Why everyone say Zen 5 is bad ?

[iameatingjam]

Yeah, the i9 was a joke. Take £30-40 off of the 11900K's price, and you got an 11700K. Take another £30-40 off, and you got an 11700 non-K, which was physically identical to all of them, just with slightly lower clocks and a locked multiplier. I got that chip and never regretted it for a single second. In fact, I still have it in my living room HTPC. It's highly versatile thanks to its default 65 W power limit that can be tuned up to the sky, and its integrated Xe graphics does 4K output and videos like a charm without needing any extra power. Even when Alder Lake launched, I never thought about ditching the 11700 for a second.

I liked my 11600k too. Until it couldn't play xenoblade chronicles at 4x resolution on Yuzu at 60 fps. I know resolution is usually a GPU thing but I already had a 4090 so I knew it couldn't be that, so upgraded to a 13600k and it played like butter. Honestly though. Those mid-grade 11th gen chips were a good deal for gamers at the time. (as long as we are talking native games.... see damn you... you've got me doing it again).

 

[oxrufiioxo]

I mean us not liking or being disappointed with Zen 5 is first world problems sure but if Arrow Lake comes out and loses at times to Raptor lake I will be like WTF as well.

Would be crappy if we were stuck with only Nvidia giving us large performance gains every couple years lol

 

[AusWolf]

I had a 3900X and 9900K at the time so it was a nothing burger.... Also the 5800X with really tunned Bdie was noticeably faster in games I tried and unlike my 9900k that could do 4000 CL14 in it's sleep the 11900k which was only a single sample sucked past 3600mhz because of the gear 1/2 BS it was such a weird generation likely due to being stuck on an outdated process node.

I actually purchased a 5950X/5800X not long after it I had some hands on with the i9 lol.

I never cared for RAM tuning, so for me, the 11700 was fine. I even run my 7800X3D with 4800 MHz JEDEC for better boot times and lower power/heat.

I understand that with a 9900K on your hands, 10th or 11th gen wasn't interesting.

 

[danc]

Amd can use 4 IF links in zen5+ if they wanted to.

 

[iameatingjam]

I never cared for RAM tuning, so for me, the 11700 was fine. I even run my 7800X3D with 4800 MHz JEDEC for better boot times and lower power/heat.

I understand that with a 9900K on your hands, 10th or 11th gen wasn't interesting.

Especially when you play 4k at a fixed frame rate like me. Ram tuning is so... why? Its not going to matter in the least. Yet I still have gone diving in over and over. And even though I set my memory back to default xmp to take some stress off my cpu... I still couldn't help myself but to sneak just a couple timing improvements in there.

Just, if you don't like memory tuning, don't start, thats my advice to you

(Off topic but can we can get some more emoticons in here. I feel the selection is very limited and missing key things. And no I refuse to call them emojis. I'm the grumpy old man who calls things what they were called in my youth.)

 

[Zach_01]

Although I did have hands on with the 7800X3D did multiple builds with it felt it was fine used it for a week at home and said nope so you never know till you actually own something I guess.

Can you explain please why?
Honest curiosity...

 

[oxrufiioxo]

I never cared for RAM tuning, so for me, the 11700 was fine. I even run my 7800X3D with 4800 MHz JEDEC for better boot times and lower power/heat.

I understand that with a 9900K on your hands, 10th or 11th gen wasn't interesting.

On the amd side it mostly has to due with most boards just really sucking at setting timings I've seen 5-10% difference with just expo on a 6000 CL30 kit on different boards with a 7700X

It was the same on Am4 boards they just suck at timings.

I don't remember it being as bad on the intel side with just running higher frequency being enough.

 

[AusWolf]

On the amd side it mostly has to due with most boards just really sucking at setting timings I've seen 5-10% difference with just expo on a 6000 CL30 kit on different boards with a 7700X

It was the same on Am4 boards they just suck at timings.

I don't remember it being as bad on the intel side with just running higher frequency being enough.

Timings, shmimings... Bah... When I don't even feel a lick of difference between 4800 and 6000 MHz, why would I care about timings?

 

[oxrufiioxo]

Can you explain please why?
Honest curiosity...

Just what I use my cpu for outside of gaming it was much slower than I expected it to be... I was only expecting a slight drop from my 5950X with much better gaming performance but outside of gaming the 5950X felt much better and was much faster at the compression stuff I do. The 7950X3D costing me 120 usd more was a factor, also a buddy of mine bet me I would like it better and even offered to pay for it if I didn't was also a factor, I used both for about 2 weeks and decided the 7800X3D as not for me at least in my primary system.

I actually was still going to keep it and have it replace my 5800X but a buddy wanted it for what I paid for it so now I have an extra am5 board, it was free so not a big deal. I got it free because I couldn't get a couple usb to work properly but after a windows 11 update they started working super random but they had already sent me a replacement and told me to keep both.

Previously when setting up 7800X3D based systems I only assemble and stress tested them and ran a couple benchmarks to make sure they are performing properly in gaming so not enough hands on time I guess.

Timings, shmimings... Bah... When I don't even feel a lick of difference between 4800 and 6000 MHz, why would I care about timings?

It only matters when the gpu is fast enough to hit cpu limits and that's in most games for me lol... The thing is it's not the whole game that is the issue only specific parts that are more cpu limited than others so the extra % of free performance is worth it for me lol.... More so on non X3D chips though.

Especially in witcher 3 Next gen where even a 7800X3D/7950X3D can drop below 60fps.... It's probably my most replayed game.... Pretty sure I've seen every ending and variation at this point lol.

 

[AusWolf]

It only matters when the gpu if fast enough to hit cpu limits and that's in most games for me lol... The thing is it's not the whole game that is the issue only specific parts that are more cpu limited than others so the extra % of free performance is worth it for me lol.... More so on non X3D chips though.

Fair enough. My monitor is "only" 144 Hz, but it's a 3440x1440 ultrawide. I don't even care if I reach 144 FPS, anything above 50-60 is fine for me. I always go with maximum visual fidelity possible with my low performance target in mind, so I'm almost always GPU limited. I don't think I'm gonna feel the difference in RAM speed, ever. So to me, personally, it's much more beneficial to keep my RAM at a lower speed and save on power and heat. Interestingly, running it at 4800 instead of 6000 MHz lowers my VSOC all the way down to 1 V, which saves me 10 W, or around 5 °C on the CPU in all situations.

 

[chrcoluk]

@ARF More cores = nice boosts for productivity stuff like software encoding and compiling, great for hypervisors.

But if we talking common consumer desktop use and especially gaming, then usually performance per core is the way to go which is why X3D as an example is outperforming previous parts even parts that have more cores.

Otherwise we could have smaller 2ghz cores with tiny cache but about 40 of them, however things dont scale in that way.

 

[AusWolf]

Otherwise we could have smaller 2ghz cores with tiny cache but about 40 of them, however things dont scale in that way.

Yes, and games would still only use 4-6 of them at a time maximum, and would run like ass.

 

[igormp]

So you're saying, if they released a CPU that had more PCI-E lanes, and a new chipset designed around that CPU, the AM5 socket would not be able to accomdate this?

Yes.

I dont think this is correct I was under the impression the CPU and chipset decided the PCI-E lane quantity and speed, not the socket.

The CPU that dictates the amount of lanes that it has. In case of most AM5 CPUs, that'd be 28 lanes, 4 of which goes to the chipset.
If you take a look at the AM5 socket pinout, you'll that there are specific pins meant for PCIe:

You can't simply add more PCIe lanes to a CPU and not change that (otherwise those extra lanes would be useless), hence why you'd need a new socket.

One workaround is to have your chipset (which only has 4 lanes to the CPU) have some extra PCIe switches, which is already done by some motherboards to allow extra slots, or even slap down another chipset to it, like what's done in x670, but all of your downstream devices would still be bottlenecked by the x4 lanes going to the CPU nonetheless.

The problem with hybrid approach is the same as the e-core issues - the scheduling is a mess. I don't think the 7950X3D is a particularly good product, and a dense core CCD chip would probably be even worse.

I'm a linux user, and it works fine there, so I don't really see those issues

there is no way one can say in their right mind that ZEN 5 is good

I can, for my applications it has a great boost and will likely save hours of work per week for me, but I gotta admit that I'm like 1% of the overall market share.

About not enough memory bandwidth. Offers for DDR5 8000 begine from $184 for 32GB, and for 48GB from $230 both with "free" 1TB m.2 PCIe 4.0 SSD. I think that in 2026 when ZEN 6 with new IF, IMC, hmm I hope totally new cIOd, memory prices will be much more attractive and will have very good reasons for upgrading or building of new PC...

Even DDR5-8000 is not really enough to solve the memory bandwidth issue, the problem is being stuck with only dual-channel. Quad-channel would help a lot with that.

 

[TumbleGeorge]

Even DDR5-8000 is not really enough to solve the memory bandwidth issue, the problem is being stuck with only dual-channel. Quad-channel would help a lot with that.

It may seem like it at first glance, but that's more bandwidth than the Ryzen Threadripper 3990X gets for all its 64 cores with 128 threads.

 

[igormp]

It may seem like it at first glance, but that's more bandwidth than the Ryzen Threadripper 3990X gets for all its 64 cores with 128 threads.

Yeah, but we're talking about a CPU from 2020 that had 100GB/s total bandwidth (~1.5GB/s per core) which people already complained A LOT, hence why TR Pro became a thing with double the channels (and the lack of RDIMM support was also a PITA), and which lacked AVX-512.
Now we have cores that are way more powerful (specially with those HUGE AVX-512 units) and still stuck at ~100GB/s with heavily clocked DIMMs (and if you opt for 2DPC for extra capacity, you will go back to ~60GB/s, or even less).

Quoting y-cruncher's creator once again:

For the 16-core 9950X to not be bottlenecked by memory here, it needs DDR5-20000. And y-cruncher is probably one of the better memory-optimized programs on this side of dense matrix. Common workloads like FFT, and BLAS levels 1 and 2 stand no chance.

This memory bottleneck is what keeps Intel competitive in the HPC space. You don't need to have a powerful CPU if it's all going to be held back by memory access anyway. Thus AMD's impressive AVX512 implementation really only shines in low-core-count SKUs or in embarassingly parallel workloads that don't touch memory.

 

[TumbleGeorge]

Yeah, but we're talking about a CPU from 2020 that had 100GB/s total bandwidth (~1.5GB/s per core) which people already complained A LOT, hence why TR Pro became a thing with double the channels (and the lack of RDIMM support was also a PITA), and which lacked AVX-512.
Now we have cores that are way more powerful (specially with those HUGE AVX-512 units) and still stuck at ~100GB/s with heavily clocked DIMMs (and if you opt for 2DPC for extra capacity, you will go back to ~60GB/s, or even less).

Quoting y-cruncher's creator once again:

Memory bandwidth per core with DDR5 8000 also will be greater than Threadrippers from 7000 series. Only the AMD Epycs from latest series in small or middle core models will have more bandwidth per core. Yes when comparing total bandwidth is different situation.

 

[Carillon]

Y-cruncher seems to be the only benchmark really bottleneck by memory bandwidth, phoronix shows big gains almost everywhere.
I remember rocket lake being 10x faster in y-cruncher using avx512 over avx2. The guy is probably the only person on earth who knows how to properly use avx512.

 

[igormp]

Memory bandwidth per core with DDR5 8000 also will be greater than Threadrippers from 7000 series.

TR 7000 should also have DDR5 8000 available, but yeah, the non-Pro still has the quad-channel limitation for the higher-core models.

Y-cruncher seems to be the only benchmark really bottleneck by memory bandwidth, phoronix shows big gains almost everywhere.
I remember rocket lake being 10x faster in y-cruncher using avx512 over avx2. The guy is probably the only person on earth who knows how to properly use avx512.

There are some others where the jump from a 9900x to a 9950x is way smaller than the jump from a 9700x to a 9900x. However, I'm not 100% sure if that's solely due to memory bandwidth, or because the dual-CCD models can do double the memory ops.

 

[Melvis]

The problem mentioned, SMT, administrator, affect every CPU not only AMD, just go look the reviews of TPU and Hardware unboxed.

I have seen the Hardware Unboxed video and it didnt show anything about intel CPU's only AMD.

ZEN 5 is, overall, about 3% faster and costs a lot more. Exemple 9700x cost 380 to 450 euros compared to the 7700X who costs 280, the 7700 can be found for around 200 euros in the tray version. The price at launch does not matter. Yes ZEN 5 is a bit more efficient, but who cares really ? The financial advantage on a year's bill of electricity is laughable.

See again you havent looked at many reviews and seem to be locked on one thing and stating the obvious, of course the older Gen CPU's are going to be cheaper and isnt worth the extra cost etc etc, this applies to pretty much every single launch of new Products, just go look at intel....

See this is what I dont get and shows how narrow minded people are......this launch is buggy AF for AMD but your still getting a performance increase and maybe more once all the bugs are sorted, then you go back and look at intels 13th to 14th Gen, literally no increase in performance at all, just like Skylake from 6th gen to 10th gen, no increase at all, just clock speed bumps and or more cores, thats it! and you dont hear people loosing there shit over that? but when AMD comes out with a new line of CPU's and its not as big a jump as everyone wanted, even with the bugs, everyone is up in arms over it.....thats just so blows my mind! and shows how greedy people are.

I've watched a shit load of reviews, and base on the results of the benchmark, there is no way one can say in their right mind that ZEN 5 is good, it is moreover overpriced compared to its performance.

No you havent or you wouldnt of come to this conclusion thats for sure, and I showed you proof and funny enough you didnt say one word about it....funny that! and again.....your to narrowed minded and dont understand that this is how things work, your comparing old vs new instead of launch vs launch where its actually cheaper.........compared to last gen at launch.

Right now, ZEN 5 is not selling and there is a reason for that, ZEN 5 is really disappointing, and at that price, a generation to skip

Which most people would anyway.......your stating the obvious here but missing the point....again!

AMD overhyped this CPU is a shameless manner, it was totally dishonest and now they come with things like SMT, Admin bug...whatever...which have an effect, but with every CPU, not only AMD, so the performance ratio stay the same, that's what I call tricks.

There is absolutely no reasons to buy ZEN 5 right now, you can find cheaper alternatives for the same level of performance.

Ill admit it was overhyped but not to the agree everyone is bitching about! if people didnt have such narrow minded minds and actually look over the fence then they might see a shameless manner......and im still waiting for you to send me this link to support your claims.....but again im still waiting.

Ill just drop another review here for you to look at and then you tell me if the % is only 3% (not including gaming) and even they say they will have to redo all Zen 5 CPU reviews because of the bugs.

 

[JWNoctis]

Yeah, but we're talking about a CPU from 2020 that had 100GB/s total bandwidth (~1.5GB/s per core) which people already complained A LOT, hence why TR Pro became a thing with double the channels (and the lack of RDIMM support was also a PITA), and which lacked AVX-512.
Now we have cores that are way more powerful (specially with those HUGE AVX-512 units) and still stuck at ~100GB/s with heavily clocked DIMMs (and if you opt for 2DPC for extra capacity, you will go back to ~60GB/s, or even less).

The worse problem is, things are still apparently stuck to ~60GB/s 1CCD / ~75GB/s 2CCD anyway, however faster you'd run the RAM, since you are already being bottlenecked by the IF at any frequency as by default it runs at 2:3 theoretical bandwidth compared to the memory controller, and if I understood it right, does not go much faster than that.

That AFAICT is still inferior to a quad-channel DDR4 Zen 2 non-Pro Threadripper. My own 1CCD Zen4 benchmarks <60GB/s on a memory bandwidth bound workload, on DDR5-6000 with 96GB/s theoretical bandwidth. I wonder what would one find if they were to, say, run a 2DPC DDR5-4800 and keep the FCLK at 2000MHz. There probably will be some heavy latency penalty, but bandwidth presumably won't suffer too much, if the above logic is correct.

Zen 6 will probably come up with something different. Widened IF that does not bottleneck and does not cause further idle power problem would be good, if they would stay on AM5 for another generation.

 

[Wolverine2349]

What does this mean? When will AMD be ready for 2 nm? In 2030?

What I want is Ryzen 7 9700X to be a 12-core (if not, then at least 10-core) with 50% more cache than now, made on the 3nm process node, being a monolithic chip.
This is my expectation.

Yes bingo exactly.

AMD keeping same exact core count since Zen 2 not good. Though that is not so much the issue.
Bigger issue is I badly want more than 8 cores on a single CCX within a single CCD.

AMD since Zen 3 has had 8 cores per CCX within a CCD which is why it improved so much over Zen 2. Since then AMD has kept it at 8 cores per CCX which is a shame through Zen 5 and given this unimpressive uplift that is unacceptable. And sadly it appears its going to be the case with Zen 6 as well at least on consumer platform per rumors. I mean they kind of need to increase core count to make it effective and stop selling 6 core parts for the lower end. Instead have 12 core CCX/CCDs and sell 24 core CPUs with dual CCD/CCXs

That way we get excellent latency across more than 8 core for core to core communication and a native more than 8 core in this case 12 core CPU.

They could cut out defective dies and sell 8-10 cores as the new low end replacement for the single CCD 6 cores and 16-20 core chips dual CCD for the dual CCD 12-16 core replacements.

But AMD is insistent on keeping 6 and 8 core and 12 and 16 core on consumer platform. And they are too cheap to do a full 12 or 16 core die because its much cheaper to just have single 8 core CCX/CDs and just package 2 of them in one chip and cut 2 defective dies out for the 6 and 12 core parts even though cross CD latency penalty is bad for thread heavy gaming in many cases. Basically AMD has one die on chiplets since Zen 3 through 5 and for the threadripper parts they just cram a bunch more 8 core CCX/CCDs into the high core count threadripper parts cause its cheapest to have only one stepping. And they do same for consumer parts just up to 2 CCDs maxing at 8 cores each.

It sucks no more than 8 cores on a single CCX and this release is embarrassingly bad as such with almost no IPC uplift in general consumer workloads and gaming. Even with a better IPC uplift still really want to see more than 8 homogenous cores on a single node from either Intel or AMD. Bartlett Lake appears to be the only hope on rumored products sadly for more than 8 cores on a single die coming anytime soon sadly. And even that is at least close to 1 year away or further.

 

[Shakallia]

I have seen the Hardware Unboxed video and it didnt show anything about intel CPU's only AMD.



See again you havent looked at many reviews and seem to be locked on one thing and stating the obvious, of course the older Gen CPU's are going to be cheaper and isnt worth the extra cost etc etc, this applies to pretty much every single launch of new Products, just go look at intel....

See this is what I dont get and shows how narrow minded people are......this launch is buggy AF for AMD but your still getting a performance increase and maybe more once all the bugs are sorted, then you go back and look at intels 13th to 14th Gen, literally no increase in performance at all, just like Skylake from 6th gen to 10th gen, no increase at all, just clock speed bumps and or more cores, thats it! and you dont hear people loosing there shit over that? but when AMD comes out with a new line of CPU's and its not as big a jump as everyone wanted, even with the bugs, everyone is up in arms over it.....thats just so blows my mind! and shows how greedy people are.



No you havent or you wouldnt of come to this conclusion thats for sure, and I showed you proof and funny enough you didnt say one word about it....funny that! and again.....your to narrowed minded and dont understand that this is how things work, your comparing old vs new instead of launch vs launch where its actually cheaper.........compared to last gen at launch.



Which most people would anyway.......your stating the obvious here but missing the point....again!


Ill admit it was overhyped but not to the agree everyone is bitching about! if people didnt have such narrow minded minds and actually look over the fence then they might see a shameless manner......and im still waiting for you to send me this link to support your claims.....but again im still waiting.

Ill just drop another review here for you to look at and then you tell me if the % is only 3% (not including gaming) and even they say they will have to redo all Zen 5 CPU reviews because of the bugs.

Won't bother to answer, your replies are beside the points made, reading comprehension is not your best asset.

 

[Assimilator]

The worse problem is, things are still apparently stuck to ~60GB/s 1CCD / ~75GB/s 2CCD anyway, however faster you'd run the RAM, since you are already being bottlenecked by the IF at any frequency as by default it runs at 2:3 theoretical bandwidth compared to the memory controller, and if I understood it right, does not go much faster than that.

That AFAICT is still inferior to a quad-channel DDR4 Zen 2 non-Pro Threadripper. My own 1CCD Zen4 benchmarks <60GB/s on a memory bandwidth bound workload, on DDR5-6000 with 96GB/s theoretical bandwidth. I wonder what would one find if they were to, say, run a 2DPC DDR5-4800 and keep the FCLK at 2000MHz. There probably will be some heavy latency penalty, but bandwidth presumably won't suffer too much, if the above logic is correct.

Zen 6 will probably come up with something different. Widened IF that does not bottleneck and does not cause further idle power problem would be good, if they would stay on AM5 for another generation.

FWIW, MLC reports a best-case of 75GB/s and worst-case of 66.5GB/s for the system in my specs. So 2-CCD Zen 5 is about as fast as quad-channel TR Zen 2.

ALL Reads        :      75066.4
3:1 Reads-Writes :      68112.1
2:1 Reads-Writes :      67747.7
1:1 Reads-Writes :      66531.8
Stream-triad like:      70028.3

 

[AusWolf]

Yes bingo exactly.

AMD keeping same exact core count since Zen 2 not good. Though that is not so much the issue.
Bigger issue is I badly want more than 8 cores on a single CCX within a single CCD.

AMD since Zen 3 has had 8 cores per CCX within a CCD which is why it improved so much over Zen 2. Since then AMD has kept it at 8 cores per CCX which is a shame through Zen 5 and given this unimpressive uplift that is unacceptable. And sadly it appears its going to be the case with Zen 6 as well at least on consumer platform per rumors. I mean they kind of need to increase core count to make it effective and stop selling 6 core parts for the lower end. Instead have 12 core CCX/CCDs and sell 24 core CPUs with dual CCD/CCXs

That way we get excellent latency across more than 8 core for core to core communication and a native more than 8 core in this case 12 core CPU.

They could cut out defective dies and sell 8-10 cores as the new low end replacement for the single CCD 6 cores and 16-20 core chips dual CCD for the dual CCD 12-16 core replacements.

But AMD is insistent on keeping 6 and 8 core and 12 and 16 core on consumer platform. And they are too cheap to do a full 12 or 16 core die because its much cheaper to just have single 8 core CCX/CDs and just package 2 of them in one chip and cut 2 defective dies out for the 6 and 12 core parts even though cross CD latency penalty is bad for thread heavy gaming in many cases. Basically AMD has one die on chiplets since Zen 3 through 5 and for the threadripper parts they just cram a bunch more 8 core CCX/CCDs into the high core count threadripper parts cause its cheapest to have only one stepping. And they do same for consumer parts just up to 2 CCDs maxing at 8 cores each.

It sucks no more than 8 cores on a single CCX and this release is embarrassingly bad as such with almost no IPC uplift in general consumer workloads and gaming. Even with a better IPC uplift still really want to see more than 8 homogenous cores on a single node from either Intel or AMD. Bartlett Lake appears to be the only hope on rumored products sadly for more than 8 cores on a single die coming anytime soon sadly. And even that is at least close to 1 year away or further.

What do you need the extra cores for so badly? I'm not picking on you, just trying to find the reason behind your statements.

 

[SL2]

Because people expect double the performance each generation.

I think you know that's wrong lol.