Zen 5 chips now sampling

[Redaktion]

After unveiling surprisingly strong data center sales in the first quarter, AMD's CEO, Lisa Su, also mentioned that the first Epyc Turin and Ryzen Strix samples are already reaching the company's partners. These Zen 5-based processors are expected to launch in the second half of 2024.

https://www.notebookcheck.net/Zen-5-chips-now-sampling.833031.0.html

[Brorin]

Holy cow. Zen 5 is coming. And here I am, still using a 6th gen Intel Core chip

[MQ.Yang]

Zen5 is a design with big differences compared to Zen4, it will finally be worth jumping to DDR5.

[NikoB]

DDR5 became obsolete many years ago. x86 cores are suffocated by slow RAM.

Look at the operating speeds of the L1-L3 cache - it is an order of magnitude faster and the x86 cores work perfectly at this speed.

Just before the advent of 8K monitors (only the Strix Halo line, let me remind you), AMD was forced to increase the data bus width of the Zen5 memory controller to 256 bits in order to at least increase the bandwidth to 120-150GB/s.

That is why they delayed introducing Display Port 2.0 for 5 years since 2019, when the new video data transmission standard was published. Apparently, even the fast memory of discrete video cards, which has long met the requirements of 8K resolution in 2D (no talk about 3D), could not ensure smooth operation with 8K content in the presence of extremely slow RAM on x86. For this reason, both AMD and Nvidia also delayed introducing full DP2.0+/UHBR20 ports for so many years. What's the point if everything again comes down to the slow system RAM, which has become a bottleneck?

Even C2D 15 years ago was capable of operating at speeds of more than 150GB/s at the L1 cache level, i.e. If it had 150GB/s system memory instead of an L1 cache, it would clearly work many times faster; processing such a stream would not be a problem for the cores.

Servers switched to HBM(3e) memory long ago, realizing the dead end of DDR5 memory.

We need 512+ bit memory buses on x86, and not the top variants - but mass-produced mid-level motherboards. But in the end, there will be an increase to 256bit only in the top-end Strix Halo series, almost 20 years after the appearance of the 128-bit controller in x86.

x86 clearly got stuck in the quagmire of slow memory controllers many years ago. Primarily due to the fault of Intel, which rested on its laurels until 2016.

It should be added that mass deliveries of Zen5 will begin no earlier than mid-2025, but rather in the fall of 2026. This is another 2 years. And Strix Halo will obviously be in great short supply again, because... AMD is increasingly moving into the server market, where profits are now much higher...

[Neenyah]

Can't wait 2029 to see one in stock and available to purchase immediately.

[A]

Servers switched to HBM(3e) memory long ago, realizing the dead end of DDR5 memory.

We need 512+ bit memory buses on x86, and not the top variants - but mass-produced mid-level motherboards. But in the end, there will be an increase to 256bit only in the top-end Strix Halo series, almost 20 years after the appearance of the 128-bit controller in x86.

x86 clearly got stuck in the quagmire of slow memory controllers many years ago. Primarily due to the fault of Intel, which rested on its laurels until 2016.

The reason there wasn't much increase was because the cost to performance ratio isn't worth it for most processor tasks. Thus it is mostly used as VRAM for GPUs who do benefit from it. Only reason you are likely seeing increases now is due to AI

[NikoB]

Light operation of 1-3 8k monitors in 2D alone justified a bandwidth of at least 150GB/s many years ago.

This alone was a good reason - an ideal 2D picture in terms of clarity of text and graphics on diagonals 24-32", which we are sadly waiting for the second decade. 8k monitors were ready for mass production 10 years ago, but the speed of RAM did not allow them to run smoothly work without friezes.

AI is just an absolutely fake and meaningless topic at the levels of performance, volumes of RAM and disk space that exist now and will be in the near future. Neural networks, even with 1TB of RAM, have virtually no real use for ordinary people, except for extremely simple tasks. What is truly complex requires hundreds of terabytes, petabytes of RAM and disk space (and even with such volumes available in large data centers, modern models are in fact primitive), which I wrote about many years ago, including here. I have already read the opinions of people directly working with AI, which confirm my point of view. Approximately since 2015, I wrote on a number of specialized forums about the inevitable collapse of the idea of a real "autopilot" in the next at least 20-25 years. This is what happened in practice.

At the current level of technical processors and computing capabilities, there is no practical benefit from neural networks on mass-produced PCs - there is simply no room for deployment of them. This is another marketing ploy by corporations to increase sales and raise prices even higher. The same thing happened with the widespread dominance of "nanotechnology". Who remembers them now? Every company from a Chinese basement always wrote on their garbage product that they use "nanotechnology." The herd was buying. It worked. It will work now with the "AI" fiction.

But those in power are receiving increasingly powerful, actually working tools for analyzing and making specialized decisions (and in real time) for managing and manipulating the crowd at the level of data centers and collecting big data, both official and openly criminal, about the population. Technologies that are completely inaccessible to ordinary people, but only increasing the gap between the powerful stratum and ordinary citizens... this is an increasingly sad growing trend. The frogs are cooked gradually and carefully (quietly eliminating any human rights activists along the way) so that they do not panic and jump out of the concentration camps being created. The world of neo-feudal lords is coming in all its glory.

Now the "right-wing" FoxNews is posting "opinions" that persuade the public and politicians not to limit the development of "AI" in large data centers. Everything is for your benefit. Well, yes, well, yes..."uniparty" in all its glory.

[A]

Light operation of 1-3 8k monitors in 2D alone justified a bandwidth of at least 150GB/s many years ago.

This alone was a good reason - an ideal 2D picture in terms of clarity of text and graphics on diagonals 24-32", which we are sadly waiting for the second decade. 8k monitors were ready for mass production 10 years ago, but the speed of RAM did not allow them to run smoothly work without friezes.

The fact that TVs did not go 8k says otherwise, most monitors aren't even 1440p yet nor 4k (pretty much most tvs are already 4k)

Unless your face is sticking in the monitor, by average viewing distance even 32" 4k should be plenty

AI is just an absolutely fake and meaningless topic at the levels of performance, volumes of RAM and disk space that exist now and will be in the near future. Neural networks, even with 1TB of RAM, have virtually no real use for ordinary people, except for extremely simple tasks. What is truly complex requires hundreds of terabytes, petabytes of RAM and disk space (and even with such volumes available in large data centers, modern models are in fact primitive), which I wrote about many years ago, including here. I have already read the opinions of people directly working with AI, which confirm my point of view. Approximately since 2015, I wrote on a number of specialized forums about the inevitable collapse of the idea of a real "autopilot" in the next at least 20-25 years. This is what happened in practice.

Most people don't need a self aware AI, even the ML that we have is more than enough for what most people need. Which is tools to assist them like help grammar, organize/search for basic data, create fan art for the desktop/avatar for a forum, code assisting, better voice recognition, better OCR and translations and etc. You don't need 1TB of ram for that, even 8gb is enough

[NikoB]

That's why there is no 8K, because for 5 years now we have been dynamized with the interface for connecting them (despite the fact that even DP2.1 is already morally outdated, even for smooth scrolling of text in browsers, since 80Gbps is not enough for at least 120Hz frame frequency, it is necessary 160Gbps, and this further increases the requirements for RAM speed).

4k TVs mostly operate in 4:2:0 mode; rarely do any models support 4:4:4 monitor mode. And all 8k TV today is in fact fake 8k, because... HDMI 2.1 is not able to transmit even a regular 24-bit video signal at 4:4:4@60Hz without lossy compression (DSC), much less at frame rates above 60Hz. Not to mention support for lossless HDR10 (30bit) 4:4:4@60Hz, which requires more than 70Gbps.

Only computer monitors make sense in 8K, because... 8k cinema really makes little sense on diagonals smaller than 200" at home. But perfectly clear text in the near zone is required even on a 22" monitor, which cannot even provide 4k, because at 16:10 there is only 205 ppi, but you need at least 220-230 and preferably close to 300, which gives only 6k on such a diagonal, but 6k is not a multiple of 4k and fhd, therefore it is technically inconvenient (although Dell released exactly this one earlier, pointing out that professionals just like ppi in the region of at least 220-230+).

With the advent of 8k monitors, the issue of cloudy fonts and visible pixelation of the picture even in the near zone of 30-35 cm will be practically closed (the eyes will no longer spontaneously refocus from objects on the screen to their pixel structure, which will sharply reduce fatigue in people with good vision), t .To. even at 32" 16:10 we get at least 283 ppi. It is unlikely that anyone will need 16k in practice (as a multiple of 8k, 4k and fhd), even if in the future this becomes easily achievable technically.

Further increases in pixel density on monitors become impractical. But it is important in VR in matrices with a large viewing angle, but this is a completely different topic, where the technologies are still in their infancy in terms of user comfort. Which was the reason (as I predicted on this forum at the time Apple released its VR helmets) of their next failure.

Most people are so technically illiterate that they don't understand that 8k is an almost perfect picture with no visible pixelation even close to it. This is exactly the permission that absolutely everyone needs to preserve their vision and minimize fatigue.

[JUAN_pcbox]

AMD Zen 5 is a gigantic leap compared to Zen 4, 2 years have passed since Zen 4 and they have improved their entire microarchitecture.
Zen 5 will have improvements in instructions per clock (IPC) of up to 30%. If that turns out to be true, it would be the largest intergenerational IPC increase since the debut of AMD's original Zen generation in 2017. Zen 5 may add support for FP16 AVX512 instructions. It can also dramatically increase CPU cache across the board, including L1, L2, and L3. and for gamers new RDNA 3.5 graphic architecture
And for me the most important thing about AMD Anyone who bought a high-end Ryzen 1000 generation motherboard was able to simply plug in an AMD Ryzen 3000 or even 5000 CPU with some BIO updates and continue with the same memory, power and everything. Otherwise, that is, the same board for almost 5 years and while Intel only offers 2 years, which I find embarrassing.
And for this year the star product and already a best-seller in the US is MINISFORUM V3 for those who want it with mobility, performance, power and minimal consumption. I look forward to seeing a comparison with the future Surface 10 on this website.