AMD Announces Ryzen 7000 Series "Zen 4" Desktop CPUs - Linux Benchmarks To Come

[ms178]

This says Zen 4 will support everything Rocket Lake implements + BF16.

https://en.wikipedia.org/wiki/AVX-512#CPUs_with_AVX-512

Zen 5 will probably increase throughput and decrease latency, if it avoids breaking each instruction into two 256-bit chunks.

Thanks for the hint, that means VP2INTERSECT and FP16 are missing (which are supported unofficially in Alder Lake and officially in Sapphire Rapids). While the former doesn't seem too important (due to https://arxiv.org/abs/2112.06342), the latter could provide a nice performance boost in some areas.

[ms178]

Do we know for sure Zen 4's AVX-512 is using the 2x256 method?

It came straight from Mark Papermaster's mouth in the presentation yesterday. It can't get more official than that.

[numacross]

[...]
which are supported unofficially in Alder Lake
[...]

Unfortunately only early revisions of Alder Lake support AVX-512. At first Intel disabled it via microcode, but in the end started to fuse it off physically.

It came straight from Mark Papermaster's mouth in the presentation yesterday. It can't get more official than that.

Fair enough, thanks

[WannaBeOCer]

Zen 4’s IPC is 2% faster than Alder Lake. Intel announced Raptor Lake will have 15% faster IPC than Alder Lake and ~40% multi-thread increase performance compared to Alder Lake. Looks like I’ll be going with Raptor Lake.

[numacross]

As Birdie mentioned, ECC is supported with the entire range of Alder Lake, if you simply use a motherboard with the W680 chipset. This is an unprecedented move*, for Intel.

After looking into this further it turns out they have not enabled ECC with the entire range, but instead segmented it (yet again). For example i5-12500 supports ECC, but i5-12400 does not. The list on Wikipedia has ECC parts in bold, and it looks that only 40% of desktop SKUs and ~9% of mobile ones have it enabled.

[coder]

I was about to post this, but it's a claim without a source. I was not able to find anything useful on AMD's official site, it's probably still under NDA.

If you look above the table, the Zen 4 entry cites several sources, including a TechPowerUp interview with Robert Hallock, in which he specifies VNNI and BF16 will be supported. BF16 seems the only controversial part, since everything else is just replicating what Intel has supported since Ice Lake. If we're lucky, the current list is still incomplete!

Also, it would be a little surprising for AMD to keep this stuff under wraps. Intel usually publishes ISA support of their products a couple years in advance, so that more software is ready by the time it launches.

Do we know for sure Zen 4's AVX-512 is using the 2x256 method?

https://www.anandtech.com/show/17552...coming-sept-27

[Dukenukemx]

All of the Apple Mac chips we've seen so far are laptop-oriented. We still have yet to see their ARM-based successor to the Mac Pro.

I don't know if we'll ever see the successor the to Mac Pro.

Also, they haven't yet moved beyond the Firestorm core, which first launched in the A14 phone SoC, about 2 years ago.

That's because Apple isn't a hardware manufacturer. In the 2 years since the release of the M1, we've seen faster M1's, including the M2. Since AMD 2020, AMD has finally updated their laptop chips to Zen3 plus RDNA2, which is called Rembrandt. In 2020 AMD released the 4000 series which were laptop only CPUs. Now AMD is about to release Zen4+RDNA3, and next year we'll see their laptop-oriented versions. AMD's been through three core revisions in the past 2 years while Apple is just barely on their second. Apple can't keep up.

In other words, I wouldn't count them out, on the performance front. And I'm pretty sure they still beat Zen 4 in perf/W.

Probably because Ryzen 7000 series are not mobile parts. These chips are focused entirely on performance with power efficiency being secondary. You don't reach 5.7Ghz and expect power efficiency. You won't find a Ryzen 7950X on a laptop. These chips advertise sockets, which Apple doesn't do, as well as removable DDR5, which again Apple also doesn't do. Wait until next year when AMD has their 8000 laptop chips based on Zen4+RDNA3. There's a reason why I called it the beginning.

[coder]

the latter could provide a nice performance boost in some areas.

I prefer fp16 to bf16, but the industry seems to disagree. IMO, bf16 lacks enough precision to be useful for much besides deep learning, whereas fp16 strikes a better balance between precision and range.

I think OpenGL (or perhaps D3D) is what pushed fp16 into IEEE754-2008 (the last part is the year of the revision), but we didn't see GPUs seriously implementing fp16, until they started using it for packed arithmetic aimed pretty squarely at deep learning.

Fun fact: Intel first added support for fp16 in Ivy Bridge, but they only went so far as to add fp16 conversion instructions. Perhaps that was anticipating broader use in GPUs.

[coder]

Unfortunately only early revisions of Alder Lake support AVX-512.

Intel stated that Alder Lake's AVX-512 was never validated. It's not entirely clear whether they meant the at the design level, or merely to ensure chips rolling off the assembly line didn't have any manufacturing defects in the parts specific to AVX-512. Once I read that, I decided I wouldn't risk using it, even if I could.

However, the main relevance of the Alder Lake row is that it should match what Sapphire Rapids officially supports.

[coder]

Intel announced Raptor Lake will have 15% faster IPC than Alder Lake

That can't be right, because the P-cores are essentially the same as Alder Lake's. The only thing that should help Raptor Lake's IPC are bigger L2 caches and tweaks to the ring bus. Those are probably worth a few % at best, and nowhere close to 15%.

Perhaps what you're thinking of is single-thread performance, which is approximately the product of IPC ratio and clock speed ratio. If that's what you mean, then you have to compare it with AMD's claim of 29% faster single-threaded performance than Zen 3. Using your figure of 2% better IPC than Alder Lake would give Zen 4 a 16.4% single-threaded advantage over Alder Lake.

Looks like I’ll be going with Raptor Lake.

Since you can't buy either, right now, might as well wait and see. What if Intel's stated numbers require a custom water cooling loop + a $1k external chiller to achieve, in real life?