackwell GPU, and TensorRT, resulting in faster rendering. Photoshop has been "reimaged with GPU-accelerated compositing at its core," which helps with the performance of live filters, HDR, and other effects. AI features get a boost from TensorRT.
There is a bit of chicken and egg and history rewritting that can occur and how much it is pushed-forced by consummer. Take the music and movies industry, their peak in revenues was the 90s CD and the 00s dvd era, by far, both tried and obstined themselve trying to keep the sell you a CD or a DVD for the same price of many months of Netflix/spotify and would have love for things to never change, they were forced into subscriptions models by consumer and tech change.
Take this for an example, yes steam giant success (with is you own nothing business model)... was it soemthing the gaming industry was trying or something that consumer actively did choose versus going to bsetbuy to buy and own a game....and forced everyone into that model.
Maybe, but the ISA would have nothing to do with it, Strix Halo being soldered is for ram need with the large memory bus it has, x86 playstation has well, many arm are socket, many arm are soldered, many x86 are socket, many x86 are soldered, the soldered one tend to be for the same reason in both ISA, ISA is virtually irrelevant to any of this (at least I do not see any link), arm is maybe cheaper to make that way than a x86 for licensing reason ? but not really technical one.
Last edited:
DukenukemX
[H]F Junkie
- Joined
- Jan 30, 2005
- Messages
- 11,195
It was never two companies. AMD and Intel are just the most successful desktop CPU's. Every other CPU company that tried to fight AMD and Intel would eventually lose because they never felt like putting in the same amount of commitment as AMD and Intel do. ARM's been around as long as x86, but ARM was never a good choice until relatively recently. Windows RT has been around since 2012 which runs exclusively on ARM. To give you an idea, it took ARM a while to go 64-bit.
Apple had the choice to go with AMD and went against this. Nobody forced Apple to stick with Intel.
Apple's success with ARM is going to depend on Windows success with ARM. As long as Windows continues to be dominated by x86, then ARM on MacOS has an uncertain future. Most programs will still be made for Windows first.
I wonder if LPCAMM could be coupled (somehow) with physically smaller sockets (or maybe the memory modules go on the back of the motherboard) to get rid of the need for soldered-on ram?
need here will tend to be how much more people are ready to pay for sockets, Nvidia can do quite a lot with sockets on their AI system with LPCAMM type of connexion, people are ready to pay a lot for the ability to change one of those gpu if they fail to keep the 8 GPU rack going, the more complex routing created by socket can end up make motherboard that cost so much more than may has well put all the best ram the socket would support soldered instead for that price on it and gain the advantage at the same time as saving price.
Going 10-12 layer board down to 8 save a lot of money, not having socket save money, a lot of test-quality control made right away save money and so on. Maybe they will, but non-soldered ram on GPU never caught-on and if we start having soc that rival them with high bandwidth, could be the same that will happen, memory could end up on package like Apple do anyway....
An open ISA for x86 would accomplish a lot. Then the Samsungs of the world can support that market, innovate, etc.
DukenukemX
[H]F Junkie
- Joined
- Jan 30, 2005
- Messages
- 11,195
CPU architecture doesn't really matter in this day in age, besides compatibility.
ARM is in a funny situation because yes anyone can license ARM, but not everyone will advance ARM's design. Not only that but ARM is now making their own chips, which can't sit will with everyone who licenses ARM. As far as I can tell, only Apple and Qualcomm have engineered their own ARM design, and at the detriment that ARM themselves could eventually improve and license a design that will outperform them. Nvidia's Vera is just an off the shelf ARMv9 with a boat load of bandwidth. As far as I'm concerned, Nvidia's Vera is just another Fujitsu A64FX processor.
That could easily change based on what AMD's and Intel's plans are for the future. If they don't license out x86 then they could start making small cell phone like chips that could be as power efficient as ARM. I have a feeling that AMD's Zen6 could finally be an x86 design that matches Apple's ARM CPU's in power efficiency. Notice I said Apple and not Qualcomm and Nvidia? There's a very good reason why Nvidia didn't allow Phoronix to have access to power consumption numbers, and I doubt it's because Nvidia didn't have power measuring ready. I really doubt Nvidia's Vera is power efficient, because they would have allowed Phoronix to test for it.
The same tools that allowed them to switch to ARM will also make it easy to switch back. I would think they'd want to switch back to Intel specifically since they're the gold standard when it comes to things like networking, storage, and etc.
AMD is just as guilty as Intel. AMD fucked up with Bulldozer and it still hurts them to this day.
Server stuff is whatever, but consumer stuff is just now getting it with Snapdragon X2's. Apple still hasn't adopted SVE.
Doubt it.
Past 5-8 years we've been dealing with Crypto and now AI. Eventually this insanity will stop and the market will correct itself.
If x86 is doomed because of a dead desktop market, then so is ARM.
Intel won't either. Remember that Intel tried to leave x86 with Itanium. You'd see x86 heavily retooled before AMD and Intel jump ship to ARM. ARM's done it, which is why ARM is improving fairly quickly.
If we're going to be a Debbie downer, then I'd bet on the AI data center market going to end up declining and not the retail PC market. You under estimate people's abilities to hold onto computers for years. You'd likely see people owning older and slower computers. You know, like I'm doing.I'm sure the next gen that is already well tapped and in production is safe. But retail PC sales are going to continue shrinking. Unless the AI thing completely explodes not just deflate a little, I don't see much changing either. Desktop compute will continue to decline to a point where it just isn't important. We have a future of cheap APUs, over priced APUs... and probably not much else consumer grade to look forward to. Not to be a Debbie downer.
They were not power tuned yet being the reason given by phoronix/nvidia... tdp number, specially with lpddr5x/all cores on the same chiplet do point for to be a really power efficient platform, like grace was (https://www.phoronix.com/review/nvidia-grace-epyc-turin/2).
Grace was using more standard core off the shell with a boat load of bandwith, vera is a bit custom.
They have a license to do it and did do some custom stuff, a big one is to have SMT (fancy spatial kind.. that kind do quite a bit) to have 2 thread per core, it is a custom x925 not just the off the shelf design:
An other one, being nVidia in the AI moment, native FP8 support for matrix and others (first arm cpu with it apparently).
SVE2 support is bigger than usual when those core are in phones, with 6x128 bits I think, the fabric it sits on and what it mean for the L3 is a significant one, there is a reason they made version of GCC/LLVM-clang for it.
must be a bit longer than that, big cortex-a710 had those, non-apple phone tend to had those for a while, any armv9 core since 2022 will tend to have it, phone will tend to be obligatory.
Switch2/Apple being on Neon is 2 big one for sure, neon stuff should run without problem on chips with SVE support at least
Last edited: