Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[Rekluse]

Double Post

 

[uzzi38]

Why are you assuming everyone needs a powerful dGPU, just because they have a powerful CPU? They handle completely different types of workloads.

Well then, why don't you list some consumer workloads that require 8+ core CPUs that don't also benefit from having a dGPU and we can determine how important that market is?

Also, Intel does not agree with you either. Because otherwise they would not have made sure the ARL desktop CPUs will comply with the 40 TOPS AI-PC requirement without having to add a separate dGPU. I.e. it's an indication that they expect a lot of ARL desktop systems to be built without dGPU.

That 40TOPs numbers is reached by a combination of CPU and NPU TOPs, it's not just NPU. The NPU only makes up a small portion even. No real workloads will utilise both at the same time in that fashion, so really hitting 40TOPs in that way is really just a marketing thing more than anything else.

Also p.s. Intel doesn't ship their mobile products on desktop the way AMD does, so obviously the desktop products will be designed to be used that way too.

 

[Thibsie]

That is just so incredibly dumb for fixed hours work. They do it here as well. Awful loud machines though with annoying fans. On permanent dock. 3x as expensive. But when I request a nicer monitor or a nicer mouse, nope. IT people are dumb here.

Not surprised. Same kind here.

 

[JustViewing]

Well then, why don't you list some consumer workloads that require 8+ core CPUs that don't also benefit from having a dGPU and we can determine how important that market is?

Non gaming enterprise software development? I know many use Laptops for software development, but a powerful multi-core CPU makes a big difference in productivity.

 

[Hail The Brain Slug]

Non gaming enterprise software development? I know many use Laptops for software development, but a powerful multi-core CPU makes a big difference in productivity.

Not just enterprise. Most fields of software development besides games have little to no use for a powerful dGPU.

 

[uzzi38]

Non gaming enterprise software development? I know many use Laptops for software development, but a powerful multi-core CPU makes a big difference in productivity.

Depends on what you do even with software development. I'm a web developer myself, and outside of my specific role in working on E2E tests (where more cores comes in handy as it lets me do parallel runs at the same time), the rest of the team don't really have much use that kind of parallelism. And if they did, chances are they'd try offloading it to Azure pipelines instead.

Very few developers are doing things like compiling massive codebases on the regular, which is where that extra CPU brunt would come in handy.

But I do see where you're coming from, that is a valid usecase. My point is just that it's probably a bit more niche than just a generic "software developers" answer.

 

[Mopetar]

At the end ? Seemed to me it went on long ago.

Your post seems to imply it was over.

I don't know if we'll see the end for another 20 pages, and that's probably being optimistic.

 

[Mahboi]

Non gaming enterprise software development? I know many use Laptops for software development, but a powerful multi-core CPU makes a big difference in productivity.

The amount of applications that extensively use multithreading is quite small and specialised, aka not a market.
There's a lot more apps that do not necessarily multithread but just have easy parallelisation, I.E compilers, web servers or crawlers, zipping/unzipping large batches of files, etc. Possibly databases but that's a can of worms I'd rather not open.

The difference in productivity in 100% of these applications is not a market for desktop or small PCs. It's distant machines where you don't care for speed, or distant machines where you care for speed and then you just pay for servers.

Not just enterprise. Most fields of software development besides games have little to no use for a powerful dGPU.

I hate to be the Jensen here, but:
yet

 

[DisEnchantment]

Non gaming enterprise software development? I know many use Laptops for software development, but a powerful multi-core CPU makes a big difference in productivity.

Most companies giving out flex working arrangements or work from home arrangements issue laptops. What doesn't run locally fast enough is offloaded to cloud.
But sometimes you need to run things locally. Don't wanna be left with E cores for that.

But I do see where you're coming from, that is a valid usecase. My point is just that it's probably a bit more niche than just a generic "software developers" answer.

Enterprise and engineering field is quite diverse, there is lot of variety in needs and types of HW. Web development is one of those fields where you can make do with a Macbook Air.

Multi core performance without compromising single core performance is definitely very common in most of the engineering world.
Build your Android image, compile your VHDL code, build something in Unity, simulate your SPICE models, build Yocto image, Execute your MATLAB Simulink models etc. all these need lots of compute power which certainly can be offloaded to cloud, but very often you need to execute these workloads locally and you don't want to be left with a Macbook Air.
This field also pay top dollars for supported HW.
For each of my engineers for instance I am provisioning AWS instances with 8K+ USD MTD and laptop replaced every 36 months with IT available 24/7 to support on HW/SW issues.
I moved most of my guys to 9374F instances but on client side there is lot of complaint. And our supplier has no AMD options on client side.

 

[Mahboi]

In AMD's case most systems that will be running without a dGPU will likely be mobile APU based anyway.

Mobile APUs are generally better suited to office PC style usecases than stuff like Raphael. They idle lower, those kinds of office style PC have no real usecase for the handful of cores you get on mobile anyway and of course now they'll have NPUs. That's just the more obvious solution.

I feel like you're heavily overestimating how much of the PC market is tower PCs running high end 12+ core CPUs without dGPUs. It's a very, very, very small subset of systems.

I'd even stretch it a bit further and say that in the next 10-15 years, we will see an increase in SFF, mini and laptop PCs in consumers to the point that APUs with a minimal amount of graphics will be the common option for 90% of non gamers.

I've already had my old man change from his 3000k era CPU to a Cezanne APU in a Minisforum box, and he's had basically nothing bad to say about it. Shoved 2 4K monitors on it and is as happy as can be. I expect a lot of people will go with the mini PC + one or two 4K 75Hz monos in the next 10 years...

 

[carancho]

Well then, why don't you list some consumer workloads that require 8+ core CPUs that don't also benefit from having a dGPU and we can determine how important that market is?



That 40TOPs numbers is reached by a combination of CPU and NPU TOPs, it's not just NPU. The NPU only makes up a small portion even. No real workloads will utilise both at the same time in that fashion, so really hitting 40TOPs in that way is really just a marketing thing more than anything else.

Also p.s. Intel doesn't ship their mobile products on desktop the way AMD does, so obviously the desktop products will be designed to be used that way too.

Heavy duty Excel files that take half a minute to recalculate. Non-accelerated data science tasks. These need very fast single core, high memory bandwidth, huge memory size (datasets that need to fit in RAM calling for at least 32GB), and to a lesser extent multithreading. No use for GPU or NPU. You'd be surprised at how many corporate jobs involve uses like these, think finance. Looking forward to Strix Halo for this.

I'd love to see more benchmarks with those markets in mind, but I guess the consumers of the content are all gamers?

 

[carancho]

I'd even stretch it a bit further and say that in the next 10-15 years, we will see an increase in SFF, mini and laptop PCs in consumers to the point that APUs with a minimal amount of graphics will be the common option for 90% of non gamers.

I've already had my old man change from his 3000k era CPU to a Cezanne APU in a Minisforum box, and he's had basically nothing bad to say about it. Shoved 2 4K monitors on it and is as happy as can be. I expect a lot of people will go with the mini PC + one or two 4K 75Hz monos in the next 10 years...

I agree, the mini PCs from the like of Minisforum are very attractive. At some point the DIY DT market demand will be so low that the supply of desktop parts should dwindle. That market will remain viable for the ultra high end only.

 

[Mahboi]

I agree, the mini PCs from the like of Minisforum are very attractive. At some point the DIY DT market demand will be so low that the supply of desktop parts should dwindle. That market will remain viable for the ultra high end only.

Once bothered to run a home server type of thing on a Raspberry Pi 4.
Twas a fun thing to tinker with, but frankly, today you can shove a Lucienne/Barcelo thing for $350 and get a full SSD and 16Go DDR4 on the thing. It's 3 times the price and 15 times the performance.

Those micro boxes are the future for a lot of low power users. If only there was less of a staircase between terrible Gracemont e cores and Zen 2 cores, there's probably a market for $250 cheap home servers in 100 million homes for home cinema/automation or small businesses like pizza parlors running their own in-house website. But everyone just pays the Cloud, obviously. It's just my idiosyncrasy to love self-owned miniboxes.

 

[JustViewing]

Depends on what you do even with software development. I'm a web developer myself, and outside of my specific role in working on E2E tests (where more cores comes in handy as it lets me do parallel runs at the same time), the rest of the team don't really have much use that kind of parallelism. And if they did, chances are they'd try offloading it to Azure pipelines instead.

Very few developers are doing things like compiling massive codebases on the regular, which is where that extra CPU brunt would come in handy.

But I do see where you're coming from, that is a valid usecase. My point is just that it's probably a bit more niche than just a generic "software developers" answer.

I also do software development as profession as well as hobby, you can certainly work with 4 cores but is it a torture especially if you are used to 8+ cores. You can run few services remotely to ease the local load, but running all of them locally improves productivity (otherwise when you fix a bug, push the code, wait for build server to complete etc..). There are nearly hundred back-end services, database servers, Web server, Multiple Opened IDEs, in these types of scenarios having multi-core CPU helps a lot. At least for me.

 

[StefanR5R]

I also do software development as profession as well as hobby, you can certainly work with 4 cores but is it a torture especially if you are used to 8+ cores.

When corporate IT writes a call for bids for workplace PCs for staff like yourself, would they put both
– more than 8 cores
and
– Microsoft's new sticker
into the requirements list?

 

[JustViewing]

When corporate IT writes a call for bids for workplace PCs for staff like yourself, would they put both
– more than 8 cores
and
– Microsoft's new sticker
into the requirements list?

Unfortunately 8 core is the maximum for now. But my work laptop is only 4 cores 8 Thread Intel and it is pain to use compared with my PC 5950X. No idea about the sticker.

 

[Joe NYC]

That 40TOPs numbers is reached by a combination of CPU and NPU TOPs, it's not just NPU. The NPU only makes up a small portion even. No real workloads will utilise both at the same time in that fashion, so really hitting 40TOPs in that way is really just a marketing thing more than anything else.

That's something I was wondering about: What TOPs number could be achieved with Zen 5 CPU alone? Let's say 8 core and 16 core.

With quite capable AVX512, it should be making a dent.

 

[Doug S]

NPU are nothing but MADD engines with operation/ memory /io aligned to GEMM. You could probably run AI on a ati 9700 pro , very small model with terrible performance but you could probably run it.

Given the rather large overlap between what GPUs need to do and what NPUs need to do, I suspect they will be combined. That's already effectively true for Nvidia's high end discrete stuff, but I would expect say Apple will combine them before long. They haven't had reason to up until now because their current NPU is so small it wouldn't have been worth complicating their GPU. Now that they are rumored to be expanding the size of their NPU, they will have reason to look at combining them to conserve die space - you get more peak NPU and more peak GPU when a block can do both. But it is more complicated as you need to be able to flexibly partition that combined unit so part of it can do AI and part of it can do graphics simultaneously.

Intel & AMD will have the same incentives, both are new to the NPU game so they'll be separate at first but I'm sure they both see the logic in combining it with the GPU. But that will take longer to do right so it makes sense to keep the NPU separate to more quickly get their first designs out the door and jump on the "AI PC" hype train Microsoft and PC OEMs want to ride.

 

[inquiss]

Heavy duty Excel files that take half a minute to recalculate. Non-accelerated data science tasks. These need very fast single core, high memory bandwidth, huge memory size (datasets that need to fit in RAM calling for at least 32GB), and to a lesser extent multithreading. No use for GPU or NPU. You'd be surprised at how many corporate jobs involve uses like these, think finance. Looking forward to Strix Halo for this.

I'd love to see more benchmarks with those markets in mind, but I guess the consumers of the content are all gamers?

I don't disagree with your use cases being valid requirements for single thread performance. That said, these use cases are a niche in a niche of corporate needs. Most stuff is word processing accessing the web and some slides. Even within a finance institution 99% of what's done can be done on a potato.

 

[dr1337]

The amount of applications that extensively use multithreading is quite small and specialised, aka not a market.

The fact anyone would say this after zen 1 with its worse single core performance managed to claw away decent market share from intel, is beyond me.

If you do anything other than use chrome and play fortnite, you will eventually find yourself in a MT bottleneck.

I don't disagree with your use cases being valid requirements for single thread performance. That said, these use cases are a niche in a niche of corporate needs. Most stuff is word processing accessing the web and some slides. Even within a finance institution 99% of what's done can be done on a potato.

I have worked inside multiple financial institutions and I never ever saw anyone working on a potato machine. In fact everyone that actually had a desk had tower desktops so they could drive > 3 displays. The sheer amount of compute used in these markets really is more than most people would assume at face value. And I should say that this doesn't implicitly speak about CPU performance no, but there is 0% chance these companies buy anything less than the best for their employees.

Like really I don't think anyone in the world expects less than 16c from AMDs next flagship.

Anyone that thinks MT doesn't matter is either very ignorant or just pretending, perhaps even terminally online. Even outside of finance there are tons of uses for MT performance and I find myself (slightly) bottlenecked every day. But until costs makes sense for client/edge uses, its not worth the investment most of the time for peak MT perf. Mistaking prices/barrier to entry for market demand is a fools errand IMO. The market always wants more performance, but it never works if the progress:cost ratios go up. There is a reason after all that OEMs aren't shipping tons of threadripper or xeon, as much as you or I would love to play with one.


And also, in this thread, people are talking about 40% gains in FP performance with zen 5 like it is a big deal. That kinda stuff is even more niche than overall MT or ST perf, yet it seems to be very important for a lot of people. I think consumer workloads (from all sales angles) and their impact on relative marketshares shouldn't be underestimated.

 

[HurleyBird]

The fact anyone would say this after zen 1 with its worse single core performance managed to claw away decent market share from intel, is beyond me.

Anyone that thinks MT doesn't matter is either very ignorant or just pretending, perhaps even terminally online.

I generally agree, but I also think eventually you reach a point where a sufficient deficit in ST just makes any advantage in MT irrelevant.

If a product gets 0.9x ST/1.2x MT, that's viable.

But if a product gets 0.66x ST/1.66 MT, that's a really touch sell tbh.

I'd take 8 Zen1 cores over 4 Skylake cores any day of the week. But I'd also take 6 Zen4 cores over 32 Zen1 cores just the same.

If Intel is relatively close in ST, say up to around 15%, then I agree with you. If Adroc and Kepler have it right and AMD has something like a 25%+ ST advantage, then I agree with them.

 

[inquiss]

The fact anyone would say this after zen 1 with its worse single core performance managed to claw away decent market share from intel, is beyond me.

If you do anything other than use chrome and play fortnite, you will eventually find yourself in a MT bottleneck.

I have worked inside multiple financial institutions and I never ever saw anyone working on a potato machine. In fact everyone that actually had a desk had tower desktops so they could drive > 3 displays. The sheer amount of compute used in these markets really is more than most people would assume at face value. And I should say that this doesn't implicitly speak about CPU performance no, but there is 0% chance these companies buy anything less than the best for their employees.

Like really I don't think anyone in the world expects less than 16c from AMDs next flagship.

Anyone that thinks MT doesn't matter is either very ignorant or just pretending, perhaps even terminally online. Even outside of finance there are tons of uses for MT performance and I find myself (slightly) bottlenecked every day. But until costs makes sense for client/edge uses, its not worth the investment most of the time for peak MT perf. Mistaking prices/barrier to entry for market demand is a fools errand IMO. The market always wants more performance, but it never works if the progress:cost ratios go up. There is a reason after all that OEMs aren't shipping tons of threadripper or xeon, as much as you or I would love to play with one.


And also, in this thread, people are talking about 40% gains in FP performance with zen 5 like it is a big deal. That kinda stuff is even more niche than overall MT or ST perf, yet it seems to be very important for a lot of people. I think consumer workloads (from all sales angles) and their impact on relative marketshares shouldn't be underestimated.

I work in the finance industry too. The data science teams have pretty chunky builds of tower PCs and some even chunkier on site servers that they use to log into for bigger modelling. Anything needing vast resources is often done in the largest off site infrastructure. The headcount of the people, even within a finance institution that are doing this modelling is tiny compared to the number of support staff, marketers, sales people, product people, legal, and front line staff.

So I'm not saying that the people you're pointing out don't need big machines, they do. I am saying though that even though you and I swim in this world, if you look at the headcount in those companies it's a small pool that actually needs beastly machines. So, globally, finance is a niche and the people within finance headcount that need these machine is also a niche.

If they need bigger and better systems they buy thread ripper or servers. If you have a need for top performance you don't buy the client platform. Core basic business machines don't need much grunt because most people don't need it.

The next flagship client is 16 cores. No more.

 

[Glo.]

I am working in finance as well and I can tell you guys - You DO need powerful machine to handle everything. Multiple monitors: 3 or more, large amount of RAM, 64 GB is the bare minimum, at least 12 CPU cores, best - 16.

Mac Studio is good enough for people like me, but for people who are running multibillion funds - you need the horsepower for modelling.

 

[inquiss]

I am working in finance as well and I can tell you guys - You DO need powerful machine to handle everything. Multiple monitors: 3 or more, large amount of RAM, 64 GB is the bare minimum, at least 12 CPU cores, best - 16.

Mac Studio is good enough for people like me, but for people who are running multibillion funds - you need the horsepower for modelling.

Sure, you find yourself within the set of people that need powerful computers but who are not running anything big enough where the really big on or off prem devices will be needed. I'm not saying people like you don't exist, they clearly do and you're one of them.

That said, if you look at a global bank and their IT needs, how many of the people within that organisation will actually need that grunt. Your department might, but what proportion of the headcount does. A global bank might have 100k employees, how many are modelling vs doing fairly basic tasks. How many of those with big data needs are going to process that on their PC Vs having that done on specialist hardware.

If you want more multi threaded performance you probably have the cash for a real multi threaded machine. Most corporate laptops just need to run the internet and office products at a decent chop.

 

[Glo.]

Sure, you find yourself within the set of people that need powerful computers but who are not running anything big enough where the really big on or off prem devices will be needed. I'm not saying people like you don't exist, they clearly do and you're one of them.

That said, if you look at a global bank and their IT needs, how many of the people within that organisation will actually need that grunt. Your department might, but what proportion of the headcount does. A global bank might have 100k employees, how many are modelling vs doing fairly basic tasks. How many of those with big data needs are going to process that on their PC Vs having that done on specialist hardware.

If you want more multi threaded performance you probably have the cash for a real multi threaded machine. Most corporate laptops just need to run the internet and office products at a decent chop.

I actually was confirming what you are saying. IMO, we are not bound by comptue anymore, the RAM, memory and latency is the bigger bottleneck.

Most people in financial markets are like me, lone wolfs. There are however in financial world people that do need horsepower, hence Threadrippers, stupd amounts of RAM, and incredibly large display setups(6x 42-43 inch monitors/TVs).

The needs for the incredibly powerful setups is simple. Beaurocracy. You need proof, statistical, mathematical, empirical proof for thesis to pull the multibillion dollar trigger over investment, or cutting moving the capital to new ventures.

And that will not be done on said Mac Studio. For the rest of people - you are correct. Only the people who are actually in the Financial Modelling, FA, etc that do need the horsepower. Everybody else can make do with MacBook Air.

Me? I need Mac Studio, 64 GB of RAM, and at least 3 displays .