So they cut down the depth of several structures (I'm guessing at the backend/execution engine), and core area still increased by 26%, while Fmax is the same on a buffed node? How much larger is the front end then? Also, if the rumor about the dual uop cache is true, I'm guessing that ate up a lot of the transistor/area budget...View attachment 101536
The bean counters are counting the mm2 at AMD.
Then there are lamentations and the weeping and gnashing of teeth for thenonbelievers of the bigly IPC gainz
TBF LNC looks way less radical than Zen 5.About the same
After the past 3 years It does feel like there’s sort of a hard limit being hit.About the same
I think they are both equally disappointing. Still, the discrepancy between what AMD explicitly told their partners about the SpecINT aka the fabled 32% figure and the figure they gave us, lingers. How can such a radical redesign only yield such a mediocre performance is beyond me.TBF LNC looks way less radical than Zen 5.
AMD internally only said 10-15%. That is also what they communicated with partners.I think they are both equally disappointing. Still, the discrepancy between what AMD explicitly told their partners about the SpecINT aka the fabled 32% figure and the figure they gave us, lingers. How can such a radical redesign only yield such a mediocre performance is beyond me.
Past behaviors is best predictors for the future and the future doesn't looking promising unfortunately. Diminishing returns I guess.Hopefully next generation gets back to upper teens for IPC improvements.
They actually said 10-15%+. The 32% figures and the likes were given to hyperscalers.AMD internally only said 10-15%. That is also what they communicated with partners.
So is the 32/40% figure correct for Zen5 CPUs used for hyperscaling? In that case, how do those CPUs differ compared to the ones used for DT, which can explain the huge performance increase difference between DT vs hyperscaling?They actually said 10-15%+. The 32% figures and the likes were given to hyperscalers.
I would run a youtube channel and a patreon account in parallel if I know such deeds.So is the 32/40% figure correct for Zen5 CPUs used for hyperscaling? In that case, how do those CPUs differ compared to the ones used for DT, which can explain the huge performance increase difference between DT vs hyperscaling?
Things come to mind are (Zen4 Hyper-scale vs Zen5 Hyper-scale)So is the 32/40% figure correct for Zen5 CPUs used for hyperscaling? In that case, how do those CPUs differ compared to the ones used for DT, which can explain the huge performance increase difference between DT vs hyperscaling?
What's the context?So is the 32/40% figure correct for Zen5 CPUs used for hyperscaling? In that case, how do those CPUs differ compared to the ones used for DT, which can explain the huge performance increase difference between DT vs hyperscaling?
It's specint socket rate.If it's 32% performance improvement at same power budget or TDP, process improvement is factored in.
If it's 32% ipc improvement, workload might be FP or vector biased.
To me it's just good old AMD coming back. People who lived through the hype cycles of K9, Barcelona, Sandtiger, Bulldozer (and, of course, Tunnelborer), or K12 know this.if true, Apple got an avg improvement of 7% in 7 months and achieved more than AMD did in 22 months. This makes M4 look very good.
This is embarrassing. Like I think it’s fair to say AMD lost talent. I mean it looks like it. The cadence slowed down as well after Zen 3.
That's 10 years like dead.To me it's just good old AMD coming back
Who is this guy? Is he credible and in a position to know?
I’m guessing the numbers provided to OEMs were the 10-15% range we’ve seen before.
The floating-point performance of the desktop should be slightly stronger. All Strix Point cores cut the SIMD throughput in half, and even damaged some AVX/SSE instructions, making it inferior to Zen4.
He said it was multiple tests, I’m sure Spec is one of them though. We’ll find out how it does in a month or so from other sources. My guess is ~10% int and ~15% fp.iirc he did IPC/ST test for many CPUs include Zen4 and I saved it with luck. It showed Zen3 to Zen4 SPECINT has only 4% uplift. But yeah it's still on par with Goldencove. Also bigger L3 cache/V-cache variant has higher IPC.
I'm not calling him or SPEC unreliable but at least it's either SPEC as industrial standard is very stringent, or we should be very careful when we see SPEC results no matter from thirdparty or official especially when comes to compiler concern.
OTOH he also mention Strixpoint SIMD unit being serious cut down compared to GraniteRidge.
View attachment 101570
The top configuration 5.1GHz should be just over 3000, and the 5GHz I measured was just 3000. The performance improvement of GB integers is greater than SPEC.
That figures. I think he usually does his testing in Linux. I expected a range of 2700-3000 for Strix on Windows OS depending on SKU and the chassis it is in. That 2800 score from last week was probably a real run just in suboptimal conditions with a bunch of background processes running.
Some more info from him regarding Geekbench 6:
AI AI AI
From David Huang, for Strix, seems they cut the FP a lot, including reducing AVX/SSE throughput.So they cut down the depth of several structures (I'm guessing at the backend/execution engine), and core area still increased by 26%, while Fmax is the same on a buffed node? How much larger is the front end then? Also, if the rumor about the dual uop cache is true, I'm guessing that ate up a lot of the transistor/area budget...
Some more info from him regarding Geekbench 6:
if true, Apple got an avg improvement of 7% in 7 months and achieved more than AMD did in 22 months. This makes M4 look very good.
This is embarrassing. Like I think it’s fair to say AMD lost talent. I mean it looks like it. The cadence slowed down as well after Zen 3.
Yeah, looks like the minds behind Zen1 and Zen 3 are now working in Tenstorrent, like someone was suspecting a few pages ago.if true, Apple got an avg improvement of 7% in 7 months and achieved more than AMD did in 22 months. This makes M4 look very good.
This is embarrassing. Like I think it’s fair to say AMD lost talent. I mean it looks like it. The cadence slowed down as well after Zen 3.
This is a reductionist take. Company's performance is tied to abilities of engineers working in it, and they come and go. Saying that the current issues are the same as those of 15 or more years ago is just silly, as it's the same company in name only.To me it's just good old AMD coming back. People who lived through the hype cycles of K9, Barcelona, Sandtiger, Bulldozer (and, of course, Tunnelborer), or K12 know this.
Someone clone Keller and GWIII.So there is a shortage of talent in the CPU industry.