Finally. That shall have been done long time ago.
The Intel Execution in [2024]
- Thread starter Shifty Geezer
- Start date
Interesting.
So rather than "fixing" the AVX512 power consumption challenge, they just decided to scrap it? Someone help me out because I'm not knowledgeable in this space: how often do AVX512 instructions show up in regular, consumer-grade software? I feel like it's probably a very low, bordering on tiny percentage... Anyway, to @Scott_Arm 's point, AMD figured out how to get AVX512 "working" in their CPUs, so why not Intel? I also agree this means yet another reason they could make sense for the next generation of console CPU...
So rather than "fixing" the AVX512 power consumption challenge, they just decided to scrap it? Someone help me out because I'm not knowledgeable in this space: how often do AVX512 instructions show up in regular, consumer-grade software? I feel like it's probably a very low, bordering on tiny percentage... Anyway, to @Scott_Arm 's point, AMD figured out how to get AVX512 "working" in their CPUs, so why not Intel? I also agree this means yet another reason they could make sense for the next generation of console CPU...
DegustatoR
Legend
What makes you think that they've decided to scrap AVX512?
This seems to me more about getting AMD and the rest on their side with implementing x86s.
Guessing you didn't read the posts about five up from mine? Since it looks like you haven't heard the news, Intel ditched AVX512 starting in Alder Lake (the 12th-gen series) and in all consumer processors since then. My reply was specifically in the context of AVX512 having gone missing on the Intel side (some theories suggested it might come back, seems those were wrong) yet AMD continues to provide it.
DegustatoR
Legend
Oh, sorry, I thought this was about todays news which are right above your post.
I know. They haven't really ditched AVX512 though even in these, it's more that they didn't add it to the E-cores and are thus forced to disable it on the P-cores too. IIRC the same P-cores in DC parts support AVX512 just fine.
My guess here would be that they will add it to E-cores eventually - as APX10 most likely or whatever was that name for a common set of all AVX512 instructions - when they will be less affected by comparatively high power draw and the die size implications.
DavidGraham
Veteran
This is one example among many of the kind of mentality that ruled over Intel, which made them lose many opportunities (iPhones, mobile, dGPUs, GPGPUs, AI, etc).
Even if they had "own internal version of x86-64" x86-64 was still AMDs tech which Microsofts commitment "standardised"
AMD announced the x64 extension initiative in 1999, produced a spec in 2000, presumably sometime around then Intel would have had to begun work, AMD's first x64 chip comes out in April 2003, Intel's Prescott (with x64 disabled) was February 2004. According to the Dave Cutler interview the Windows x64 initiative was really more of a side project that he spearheaded and wasn't a company "commitment" until after it was done and after it started gaining traction internally. Let's not pretend that Intel didn't start moving on this until after MS committed to it. The post was about Intel being subject to internal politics which slowed their move on x64. Interestingly MS had their own version of those politics due to different teams with their own products, own kernels.
Last edited:
DegustatoR
Legend
Intel was hoping to transition in full to IA64 back then.
Yeah, reading through the Pentium Chronicles and the various talks and interviews Colwell has given, it paints a messy picture of Intel. Different teams working on projects that were siloed off from each other, and inevitably you get contentious politics when one excels and another lags. Probably an inevitability for large companies backing multiple horses in different, albeit converging, races. Especially back then when they were in the height of Dennard and Moore's Law scaling -- very hard to place the right bets 5+ years out.
Intel bet on Itanium hard. Ultimately it failed but it succeeded in what Intel and HP (mostly HP) originally set out to do: to kill other RISC server CPUs. Many vendors at the time dropped or significantly reduced their developments on their own RISC CPU because they were too afraid of Itanium. Some were simply pushed out. So business-wise Intel was correct. The problem was that Itanium is simply not a good ISA, and they couldn't bring a good enough product. If Itanium was a good ISA and thus Intel were able to bring up good Itanium CPU we might not be using x86-64 now.
As for why Itanium is not a good ISA, that's another story, but the general idea is that Intel is never really good at designing ISA. The best ISA Intel designed were probably i960 and i860 but both have their own problems (though to be fair all ISA have problems).
As for why Itanium is not a good ISA, that's another story, but the general idea is that Intel is never really good at designing ISA. The best ISA Intel designed were probably i960 and i860 but both have their own problems (though to be fair all ISA have problems).
Arrow Lake is an absolute disaster, going by reviews. Stability and inconsistency issues, lackluster performance in most applications, efficiency is a mixed bag as I'm not sure N3B is really quite as suited for high clock speeds as Intel's more mature nodes can be.
It's so bad that I fear people will take the wrong lesson and start talking about Zen 5 as if it isn't actually the big disappointment is still is. Seems overall, consumers are just getting gut kicked for this generation of CPU's, and with likely no reprieve til 2026.
Will probably take some time before the blame gets properly tracked down for where Arrow Lake went wrong, but I also expect that there's gonna be plenty of pointing to the tile package design and honestly, that's easily gonna be my guess as to a big chunk of the problems here. I've been very critical of Intel's tile strategy for a while now, purely on the basis of unneeded complexity and cost with minimal upside in terms of reuse or scalability. And now if there's further issues relating to getting it to work well with Windows and performance issues, it might well look like one of the worst decisions they've made in a long time. And it was a decision. There was nothing keeping them from simply sticking with simple, monolithic designs other than their own hubris. Or heck, I wouldn't have blamed them at all if they straight up just copied AMD's simple and clever chiplet strategy.
I just have to imagine 8P Lion Cove + 16E Skymont on a monolithic Intel 3 process would have been so much better for Intel all round.
It's so bad that I fear people will take the wrong lesson and start talking about Zen 5 as if it isn't actually the big disappointment is still is. Seems overall, consumers are just getting gut kicked for this generation of CPU's, and with likely no reprieve til 2026.
Will probably take some time before the blame gets properly tracked down for where Arrow Lake went wrong, but I also expect that there's gonna be plenty of pointing to the tile package design and honestly, that's easily gonna be my guess as to a big chunk of the problems here. I've been very critical of Intel's tile strategy for a while now, purely on the basis of unneeded complexity and cost with minimal upside in terms of reuse or scalability. And now if there's further issues relating to getting it to work well with Windows and performance issues, it might well look like one of the worst decisions they've made in a long time. And it was a decision. There was nothing keeping them from simply sticking with simple, monolithic designs other than their own hubris. Or heck, I wouldn't have blamed them at all if they straight up just copied AMD's simple and clever chiplet strategy.
I just have to imagine 8P Lion Cove + 16E Skymont on a monolithic Intel 3 process would have been so much better for Intel all round.
The Core Ulta performance issues can't be fixed with drivers. In fact the APO in some cases even worsens the performance
What makes you concretely confident of that? Also, I didn't mention drivers. I was more referring to the applications themselves needing changes.
Of course we shouldn't be wondering about any of this as Intel should have solved all of this before launch.
DegustatoR
Legend
I'm pretty sure that a whole bunch of ARL performance issues can be fixed with BIOS and Windows updates. This should be enough to put it at parity with RPL-R and Zen 5 - which is an okay result considering that it also shows 2X perf/watt improvement over RPL-R and is about on par with Zen 5 in this.
Whoa. I haven’t seen anything to indicate 2x perf/watt over raptor lake. What’s the source for that?
DegustatoR
Legend
Intel Core Ultra 9 285K, 7 265K & 5 245K vs. AMD Ryzen im Test: Gaming-Leistung
Intel Core Ultra 200S im Test: Gaming-Leistung / Leistungsrating FPS und Frametimes / Core Ultra mag RAM-OC / Anno 1800
www.computerbase.de
They also have a summary graph there.
Fully threaded (production) workloads seem to fare a bit worse but there is still a visible perf/watt gain.
