Leaker channels, MLiD, Coreteks etc. hype releases up too much pre launch. I guess it gets views and buys them beer. They then go on to say, after launch.. “didn’t meet expectations and the fault is”…cue another leak.. more views, more beer.
Not only that, but they almost
need to find someone to "blame" for it falling short of expectations, so that people forget it was
them who helped hype up those expectations.
I wish there was some kind of blowback for these folks, but they're operating in a zone largely consequence-free zone and with the incentives heavily weighted towards being sensationalist. Perhaps the best we can hope is that people eventually learn to take them with a grain of salt and stop paying so much attention to them.
Intel was milking the same basic design, improved iteratively, up to the 9900k then a change in design philosophy was implemented, P cores and E cores. This saw fruition with the 12th gen.
AMD have been milking Zen and improving it iteratively but the core design must be well within the area of diminishing returns regarding improvements.
I’m not criticising either company there, the big gains are made early in the lifetime of a particular concept, as the low hanging fruit is knocked off gains become harder, more obscure.. then comes revolution and a new concept.
Wow, that's a thoughtful perspective.
What I see happening is that AMD has been dipping its toe into the hybrid realm, using its C cores. They're not as ambitious as Intel's approach, but given how Zen is already more biased towards area-efficiency, I think they don't need to be as extreme as Intel's E-cores. Indeed, Strix Point does look quite good, compared to Meteor Lake. Both with similar thread counts and process nodes.
Meanwhile, Intel is trying to paper over some scheduling hazards of their hybrid architecture by trying to narrow the gap between P-cores and E-cores, taking them in a direction closer to AMD's C-cores (performance-wise, not architecturally). If this actually
hurts perf/area and perf/W, as I expect, then I think this situation is unfortunate.
The real problem is software, but that's such a slow-moving behemoth that I don't see either company even
trying to alter its trajectory. To be clear, what I mean is that software uses somewhat antiquated constructs to employ multiple cores. We need some real investment in OS and API technologies, here. Threads need to stop being the basic building block for multi-core scaling.
what was there with the 3000 to 7000 Zen chips regarding latency was good, perhaps with some new instructions or configuration the inter CCD latency was seen to cause an instability? Increase latency to mitigate ?
I've heard inter-CCD latency increased due to more aggressive core parking. As I think about it, I'm not sure if that theory holds water, given how I believe those benchmarks work. Indeed, ChipsAndCheese doesn't give that argument any play. In fact, they don't even speculate why:
AMD’s desktop Zen 5 products, codenamed Granite Ridge, are the latest in the company’s line of high performance consumer offerings.
chipsandcheese.com
Perhaps we're overthinking this and it was really just something like a chip bug they had to mitigate via relaxed timing.
FWIW, I also think the effect of core-to-core latency seems to be overplayed. It'd be interesting to look for scaling discontinuities in multithreaded benchmarks vs. single- & dual- CCD CPUs. See how often and severely they really occur. Perhaps people seizing on the inter-CCD latency problem are just grasping at straws.
Facebook
Twitter
Instagram