Exactly, "the updated interconnect" that was available for some years now, and AMD chose cheaper interconnect tech because it was "good enough", as said, with Zen 6 they're finally paying up for the better tech.
It's also possible that dies for new consoles will be on chiplets with new IF. We don't know this at the moment, as no leaker has been able to penetrate this level of information. Such die could be similar to Strix Halo philosophy, but with one 12-core CCD and bigger IOD/iGPU.
They usually develop new interconnect for new platforms that support higher standards. In this case, they are preparing IF to support several segments: EPYC Venice with PCIe 6.0, future Medusa Halo, and possibly future consoles. The testing vehicle for new IF is Strix Halo. On Strix Halo, IF area is more narrow, denser and distributed, as shown in the picture. We know this because CCD is shorter and a tad smaller than vanilla Zen5 CCD. They have not disclosed details, but it has most likely to do with denser fan-out links and TSVs.
EPYC-wise, the IF is usually co-developed with whatever PCIe standard is supported on a new gen platform. IF is usually equal or faster than supported PCIe. On Zen5 CCD, IF speed is 36 Gbps per single link, whereas PCIe 5.0 is 32 Gbps. On Zen6, IF single link will have ~64 Gbps, if not more, and PCIe 6.0 on Venice platform will also bring 64 Gbps per lane. The speed of IFOP is similar to xGMI on IFIS. Plus, specific EPYC SKUs allow GMI-Wide up to 72 Gbps, with two x16 IF links to each CCD, for specific workloads and data center clients. Such new IF, 'GMI4', could feature in new consoles if Sony and Microsoft are happy that monolithic dies are not necessary. Again, we don't know this at the moment.
There was no need to have faster and more expensive IF before, as chips are very competitive in current market with what they have. Less is more. Besides, cost would increase unnecessarily. And more complaints that final products are even more expensive. So, do you want a good and affordable product in its class or prohibitively expensive product with diminishing returns? You can't have it both ways.
Note the differences between the expensive packaging used for the Strix halo (seen in hard to find $2000+ devices with zero upgradability due to everything being soldered), similar in appearance to something like an ARL or Apple M series chip with it's tiles placed as close to each other as possible.
This is nonsense. Note that there is twice as much to pay for similar Apple system. Note that Intel does not offer such silicon at all. Does Nvidia offer anything in this segment at this price point? No. Qualcomm? No. Anyone else? No. And somehow I hear you complaining. One can literally order now the top SKU 395 for $1,500 from GMK. Any better offer from anybody else with similar silicon?
SSDs and WiFi chip can be upgraded, even more modules can be upgraded on Framework Strix Halo, including different IO ports. I personally do not enjoy soldered platforms, but this is in comparison with Apple, who will cut your clicking hand off by $1,000 if you dare to click on 16GB of extra RAM module and extra 2TB SSD, from their store only, it should be said. It is Apple's platforms that are 100% not upgradeable by users who just want to instal standard components available in open market, just like they do in DIY desktop space. Apple is the most anti-consumer company in terms of selling closed platforms. There is no doubt about this. It's not surprising that their laptop segment has been stagnating in recent years, as buyers are increasingly tech savvy and want to have more options to exchange components on their own rather than pay a fortune and lose their clicking fingers in exchange for propriatery parts.
M chips with tiles? Are you talking about Ultra chip? Apple chips are monolithic. There are no tiles.
And the cheap packaging used for the 9955X3D, with lower density interconnects forcing further physical separation, leading to the latency issues X3D tries to mitigate and the IF drawing idle power that cannot be lowered to what competing monolithic chips can achieve
You act as if you re-discovered America yesterday. Monolithic chips, of course, have their advantages in terms of latency and idle power, but also drawbacks, such as lower yields as die size increases. Silicon production is a balancing game of economics and technological viability. There is no perfect solution. Trade-offs simply need to be good enough for a product to be successful. Entire industry is slowly moving towards chiplets in increasing number of segments, and AMD has championed chiplets to their great success. Even Nvidia will have a chiplet-based GPU die on Vera Rubin Ultra.
You should look more closely into physically close tiles on Arrow Lake and ask questions about latency that has had a catastrophic effect on gaming prformance. Physical distance between dies is not everything. Also, I will take idle power any time than 275HX CPU that almost melts laptops, being allowed to run at 99 degrees, guzzling over 200W at CPU socket power.
As regards to 9955X3D, here is almost 5 hour session below testing the same monster gaming laptops, one with a 'terrible' CPU 9955X3D and another one with 275HX that...appears to have its tiles, quoting you: "placed as close to each other as possible". I'd recommend you to watch it. It's very informative. Ryzen is on average 16% faster in 35 games while using less power; 3% faster in GPU-bound games and whopping 55% faster in CPU-bound games.
https: //www.youtube.com/watch?v=kB1APliFiYA
It amazes me that more advanced CPU on paper, Arrow Lake 275HX, is not able to beat Ryzen in a halo laptop, and how AMD still manages to win here. Intel folks must be furious that they paid so much for this architecture and cannot defeat AMD, even when the CPU guzzles 200W. Astonishing! There are questions to be asked here about power-to-performance scaling, and idle power is the least important one in this laptop segment.
Zooming out, AMD's margins on revenues are 54% (Intel 34% only), they produce client chips on much cheaper N4 node (Intel pays a fortune for N3), and somehow Intel is still not able to beat them, not only in halo laptops, but across segments and with tiles as close to each other as possible, to use your words. Something stinks in the kingdom of Denmark... The new CEO now mandates minimum 50% margins on any future architrectures, as they have been recklessly runninng the company in an unsustainable way for several years in order to come across as competitve at any price. It cannot work like this in a long run. Check it out. Your colleagues wrote about it.
Intel's tale of financial difficulties has been told for many quarters now, and the company is slowly paving the way to profitability through workforce reduction, new aggressive product roadmaps, and, as of now, a 50% gross margin requirement before entering production. At Bank of America's...
www.techpowerup.com
In addition, much faster interconnect fabric on Arrow Lake somehow manages to have latency too high, impacting gaming and throttling Gen5 SSD speed by 2GB/s. If I was Sony or Microsoft, putting aside compatibility debate, I would not have confidence to give Intel contract at this junction to develop my new consoles. No way, by looking at issues they have on client products.
Such issues are making hundreds of thousands of buyers switching to AMD. This trend is clearly visible in any quarterly numbers. Intel have consistently been losing market share in last 7 years. Slowly, but consistently going down. Their Xeons are now two generations behind in average performance. They have nothing to offer in console segment as their iGPU does not scale enough beyond smaller designs for laptops and handhelds. Arrow Lake mega-APU project was cancelled a few years ago. They can barely release a desktop GPU in small quantities. Nvidia has nothing to offer for big consoles either. Why would they bother if they earn tens of billions in data center? It's not surprising that Sony and Microsoft do not work with them on development of new consoles.
