AMD Zen 5 "Granite Ridge" CPUs Reportedly in Mass Production

[Chry]

To build an additional PC, I hope?
I sure hope that you're not upgrading CPUs "every" generation and then selling the old, that would be a tremendous waste of money long term.

For the same PC. The money is not wasted if it gives me psychological satisfaction. And saves me accumulative 5ms on some tasks.

 

[Vader]

Advanced noise reduction on microphones, live green screen, "magic eraser" for photos and so on. Small but sometimes useful consumer-level tricks that have been on the phones for the last couple of years but aren't yet built-in into windows. I imagine they will be with win12 later this year.

Correct me if I'm wrong, but isn't the point of the NPU to do these tasks at a higher efficiency, rather than enabling them at all?
In desktops this was always possible, most likely by using a dedicated GPU. Not to say I don't welcome these developments, but this is more impactful in mobile devices.

 

[trparky]

Correct me if I'm wrong, but isn't the point of the NPU to do these tasks at a higher efficiency, rather than enabling them at all?

It probably comes down to the fact that trying to brute force the tasks using general purpose x86 instructions would be so massively inefficient that you might as well just say screw it. It's the same reason why we hand off certain jobs to GPUs and dedicated hardware to do things like video encoding and decoding, because doing it with general purpose hardware would be too computationally expensive.

 

[las]

Finally, can't wait for official dates

 

[kapone32]

Computex 2024 should be good for Ryzen.

 

[efikkan]

Hopefully memory controllers have been improved with new Gen of CPUs and they improve memory training speed.

As you should know, memory training happens because the memory controller degrades and can no longer operate reliably at the current setting. Even if AMD improve their controller, it's still a losing battle with OC memory. Many such kits have very low SPD profiles, which is what you'll ultimately have to fall back on. OC memory is by far the main culprit when it comes to causing stability issues for custom builds.
Unless you're building a PC for the purpose of benchmarking, stick with quality memory with the highest JEDEC speed rated for your CPU, then all of these problems will go away.

Why not? I will do it myself, but probably year after the launch when I can sell my current 7600x and get 8600x/9600x with extra $50. The performance jump between 5600x and 7600x in SC and MC is >30% and the later is released 2 years later. I expect similar improvement with Zen5. What we can say about the the Alder Lake owners who upgrade from 12th to 14th gen to get less than 10% SC improvement.

Also when I resell the 8600x/9600x after few years It will have better value than 7600x, probably $25-30 so its almost free upgrade

So, you rely on selling your old CPU for $50 off when it's halfway through its warranty period? I don't know your local market, but it seems like a poor deal for the buyer to me.

For the same PC. The money is not wasted if it gives me psychological satisfaction. And saves me accumulative 5ms on some tasks.

Sure, enjoy your PC

But wouldn't skipping a generation then give you even more satisfaction when you can actually notice the difference?

I often recommend keeping the old and relegate it to a "spare" system (or give it to a family member). Having two operational PCs is very handy, both when the primary have some issues, and when a friend comes over for a round of gaming, having an "OK" secondary PC can still pack a whole lot of fun. (Not to mention, if you ever do any kind of "work", you should always have some kind of spare.)

I don't know what the "kids" do these days, but back in the days, we hooked up computers and played through LAN. Unfortunately I rarely have time for that any more…

 

[Chry]

But wouldn't skipping a generation then give you even more satisfaction when you can actually notice the difference?


I don't know what the "kids" do these days, but back in the days, we hooked up computers and played through LAN. Unfortunately I rarely have time for that any more…

Oh but I'll notice the difference even if there won't be one! And upgrading every generation will max out the excitement. Also dopaminic excitement works through expectation of getting something and if I was only planning to upgrade in the next-next generation it would be too far away to get excited about!

As for LAN gaming - unfortunately that's only for us old farts anymore. Youth are all playing through the internet.

 

[Minus Infinity]

Computex 2024 should be good for Ryzen.

That would probably be for X3D variants. All rumours are saying April announcement and probable release.

 

[usiname]

It probably comes down to the fact that trying to brute force the tasks using general purpose x86 instructions would be so massively inefficient that you might as well just say screw it. It's the same reason why we hand off certain jobs to GPUs and dedicated hardware to do things like video encoding and decoding, because doing it with general purpose hardware would be too computationally expensive.

7840u have 10 TOPS, while 8840u with dedicated NPU have 16. If 7840u is to slow to brute force the job, then the dedicated NPU is also useless

 

[trparky]

If 7840u is to slow to brute force the job, then the dedicated NPU is also useless

How do you figure that? The NPU is a separate area of the silicon. It's like the dedicated neural engine that one might find in the Apple M Series of chips.

 

[usiname]

How do you figure that? The NPU is a separate area of the silicon. It's like the dedicated neural engine that one might find in the Apple M Series of chips.

Isn't the count of TOPS the AI compute power? The extra NPU should add more TOPS so more AI performance, but that does not mean that the CPU was not capable to do AI calculations before. Especially with improvement of just 6 TOPS when extra NPU is added (10 TOPS before the NPU). Is here a problem the 10 TOPS of 7840U to be used for such AI tasks and you need dedicated NPU for this?

Actually there is also "Total Processor Performance" which is probably when the CPU cores and GPU and NPU if available are used.
7840U have 32 TOPS
8840U have 38 TOPS
It looks like the NPU does not add almost anything to the total performance, if the full cpu is used

And if you have dedicated GPU and use it for the AI calculation, you will not feel difference if your CPU have NPU or not.

 

[trparky]

Isn't the count of TOPS the AI compute power?

In a word, yes.

The extra NPU should add more TOPS so more AI performance, but that does not mean that the CPU was not capable to do AI calculations before.

OK, I simply assumed that the NPU was a special kind of silicon that does things like this infinitely more efficiently. Perhaps I fell into the marketing hype that AMD, Apple, and Intel are hyping. I don't know.

 

[TumbleGeorge]

OK, I simply assumed that the NPU was a special kind of silicon that does things like this infinitely more efficiently

Fixed for you.
Yes will increase efficiency of whole system. Plus some new(or new for me) features.

 

[efikkan]

As for LAN gaming - unfortunately that's only for us old farts anymore. Youth are all playing through the internet.

Sure, I know local hosting of games is rare these days, but isn't there value to playing with friend in the same room vs. remotely? (even if the game is hosted on a server) Isn't playing Fortnite or whatever more fun when people are physically in the same room?

Regardless, I'm at least excited about this year in terms of CPUs, as both makers are bringing a new generation (not another refresh this time), and I'm hoping to build two systems this year (HTPC and workstation). But time flies, so who knows, one of them might slip into next year.

 

[Minus Infinity]

In a word, yes.

OK, I simply assumed that the NPU was a special kind of silicon that does things like this infinitely more efficiently. Perhaps I fell into the marketing hype that AMD, Apple, and Intel are hyping. I don't know.

Here is a good benchmark I found comparing cpu to igpu to npu for AI tasks on a Core 7 Ultra 165H

• Procyon (OpenVINO) NPU: 356
• Procyon (OpenVINO) iGPU: 552
• Procyon (OpenVINO): CPU: 196

Notice that the npu is about 80% faster than the cpu, but the iGPU is much faster again.

NPU is not even really needed. If AI is important buy a powerful GPU. Of course for laptops you often don't get that choice so the iGPU will be the best for the task. NPU.


This article few day old article has a good read on NPU:

IS NPU required

 

[trparky]

Here is a good benchmark I found comparing cpu to igpu to npu for AI tasks on a Core 7 Ultra 165H

• Procyon (OpenVINO) NPU: 356
• Procyon (OpenVINO) iGPU: 552
• Procyon (OpenVINO): CPU: 196

Notice that the npu is about 80% faster than the cpu, but the iGPU is much faster again.

That's a bloodbath.

 

[TumbleGeorge]

Notice that the npu is about 80% faster than the cpu, but the iGPU is much faster again.

It is interesting in this case, what is the area of NPU/GPU logic and what is the consumption per unit of "AI" performance. It still seems to me that the NPU is at least an order of magnitude more efficient than the GPU.