AMD’s Zen 4, Part 3: System Level Stuff, and iGPU | Hacker News Reader

AMD’s Zen 4, Part 3: System Level Stuff, and iGPU

chipsandcheese.com

182 pointszdw21 comments

  5950X hit its highest clocks on cores 1 and 4. Both are located on the first CCD. My 3950X clocked highest on cores 4 and 5. Both are also on the first CCD. This is a curious pattern. Perhaps AMD is only binning one CCD to hit the maximum boost clocks on 16 core parts.

I have 5950X and saw this pattern, the fastest cores are though 2nd and 6th (starting from 0). I don't remember exact values from testing (with stress-ng and openssl speed), but it was something similar to what authors observed in their testing. Those two cores were able to reach speeds of around 4.97, the next ones 4.95 or so, with the slowest pair stuck at around 4.7GHz.

This would make sense right? You probably can't run all cores at max turbo frequency before hitting power and thermal limits. So dividing chiplets based on binning would result in good performance for 1-8 -threaded tasks while reducing chip lottery.

(But it's possible there are other reasons for the discrepancy)

The first CCD is binned for the SKU's spec, the other is whatever. This has been true for all chiplet Ryzens. Though the effect used to be more extreme, I had for example a 3900X where the second CCD only managed to run at 4.1 GHz under load, compared to ~4.4 GHz on the first CCD (and 4.6 GHz being printed on the box). That's pretty common with those.

Is there optimization being done by AMD to locate the best performing chiplet on the best possible place on the bus?

There's no bus and thus no place on the bus.

?950x chips in particular are binned that way, the first die is an "Epyc"-grade bin, the second die is a bin good enough for ?800x (or, if sold as 6 cores, both of the ?900x cores).

AMD binning independently works of functioning cores; having all 8 cores working does not mean it magically bins higher. AMD has not commented publicly on this, but it is generally believed there are two Epyc bins (one for high perf, one for low voltage, both exceptional quality), one high quality Ryzen bin, and one low quality bin.

?950x contains one Epyc quality + one high quality; x models contain high quality dies (in addition to being binned for broken cores in 6 core-based models independent of clock speed binning), non-x using lower quality dies. Threadrippers are bin-selected similarly to ?950xs.

All 6 core Zen 2 chips that were sold were 3+3, never 4+2/2+4.

The only exception to the above binning is a single 4 core non-APU Zen 2 existed, using the very few dies where an entire CCX was dead in the CCD, but the other CCX worked. This also mean half the L3 was dead (as the L3 was split per CCX, and wasn't CCD wide). With Zen 3 and 4, this chip can't exist as they are single CCX (all 8 cores) per CCD.

Die binning, however, isn't magical. ?800x record-setting overclocks are only a smidgen behind record-setting ?590x per-chiplet overclocks.

Thermos?

Could it be a simpler explanation in that the thermos of cores 1 & 4 are better than where other cores are located?

Part 1: https://news.ycombinator.com/item?id=33477119

Part 2: https://news.ycombinator.com/item?id=33518751

I'd really like to see more info on the desktop G APU's. I'm not even sure if they're making a Dragon Range desktop APU, but it's what I personally am most interested in (I get that the 7000 series have fairly meager integrated graphics as is).

I realise it's a relatively unimportant market segment for them, but some info either way would be nice regardless.

They probably will. It usually takes a while because they prioritize shipping the chip in laptops, then OEM systems. Maybe they skip the second part this time because now they have the basic igpu for the office PC market. Also with the down market they may have an oversupply which let's them start selling at retail faster.

Dragon Range is identical to Raphael that's already out.

I thought they might release a desktop Rembrandt in the $100-200 range but they haven't, probably because AM5 motherboards are so expensive. They may or may not release Phoenix on AM5 later this year.

We're allegedly going to see this in 2H '23. The 2 CU iGPU performance is about half of a 5600/5700g, so if we see an 8-12 CU model...it could completely replace the low-mid market for gpus.

Do you expect it to scale that well? If it were linear it would reach something like a RX 570, but isn't it more likely it will just be a bit better than the graphics of the 5700G? So probably not even beating a GTX 1050?

I'm honestly just happy the iGPU isn't hanging off the PCIe bus anymore compared to the APUs. Having to choose between iGPU and 8 PCIe lanes on an already I/O constrained desktop part was obnoxious.

Apparently "CCD" stands for "core-complex die", a dumb acronym that just means "chip with cores", and different from the other "chip with i/o interconnect" you also get in a module.

It is annoyingly rare to find anybody expanding this acronym.

A meaningful definition of "CCX" is harder to find.

CCX or core complex is a lower level grouping of cores. On Zen 2/3 the CCX would have 4 cores on a little ring bus. A full 8-core CCD would then have two of those 4-core CCXs connected via infinity fabric. The inter-CCD connections were also made with infinity fabric between the CCXs.

Zen 4 increases the CCX size to 8, so each CCD only has one.

I can't remember exactly how Zen 1 worked but I think it was different.

Nit: Zen 3 increased the CCX size to 8 cores.

You have the generations there a bit off. Zen 1/2 were 4-core CCXs and it was Zen 3 that switched to 8 core CCXs. Zen 2 is what introduced the CCD, but kept Zen 1's CCX.

Off topic: I am looking for the hardware programing manual of Zen IOMMU (to map bus address to physical address), I cannot find it or what???

The latest revision is from October 2022, so I assume that the many changes that have been incorporated into this revision are applicable to Zen 4:

https://www.amd.com/en/support/tech-docs/amd-io-virtualizati...

omg... I read "virtualization" missing out the "I/O" and thought it was about the virtual machine hardware support...

Thx...

Reply on news.ycombinator.com