AMD Zen 5 Technical Deep Dive

[ARF]

I know a whole lot of people who are switching from 13900K/14900K because they don't have to deal with instability anymore and they'll have a nice performance bump all around while saving power. So I guess the timing works pretty well for AMD here. I will certainly get a 9950X simply because it'll speed up my workloads and a potential 20% bump over a 7950x with some nice memory tuning would go a long way without having to go TR.

What instabilities? If this is true and global, this is news-worthy. Why don't they undervolt and underclock, apply "eco" modes in the BIOS?

 

[Dr. Dro]

Please AMD, don't fumble the pricing.
It's looking nice, but getting cocky with bad pricing would be horrible.

If these CPUs significantly outperform Intel's offerings, THEY WILL CHARGE ACCORDINGLY. There is precedent for this, namely Zen 3 and specifically the $300 Ryzen 5 5600X before the i5-12400 dethroned it as the midrange king for practically half the price.

 

[Sabotaged_Enigma]

AMD could actually pull all C-core processors for the Athlon and Ryzen 3 lines. In that end of the market, high clocks aren't necessary.

Nice thought!
I'm also wondering why they haven't made new Ryzen 3's and Athlons. Of course that's bacause there aren't so many bad dies to make them. But low-end market is still important (at least to certain users). Think they should either make a smaller monolithic die having 4 cores, or just use compact version cores like you said.

 

[ARF]

Nice thought!
I'm also wondering why they haven't made new Ryzen 3's and Athlons.

Because they sell Athlon 3000G (Q4 2019) and Ryzen 3 2200G (Q1 2018).

Greed and stupidity.

AMD thinks that they can get away with a 6-year-old CPU being sold as brand new to the masses.

No wonder why their revenues are on the decline, and they don't expect better times...

 

[mkppo]

What instabilities? If this is true and global, this is news-worthy. Why don't they undervolt and underclock, apply "eco" modes in the BIOS?

It is news..

It's been for a while and intel has done jack all about it, there's a whole other thread out there in these forums. More and more people are reporting these issues now. I have two friends, one with 13th and one with 14th with both reporting video out of memory issues with a less than 1 year old CPU. One even RMA'd their GPU lol but as we know it now, it's Intel.

Not everyone has the time, or wants to put a band aid on a $600 CPU and lower it's performance. They're RMA'ing it and going AMD because why not have more performance, less power and better reliability? Intel still hasn't identified the issue or put a fix, only thing they've done is try to pin the blame on others.

 

[Dr. Dro]

Nice thought!
I'm also wondering why they haven't made new Ryzen 3's and Athlons. Of course that's bacause there aren't so many bad dies to make them. But low-end market is still important (at least to certain users). Think they should either make a smaller monolithic die having 4 cores, or just use compact version cores like you said.

Industry has been attempting to pivot from the budget segment for a while. Intel's the only one making low-end processors on the mainstream socket left, and they're widely derided for it (Intel 300/310 being Alder Lake-based dual core processors with 2 P-cores, but retaining hyper-threading and full instruction set including AVX2 and AVX-VNNI at 3.9/4 GHz respectively). AMD hasn't released a Ryzen 3 for socket AM5 yet, with the quad-core EPYC variant of Raphael being the lowest-end chip and the 7500F being at the bottom of the ladder on the consumer Ryzen side. Might/will change in the future.

AMD thinks that they can get away with a 6-year-old CPU being sold as brand new to the masses.

No wonder why their revenues are on the decline, and they don't expect better times...

That's because they can and do. It's cheap and there's a market. Not everyone needs the fastest, latest and greatest, and the TCO of the Intel 300 processor I mentioned earlier is still higher due to the cost of LGA 1700 motherboards being higher than socket AM4 on average.

 

[ARF]

That's because they can and do. It's cheap and there's a market. Not everyone needs the fastest, latest and greatest, and the TCO of the Intel 300 processor I mentioned earlier is still higher due to the cost of LGA 1700 motherboards being higher than socket AM4 on average.

And these people get unpleasantly surprised by the new Windows update, after which their systems begin crawling like snails.
You know that this excuse can't be serious?
You need a modern CPU with a modern instruction set, modern manufacturing process to offset the power consumption, enough cores, because what was enough in 2018 is not enough in 2024, not to mention in a couple of years..

 

[rv8000]

When gaming with a 4090 at FHD, the 7800X3D is more than 25% faster than the 5800X3D in TPU's review from last year.

I was assuming an average across all resolutions. It’s all guesswork until we have a review anyways.

 

[Dr. Dro]

And these people get unpleasantly surprised by the new Windows update, after which their systems begin crawling like snails.
You know that this excuse can't be serious?
You need a modern CPU with a modern instruction set, modern manufacturing process to offset the power consumption, enough cores, because what was enough in 2018 is not enough 2024, not to meniton in a couple of years..

I doubt people buying Athlons, old Celerons/Pentiums or the new Intel Processor line care about performance first and foremost. They want a cheap system. It's for basic office work, schoolwork, maybe a child. Perhaps a small networking or file server at home. There's no need to be a flashy, performant system that is capable of running the latest and greatest experiences. I still use a Core 2 Quad system as my personal server. It does its job.

 

[mkppo]

No wonder why their revenues are on the decline, and they don't expect better times...

Their revenues are actually up quarter over quarter compared to last year for two quarters now. They also revised their forecast to increase revenues by some 500m this year and are expecting better times. Bit of misinformation there..

 

[Vincero]

I think it's yields, the c clusters are very tiny even with 16 of them in there. Also to fit 192 of them in an Epyc socket probably needs 3mn for size reasons but i'll have to look at Turin again to check if 4nm would've worked.

They've already done zen4c at 5nm, whereas the current AM5 IO die is fabbed on 6nm.
At 5nm the 4c core is 2.5mm², which if they just optically enlarged it (yeah I know, they probably wouldn't just do that) would be roughly 3mm², which is very roughly the same size as one of the infinity fabric logic blocks (GMI3 PHY) - they could probably optimise this very easily. So loose one of the GMI3 blocks and replace it with 1 zen4c core and add another 3 zen4c cores along the top edge.

In theory you end up with a hybrid IO die roughly 10% bigger than the 125mm² normal one. If they really engineered it they could probably make it only a few percent bigger as I'm sure there's extra logic that could easily rearrange.
Obviously the biggest issue would be controlling heat, etc., but they could lock the max clocks of the 4c cores to something non-offensive.

End result: Ryzen 3 base CPU... add a salvaged 4 or 6 core CCD and you could have an enhanced Ryzen 5 lineup and start knocking that ball game of thread counts Intel is pushing using P+E thread counts back into Intel's court seeing as the zen4c cores still do hyper threading whereas the 'E' cores don't... oh and still do AVX-512 (if anybody cares).

Side note: It's crazy to think a whole general purpose CPU core is actually smaller than the VCN or DCN blocks...

 

[theouto]

If these CPUs significantly outperform Intel's offerings, THEY WILL CHARGE ACCORDINGLY. There is precedent for this, namely Zen 3 and specifically the $300 Ryzen 5 5600X before the i5-12400 dethroned it as the midrange king for practically half the price.

Well, yeah, that's what a business does, but having the gpu market be replicated would suuuuuuck, it's in a healthy spot at the moment, just hold the pricing the same.

 

[Punkenjoy]

Just to clarify one point:

The difference of the power usage between AMD and intel for the work done in the I/O die is probably pretty minimal compared to the execution cores and all other thing. The handicap on AMD side is not the separate I/O die itself, but the fact that data have to transit via the infinity fabrics to the I/O cores. The distance is longer and it require more power to do so.

Having it in your die mean you can transfer data more easily with less distance and also you might not need to use a technology like infinity fabrics that SerDes witch also use a portion of the power.

 

[Vincero]

Just to clarify one point:

The difference of the power usage between AMD and intel for the work done in the I/O die is probably pretty minimal compared to the execution cores and all other thing. The handicap on AMD side is not the separate I/O die itself, but the fact that data have to transit via the infinity fabrics to the I/O cores. The distance is longer and it require more power to do so.

Having it in your die mean you can transfer data more easily with less distance and also you might not need to use a technology like infinity fabrics that SerDes witch also use a portion of the power.

Well yeah, that duplication of logic on each side of the link (IO die & CCD) is a not insignificant penalty, especially as it is essentially always active and performance focussed. If monolithic designs have say 5W of 'uncore' load, the chiplet approach must add a couple at least, increasing per chiplet.

 

[acekombatkiwi1]

It's literally written under the image?

https://imgur.com/cEfAFn4

I love TPU, only needed to go 3 questions before mine had already been asked.

A bit disappointed that they didn't look to improve temps more than just reporting the correct temp, considering it was a main discussion point on 7000 series. I know they need to keep the overall z height the same for compatibility, but could have increased the substrate thickness to get a thinner IHS.

Maybe with X3D.

 

[Shtb]

A bit disappointed that they didn't look to improve temps more than just reporting the correct temp, considering it was a main discussion point on 7000 series

AMD says it has optimised the thermal material between silicon and heatspreader for this generation. This is coupled to a tweaked layout which minimises localised hotspots that had previously caused overall temperatures to soar on Zen 4.

AMD Ryzen 9000 CPUs run up to 7°C cooler than the current generation | Club386

AMD cites a significant decrease in temperature for the same TDP. Welcome news for enthusiasts.

www.club386.com

 

[RaceT3ch]

I hope AMD doesn't plan on releasing B840, only PCIE 3.0 despite A620 having 4.0? What the hell AMD?

 

[Steevo]

The eTVB issue is separate and is not why the chips are dying... but no, AMD idle power has been several times that of Intel because they are monolithic chips. This is the same reason the APUs have lower idle power, no IOD to constantly draw power. Zen 5 is great, but this is clearly not something it has addressed thus far. AMD likely cannot address it, at least not with a cost-effective solution yet.

People keep saying it's higher, and it's not.

New Intel chips are going to have 5 chiplets what will be your excuse then?

 

[mahirzukic2]

Please AMD, don't fumble the pricing.
It's looking nice, but getting cocky with bad pricing would be horrible.

Like they did the last time. Lost a lot of would be customers. Lost me for example. I really wanted an AMD system, but had to look for value option.

 

[Dr. Dro]

People keep saying it's higher, and it's not.

New Intel chips are going to have 5 chiplets what will be your excuse then?

Is this package power or has this reviewer excluded cIOD consumption? It should be a static 10-15W for cIOD alone. Intel can idle down to a watt.

 

[W1zzard]

AMD Ryzen 9000 CPUs run up to 7°C cooler than the current generation | Club386

AMD cites a significant decrease in temperature for the same TDP. Welcome news for enthusiasts.

www.club386.com

I just went through my audio recording again, in the Granite Ridge breakout session they said that there is no change to the IHS and TIM and continued that the temps are lower due to better floorplanning to avoid hotspots and better placement of the thermal sensor. Will ping Tarinder @Club386 (we're friends) and ask if he got any other info from AMD

 

[ARF]

If they put the thermal sensor in another place, it will mislead the readings. Of course, that the silicon can withstand much higher temperatures than 95°C, but still... it is not an apples-to-apples comparison..

 

[W1zzard]

If they put the thermal sensor in another place, it will mislead the readings. Of course, that the silicon can withstand much higher temperatures than 95°C, but still... it is not an apples-to-apples comparison..

It never is. You can't put the sensor in the hottest part, because then the sensor circuitry would take up so much space that the temps won't reach maximum in that area.

What's done is to put several sensors in less "busy" locations and interpolate the temperatures, of course including "enough" safety margin

 

[Vincero]

People keep saying it's higher, and it's not.

New Intel chips are going to have 5 chiplets what will be your excuse then?

That graph looks great until you consider that Intel are at least 1 process node behind...

Ironically looking at previous APU vs CPU tests where most things are equal (e.g. the 5700G, although yes they have different cache config and PCIe performance - had to go back to zen3 as the zen4 CPU and APU chips are very different so harder to compare), whilst the idle numbers are similar, the low-load single-threaded and multi-thread performance numbers show the APU 8-core having consumption lower than the 8-core CPU and sometimes lower than 6-core CPU - all the while it's also powering a much bigger graphics block than on the AM5 IO die (and it was Vega so not exactly super efficient).

Looking back, credit where it's due, when the 14nm+++++++ process wasn't being clocked to hell, it may have been old but the Skylake>Rocket Lake designs were efficient at using it, e.g. that 9900x idle and single-thread power use, and the 11400 trading blows with 3600 despite being 14nm vs 7nmm, and invariably being the faster part...

 

[Panther_Seraphin]

Surprising there's still no plans in this generation to bring the 'c' cores into their main desktop product offering (the 8000g series being both last gen zen4 and also more of a laptop part).

I've often thought AMD could kill two birds with one stone by doing an alternate IO die configuration loosing one of the chiplet infinity fabric channels on the IO die and instead splicing on some zen4c/5c cores - sure they'll end up with a slightly bigger IO die (likely a few mm² bigger) but they could then offer lower cost quad-core AM5 CPUs using the IO die cores only (and AM5 really needs some cheap low-end parts to exist once AM4 really is dead), and have 4c + 8 core desktop products (and whatever derivative SKUs they'd want).
Admittedly the quad-core AM5 using just 'c' cores would be an odd product outside of just being a value proposition - tons of IO capability compared to the 8000g series but not a huge amount of CPU power to exploit it, but I could see home NAS / niche appliance / embedded usage scenarios where the additional 8000g Radeon CUs wouldn't be utilised and a smaller essentially 'monolithic' die with better IO options would be desireable.

I reckon there's a large amount of bare minimum office PCs being sold that utilise new Intel Core-i3 chips instead of AM4 4000/5000g CPUs due to AM4 being seen as older platform, the newer i3's having better IPC and (even though Intel's process tech is behind) comparatively acceptable power usage. Eventually that existing AM4 inventory is going to end or not be used for most new sales.

Obviously they can still make the normal IO die for the 16 core CPU configurations they currently push out - not suggesting they abandon that.

Problem is now your spinning a completely new IO die for a specific segment of the market

Currently this is why their chiplet approach works as they can mix and match parts to suit from bare office PCs all the way up to highest end server parts with very little changing between them. For AMD this is the best in terms of mass productions benefits. I believe for their whole current offerings its 2 IO dies, 2 CCD types and 1 Monolithic die to cover all these?

Instead of what you are saying in regards to reworking of the IO Die. I suspect there will be a repurpose of some of their mobile offerings to fill the low end office pcs, especially with the explosion of USFF PCs becoming more and more common in office spaces.