News Two AMD Zen 5 CPUs may receive a significant performance uplift — Ryzen 7 9700X and Ryzen 5 9600X rumored for 105W TDP option with next AGESA update

[logainofhades]

They are marginally better than their 7000 series counterparts, and cost more. Once the 7000 series supply dries up, and prices are cut on the 9000 they will be compelling options. Anyone building now, should just save the cash and get a 7000 series chip, and use the savings elsewhere.

 

[bit_user]

Greed entirely; moreoever, 9700X compares pretty closely to 7700 in performance in a lot of workloads

"A lot"? ? It's a bigger uplift than we saw from most generations of Intel CPUs, during the quad-core era.

I think you might be confused. There are cases where it performs close to the 7700X, but using much less power. Once you enable PBO, the 9700X easily outperforms even the 7700X, in those cases.

and actually has similar performance-per-watt as well (actually worse in some cases, going against the "Zen 5 is SUPER efficient" marketing notion).

Gaming is the main area where it regresses on efficiency. Gamers Nexus performed a detailed investigation where they found it's consistently spiking its clocks higher than the 7700X. In cases where the previous gen was still fast enough to get about the same work done per frame period, doing so at lower clockspeeds was a more energy-efficient strategy.

Even so, Gamers Nexus showed Zen 5's gaming efficiency is still far above that of Raptor Lake.

GN actually shows how Zen 5 is boosting in comparison to Zen 4 counterparts, and generally it's lower,

Oops, you got that backwards.

Interestingly, PBO has shown to run away on power consumption without significantly increasing performance in many gaming workloads,

Not according to this:

​

That shows only a 11.3% increase in power consumption, when enabling PBO in gaming.

 

[bit_user]

For NON-GAMING workloads how? Based on HU video the uplift in common apps like Adobe, Blender etc. is minimal almost none.

Hard to know what someone means by "almost none", but it offers meaningful increases in lightly-threaded apps and PBO provides a 22% uplift in Blender:

​

The multithreaded speedups aren't on the same magnitude as adding more cores, but that's as expected.

The problem with the more expensive models is that there is hardly any improvement over last gen. We talkin 1-5% at best.

No, this is not accurate. Single-threaded CB R24 is 12.5% faster, comparing stock with stock (9700X with 7700X).

 

[Papusan]

"A lot"? ? It's a bigger uplift than we saw from most generations of Intel CPUs, during the quad-core era.

Nowadays we have more competition, yet AMD fail to deliver. So this is even worse than in the Intel quad-core era where it was no competition. Yup, AMD's Skylake moment. Who expected this when the competition is harder than ever? Hmmm... This make a bad taste in my mouth...

View: https://youtu.be/PHCGvhmfCOw?t=586

And the low sales just confirm the bad value of 9000 series Ryzen chips.

AMD Ryzen 7 9700X & Ryzen 5 9600X "Zen 5" CPU Sales Off To A Slow Start, Only 50 Chips Sold At Germany's Largest Retailer During First Week

AMD's Ryzen 7 9700X & Ryzen 5 9600X "Zen 5" CPUs are off to a slow start as their sales figures highlight a weak reception from PC builders.

wccftech.com

 

[bit_user]

Nowadays we have more competition, yet AMD fail to deliver. So this is even worse than in the Intel quad-core era where it was no competition. Yup, AMD's Skylake moment. Who expected this when the competition is harder than ever? Hmmm... This make a bad taste in my mouth...

I think that's somewhat an exaggeration, but the way I look at it is that Zen 5 didn't increase core counts or cache, nor did it make an aggressive move on process node. So, the performance difference was a product of IPC * clock speed. The IPC increase was decent for a generation without substantial cache increases. The clock speed picture is complicated by what was mentioned above.

In case you haven't heard, there's also the Windows scheduler controversy, which has been deepened with talk of core parking.

We found the Missing Performance: Zen 5 Tested with SMT Disabled

Reviews of AMD’s Zen 5 processors this week surprised many, with lower-than-expected results. After some investigation, we discovered that turning off Simultaneous Multithreading (SMT) can yield notable performance gains, particularly in gaming. This article presents our findings, including...

www.techpowerup.com

It does sound like some of Ryzen 9000's woes could be addressed through fixing software issues and better frequency management. I still think it's not a bad core, even if it's a little underwhelming. If I were buying an AM5 CPU today, I would indeed bet on Ryzen 9000.

 

[Hotrod2go]

My early testing with 9700X in my sig rig with prime95 blend for over an hour already, HWiNFO reports 118w average (with AVX-512 disabled) using AGESA 1.2.0.0. Think Asrock have implemented this with higher TDP for Zen 5 but this is with all auto bios settings, PBO enabled with EXPO profile.

 

[Guardians Bane]

Hm... If they add an "ECO" mode of 105W and keep the current as 65W alongside, then it'll be fine. Changing the default behaviour to 105W I do not like. Specially since people complaining about the performance of these CPUs is not really interested in the MT aspect of them. AMD swinging and missing like a blind child again?

Also, wasn't this also rumoured before the launch even? Am I remembering wrong? Well, not that it matters, I guess.

Regards.

I remember that rumor also. That's the first thing I thought of when I saw this article. But that rumor was for 120w. So it makes me wonder about that being something else to try. No idea if that wattage would benefit in anyway though. I'm on a 6900hx based mini PC and so far I am probably going with the 7000 series SKUs. The price/performance is better with 7000 currently. I will admit, I was absolutely underwhelmed with 9000. But one way or the other, in building an AM5 system very soon. Just going from MoDT to full fat DT will bring plenty of performance gains for me. Lol

 

[-Fran-]

I remember that rumor also. That's the first thing I thought of when I saw this article. But that rumor was for 120w. So it makes me wonder about that being something else to try. No idea if that wattage would benefit in anyway though. I'm on a 6900hx based mini PC and so far I am probably going with the 7000 series SKUs. The price/performance is better with 7000 currently. I will admit, I was absolutely underwhelmed with 9000. But one way or the other, in building an AM5 system very soon. Just going from MoDT to full fat DT will bring plenty of performance gains for me. Lol

It's fine to skip 9000 if the value or the things you use don't take advantage of it. This has been true for a lot of generations where things are added, but are not used actively and take advantage of it (look at RTX, for example). Ry7K has AVX512, but just to tick the box and Ry9K implemented it properly and added VNNI, so it's actually good for specific workloads not many use or take advantage of right now.

The rule is always: if the CPU doesn't do what you need right now, don't buy into the "future" promises unless you can use those in the near future. Sometimes the "fine wine" effect is just coping.

I'll wait for the VCache'd siblings and check then if I go for the 9950X or the purported 9950X3D.

Regards.

 

[bit_user]

Ry7K has AVX512, but just to tick the box and Ry9K implemented it properly and added VNNI, so it's actually good for specific workloads not many use or take advantage of right now.

Phoronix tested the 7950X with/without AVX-512 and measured a 59.0% improvement (geomean) across benchmarks designed to take advantage of it.

https://www.phoronix.com/review/amd-zen4-avx512
​

So, I think you really can't say that Zen 4 added it only as a box-ticking exercise. For people who could actually use it, even the Zen 4 implementation provided very worthwhile benefits.

If you know a little about AVX-512, it's not too surprising to see some real speedup vs. the AVX2 implementation. Anandtech famously used a benchmark called 3D Particle Movement, that I think was essentially from Ian Cutress' doctoral research, which had both AVX2 and AVX-512 versions. The latter was faster by almost a suspicious amount.

 

[-Fran-]

Phoronix tested the 7950X with/without AVX-512 and measured a 59.0% improvement (geomean) across benchmarks designed to take advantage of it.

​

https://www.phoronix.com/review/amd-zen4-avx512​

​

So, I think you really can't say that Zen 4 added it only as a box-ticking exercise. For people who could actually use it, even the Zen 4 implementation provided very worthwhile benefits.

If you know a little about AVX-512, it's not too surprising to see some real speedup vs. the AVX2 implementation. Anandtech famously used a benchmark called 3D Particle Movement, that I think was essentially from Ian Cutress' doctoral research, which had both AVX2 and AVX-512 versions. The latter was faster by almost a suspicious amount.

I meant that comparatively to Ry9K. It has it and can run AVX512 stuff with a speedup, but the "proper" implementation is in Ry9K.

There's also the nuance of what of AVX512 they've implemented across, which Intel just screwed up and made super confusing, but at least AMD seems to support all the important stuff. There was a Venn diagram of it somewhere xD

Regards.

 

[bit_user]

I meant that comparatively to Ry9K. It has it and can run AVX512 stuff with a speedup, but the "proper" implementation is in Ry9K.

FWIW, I think that's a somewhat artificial distinction. Zen 1 implemented AVX in 128-bit chunks, but I don't recall people making a big deal that it wasn't "proper", at the time.

Also, AMD said the laptop version of Zen 5 (and yes, I include the full Zen 5 core, not just Zen 5C) implements AVX-512 in 256-bit chunks, like Zen 4 did (note: while AMD calls this "double-pumped"; I don't).

There's also the nuance of what of AVX512 they've implemented across, which Intel just screwed up and made super confusing, but at least AMD seems to support all the important stuff. There was a Venn diagram of it somewhere xD

The wikipedia page has a table showing which CPUs implement the different subsets.

https://en.wikipedia.org/wiki/AVX-512#CPUs_with_AVX-512
​

Basically, Zen 4 caught up with Ice Lake + BF16. Zen 5 caught up with Tiger Lake + BF16. The only thing Zen 5 seems to be missing is fp16 support, which is somewhat niche in its use cases, particularly when you already have BF16.

 

[-Fran-]

FWIW, I think that's a somewhat artificial distinction. Zen 1 implemented AVX in 128-bit chunks, but I don't recall people making a big deal that it wasn't "proper", at the time.

Also, AMD said the laptop version of Zen 5 (and yes, I include the full Zen 5 core, not just Zen 5C) implements AVX-512 in 256-bit chunks, like Zen 4 did (note: while AMD calls this "double-pumped"; I don't).

Yeah, that's fine. As you say, it's just a superficial distinction. The meaning I give it is instruction throughput. Zen5 just increased it massively. That's all.

Technically they both implement it "properly".

The wikipedia page has a table showing which CPUs implement the different subsets.

​

https://en.wikipedia.org/wiki/AVX-512#CPUs_with_AVX-512​

​

Basically, Zen 4 caught up with Ice Lake + BF16. Zen 5 caught up with Tiger Lake + BF16. The only thing Zen 5 seems to be missing is fp16 support, which is somewhat niche in its use cases, particularly when you already have BF16.

Yes, this is what I was alluding to.

Regards.

 

[stuff and nonesense]

I'm afraid some people are just way too hung up on TDP, these days.

When the first Acorn, ARM processor fired up it ran on leakage current, Vcc wasn’t connected…

 

[Roland Of Gilead]

I think this is problematic. The CPUs are being sold as 65W, so almost any cheap cooler will suffice. But one BIOS update later, and the chip will run too hot for these coolers, increasing noise and possibly throttling under heavy loads. They really should let PBO do that task for them, and launch a 9600XT or something. Sound nice in enthusiast terms, but most users won't see the increased TDP coming.

Good point!

 

[bit_user]

Phoronix tested the 7950X with/without AVX-512 and measured a 59.0% improvement (geomean) across benchmarks designed to take advantage of it.

https://www.phoronix.com/review/amd-zen4-avx512
​

Oh, sweet! Phoronix just reprised this and added data for the Ryzen 9950X!

https://www.phoronix.com/review/amd-zen5-avx-512-9950x
​

The gist of it is that enabling AVX-512 added 55.8% more performance on the 9950X, while this round of tests showed a benefit of just 41.0% from enabling it on the 7950X. Even without AVX-512, the 9950X was 15.3% faster than the 7950X; with AVX-512, the Zen 5 CPU was 27.4% faster than its Zen 4-based predecessor.

So, baseline floating point performance improved on Zen 5 (which we already knew from what AMD said about Zen 5), but the wider AVX-512 implementation resulted in even bigger relative improvements on Zen 5 than using AVX-512 on Zen 4.

The article also showed that average power consumption increased by only about 3W on each CPU, when enabling AVX-512. However, it showed that the new CPU is using about 20W less than the Zen 4-based 7950X, in each case.

 

[usertests]

A 9700X with a higher TDP sounds like the kind of thing that AMD would call a 9800X and sell for more money.

That's what a certain techtuber has recommended, except for the price increase. Something to launch after Windows/drivers are fixed so that the chips can be reviewed again:

In a way it's kind of fair, since they're using 9700X + PBO to approximate 9700X @ 105W, which is the same TDP the old 7700X had. Granted, the 9700X can use much more than 105 W under PBO, but going to 105 W probably gives you most of the gains to be had.

Does Ryzen Master or other software not allow people to set the TDP of the 9700X to 105W (different than PBO)? Ars Technica seemed to have done that in this review:

AMD should have released these as the 9600 and 9700. Direct upgrades for the respective 7600 and 7700 65W parts
Instead, they chose to shoot themselves in the foot.

There seems to be other problems, related to these chips running Windows 11 (and not just the privilege bug that has already been identified). Increasing the TDP could help but won't fix everything.

 

[bit_user]

Does Ryzen Master or other software not allow people to set the TDP of the 9700X to 105W (different than PBO)? Ars Technica seemed to have done that in this review:

LOL, the last two entries seem to be swapped.

 

[Geezen]

What I don't get is how they insist on keeping the 9600x and 9700x at the same TDP, either 65w or 100w. SPLIT them, like the 9900x and 9950x. Just bizarre.

 

[bit_user]

What I don't get is how they insist on keeping the 9600x and 9700x at the same TDP, either 65w or 100w. SPLIT them, like the 9900x and 9950x. Just bizarre.

There's no "right" answer, here. Letting the 6-core 9600X use more power per core helps it partially make up for having fewer cores. That's one way to look at it. Either way, you probably need to load up at least 4 of the cores (with 1 thread each), before the CPU reaches 65 W.

Actually, a reason to specify exactly 65 W might be due to the selection of heatsinks that enables. 65 W is a very popular TDP figure, in the PC industry. I'll bet there are lots of inexpensive heatsinks available for dissipating that much heat. If they scaled up the 9700X to use the same power per core, you'd end up with 87 W, which might push you into using a higher tier of heatsinks that would add cost for OEMs. Since most games & productivity apps don't have enough threads to benefit much from > 65 W, why increase it at all? You'd be adding cost (i.e. requiring bigger heatsink) for performance most users wouldn't even see.

 

[Geezen]

There's no "right" answer, here. Letting the 6-core 9600X use more power per core helps it partially make up for having fewer cores. That's one way to look at it. Either way, you probably need to load up at least 4 of the cores (with 1 thread each), before the CPU reaches 65 W.

Actually, a reason to specify exactly 65 W might be due to the selection of heatsinks that enables. 65 W is a very popular TDP figure, in the PC industry. I'll bet there are lots of inexpensive heatsinks available for dissipating that much heat. If they scaled up the 9700X to use the same power per core, you'd end up with 87 W, which might push you into using a higher tier of heatsinks that would add cost for OEMs. Since most games & productivity apps don't have enough threads to benefit much from > 65 W, why increase it at all? You'd be adding cost (i.e. requiring bigger heatsink) for performance most users wouldn't even see.

I'd argue that if a user isn't taking advantage of the 8 cores, they shouldn't even be buying that version. It's still a weird decision to lump 6 and 8 core parts as the same TDP.

 

[Guardians Bane]

It's fine to skip 9000 if the value or the things you use don't take advantage of it. This has been true for a lot of generations where things are added, but are not used actively and take advantage of it (look at RTX, for example). Ry7K has AVX512, but just to tick the box and Ry9K implemented it properly and added VNNI, so it's actually good for specific workloads not many use or take advantage of right now.

The rule is always: if the CPU doesn't do what you need right now, don't buy into the "future" promises unless you can use those in the near future. Sometimes the "fine wine" effect is just coping.

I'll wait for the VCache'd siblings and check then if I go for the 9950X or the purported 9950X3D.

Regards.

Right on. I

It's fine to skip 9000 if the value or the things you use don't take advantage of it. This has been true for a lot of generations where things are added, but are not used actively and take advantage of it (look at RTX, for example). Ry7K has AVX512, but just to tick the box and Ry9K implemented it properly and added VNNI, so it's actually good for specific workloads not many use or take advantage of right now.

The rule is always: if the CPU doesn't do what you need right now, don't buy into the "future" promises unless you can use those in the near future. Sometimes the "fine wine" effect is just coping.

I'll wait for the VCache'd siblings and check then if I go for the 9950X or the purported 9950X3D.

Regards.

Right on. I hope the X3D SKUs are better than 7800x3d. It be nice to see a new AMD gaming king. For me an X SKU is better. But I will wait to see actual benefits to going beyond Zen 4 (talking price/performance and actual good performance uift over 7000)

Regards back to ya

 

[bit_user]

I'd argue that if a user isn't taking advantage of the 8 cores, they shouldn't even be buying that version.

In the event that they do use 8 cores, it does actually perform better than 6, even within the same TDP.

​

So, that's 23.6% faster, at the same power, for having 33% more cores. Not bad, IMO. This shouldn't come as a surprise, if you consider that CPU cores increase power consumption at a nonlinear rate with respect to frequency. Therefore, you can run more cores in the same power budget, just by trimming top end frequencies only a little bit.

It's still a weird decision to lump 6 and 8 core parts as the same TDP.

I think you probably aren't familiar with AMD's portfolio of non-X models or Intel's range of non-K CPUs. In AMD's case, they range from 6 to 12 -core parts (7600, 7700, 7900) all at 65 W. In Intel's case, the models going from the 10-core i5-14400 to the 24-core i9-14900 are all rated at 65 W.

 

[Reality_checker]

Raising the default TDP is called a performance uplift these days?

 

[bit_user]

Raising the default TDP is called a performance uplift these days?

Benchmarks with/without PBO show that it's good for boosting multithreaded performance, but that's about it. In the graph I quoted above, the 9700X gains 14.0% from PBO, although it's using well above 105 W, by that point. So, we should expect to see something less (but not insignificant).

In the case of the 9600X, PBO yields a mere 9.4% improvement in MT scores, since it's already running at more W/core than the 9700X, as is.

 

[stuff and nonesense]

Raise the TDP, allow the processor to draw more power. This has a small effect single threaded but lets multithreaded workloads perform better.

Voltages within the cpu are better supported, slew rates are maintained, voltage levels are maintained. The transistors are allowed to work.

Single thread, the processor is mostly running well within its power limits at max clock but divide the power between many cores and the available power is spread across them. The clock that can be maintained drops. Raise the power, raise the voltage that can be applied to the cores, raise the frequency closer to the silicon limits.