AMD: Zen 4, Speculation, Rumours and Discussion

Analysis of the AVX-512 implementation.

tl;dr:
It's overall very good.

Peak FMA throughput, and peak memory throughput are half of Intel's full-fat AVX-512 implementations, but despite this it will on many loads beat them. Latencies are low, and especially shuffles and math on masks are substantially better than on Intel, and unless you are doing BLAS or something, it's possible for these "ancillary" things to overcome the deficit of raw computational throughput. It really helps that the core doesn't downclock to do AVX-512 -- in fact clocks are higher for 512 than they are for AVX256 (because frontend has half the work to keep backend occupied).

As a negative side, VGATHER* are still comically slow, and unlike Intel the core is register port limited if you are doing FMA with masked result merging.
 
Here in germany the 5600X costs 180€. That will be nearly 200€ less than the 7600X. Zen3 prices crushed hard after Alder Lake.
When was the last time new gen launch matched current gens prices at the launch of new gen?
Exactly, never. You compare launch MSRP vs launch MSRP, of course.
 
When was the last time new gen launch matched current gens prices at the launch of new gen?
Exactly, never. You compare launch MSRP vs launch MSRP, of course.
Alder Lake didnt cost much more and delivered 50%+ perf/w improvements over rocket lake.
 
Alder Lake didnt cost much more and delivered 50%+ perf/w improvements over rocket lake.
At least in Finland Core i9-11900K had dropped from 590€ at launch to around 435€ already long before i9-12900K launched at about 660€.
12900K official MSRP at launch was also $50 higher than i9-11900K at launch.

Edit: at least in Cinebench nT i9-11900K actually beats i9-12900K silly in power efficiency (around 70%). On single core (superpi) 12900K is better than 11900K, but the difference is just bit over 6%
 
Last edited:
Just seen a video on YouTube of direct die cooling and it was 20-25c drop in temps.

Maybe it's time we went back to direct die cooling like we did back in the old days.
 
Never had a problem with chipping the corners with my thunderbird, and I feel like the copper shims addressed that well enough. I'd think having an integrated shim/spacer surrounding the die like you see on many GPUs would work just as well to protect cracking dies. If there really is that large of a temp delta I find it hard to imagine that Intel and/or AMD won't do this at some point. They're both bending over backwards to eek out the clocks at the ragged edge, and that's an awful lot of performance to be leaving on the table.
 
I don't understand why a soldered copper IHS adds so much inefficiency.

When I replaced the paste inside my Core i5 7600K and 8600K with liquid metal I saw a drop like that.
 
Thick IHS, small die area, very high clocks? 5.4ghz all-core is a lot. The IHS and solder TIM are still thermal insulators relative to direct die. The 20c drop in temp also shaved off 15watts to the cpu package power, so there's also that.
 
GPU's have been direct die for years now with problems and I don't see many, if any people on forums posting about how they've cracked the die replacing the heat sink.

So I think at some point we might see it become a standard practice for CPU's again, I used to run a 3770k direct die using EK's precision mounting kit and it was easy to use.
 
So launch platform price is bad.
Not any surprise since this is new socket + new chipset + new RAM gen + new PCIE gen all in one.

Give it 6 months or so for early adopter premiums to go, B650 to come out (probably should anyway for likely teething problems to get resolved) and low-mid pricing should be looking much more healthy.
One review basically said '7600X isn't a cheap CPU so you should match it with a premium mobo (x670) & RAM... omg its super expensive' :unsure:

Haven't got to any bits with that 65W TDP stuff yet, possibly AMD should be giving it an easy mode switch & some Eco-marketing name for those of us put off by the prospect of a constantly hot, high-power using chip.

I will be interested to see how the boost clocks work out on 7600x with more normal/budget friendly cooling than these reviewers all seem to have.
My cheap Coolermaster Hyper 212 is entirely unstressed by my 5600x, supposedly is good for 150W so I have assumed it should cope well enough with a 105W TDP chip like 5800X or 7600x.
But I guess if I'm aiming for a theoretical 7800 3DX & maxing out boost I'll have to factor in a more serious cooler.
 
An in depth look at Zen 4's AVX-512:


TL: DR: it's pretty good :)
 
Last edited:
On the whole, it is amazing how much disabling PBO and CPB reduces the CPU temperature. On heavy sustained loads like Cinebench multi core, we went from 95C to just 65C - that is a 30C drop in CPU temperature!

Even relatively lighter loads like Photoshop went from temperatures ranging from 60-80C with PBO/CPB enabled to just ~42C with them off. Others like Premiere Pro, After Effects, and DaVinci Resolve were frequently hitting 95C with PBO/CPB, and dropped to 40-62C with then disabled. As a reminder, these are workloads that saw no difference in performance, yet the 7950X CPU is as much as 30C, or even 40C, cooler with PBO/CPB disabled.
...
In most of the workloads we tested, it is very clear that using Precision Boost Overdrive (PBO) and Core Performance Boost (CPB) is not worth it. Photoshop, Lightroom Classic, Premiere Pro, After Effects, and DaVinci Resolve all showed no difference in performance when we disabled these settings from their "Auto" default, yet the CPU temperature dropped as much as 30C, or even 40C!
...
It was only the extremely heavy workloads where the CPU is loaded to 100% for sustained periods of time that PBO/CPB gave any amount of performance increase. Even then, however, the performance gain was only about 10% at most, and in exchange, the CPU was often pegged at 95C, versus just 65C with PBO/CPB off. Running the CPU for sustained loads at 95C doesn't seem like a terribly great return for a meager 10% performance boost in our opinion.
 

I'm looking forward to the X3D version of Zen 4. If 7950X does not need to be 95C to have good performance, there's a good chance that a nice sized cache would be a better use of this headroom.
 
The X3D versions will almost certainly be lower tdp just like the current X3D variants. Sandwiched dies impedes heat transfer, resulting in lower tdp.

I would much prefer AMD mimicked Apple and went for maximum efficiency instead of following Intel into the space heater market. Glad Eco mode exists.

Cheers
 
When I replaced the paste inside my Core i5 7600K and 8600K with liquid metal I saw a drop like that.
That's because thermal paste has a Thermal Conductivity of about 4 to 14 W/(m·K) while liquid metal has a Thermal Conductivity of about 70 W/(m·K)
 

I wonder if disabling pbo also helps to cool my 5600x too. Currently I got it undervolted via curve optimizer and pbo still enabled.

By default, msi bios pumps way too nigh voltage and it overheats even when idling on bios ROFL.
 
Any source for ryzen 7640u having RDNA3 igpu?

Googling around, i keep finding the same table being recirculated with no link to the source.

7640u would be great as it'll probably cheaper than 77xx, but with zen 4 and interestingly RDNA3 igpu
 
Any source for ryzen 7640u having RDNA3 igpu?
This, I guess (and other earlier slides). Their lineup reveal says that 7x40 is "Phoenix" and earlier roadmap slides says "Phoenix Point" has RDNA3. It's a tiny bit of a jump, but not a long one.

AMD-Mobility-CPU-Roadmap.png
 
Back
Top