Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[CouncilorIrissa]

As I wrote above, if you look at the ray tracing score, you get more or less what you expected.

As I wrote dozens of times: don't concentrate on aggregated scores, study subtests results. GB6 global MT score doesn't reflect perfect scaling at all, that was a huge departure from GB5, and IIRC this was explained by PrimateLabs.

The GB6 MT score change bugs me a little. I understand where they're coming from. It's just that people have a mental model of what a MT score usually means (embarassingly parallel workloads), and they hijacked the term and replaced it with something else entirely.

I think Userbenchmark of all places had the right idea: they had separate scores for single-core, quad-core and multicore performance. Too bad the site is garbage otherwise.

 

[igor_kavinski]

60W: https://forums.anandtech.com/thread...ranite-ridge-ryzen-9000.2607350/post-41246800
90W: https://forums.anandtech.com/thread...ranite-ridge-ryzen-9000.2607350/post-41248191

 

[StefanR5R]

I'm looking forward to the SP6 Zen5 products. I wonder if, for this generation, AMD will bless us with a 4 CCD full fat Zen5 product with 32 Zen5 cores running at decent speeds? It would be a nice replacement for older Threadrippers out there. Even the 64 core Zen5c parts would provide a lot of grunt for workstations that use AVX-512 heavily. There are some nice SP6 workstation ATX boards out there...

Asymmetric architecures are probably fine for low performance edge servers, but this might not be good for workstation performance. Also, current 8004 processors top out at 200 W. SP6 is meant to be a low power socket.

7k TR's are SP6 for both platforms and they're 350W just fine.

Are SP6 and sTR5 power pinouts the same though?

(Purely going by what's taken up by I/O, SP6 could actually have at least as many power pins as sTR5.
SP6 = LGA4844 = 96 PCIe 5.0 lanes and 6 DDR5 channels
sTR5/WRX90 = LGA4844 = 128 PCIe 5.0 lanes and 8 DDR5 channels)

 

[StefanR5R]

Does RAID5 mean that Zen 5 has hardware RAID support???

NVME RAID 10 and 5 look to me as if these are features which were developed with EPYC 4004/4005 in mind.

What about that NVMe RAID 5? AMD not including it on 7950X product page means that Zen 5 has special hardware instructions to speed up XOR operations?

The wider vector pipelines should help with that, shouldn't they?

 

[Josh128]

View attachment 102731

60W: https://forums.anandtech.com/thread...ranite-ridge-ryzen-9000.2607350/post-41246800
90W: https://forums.anandtech.com/thread...ranite-ridge-ryzen-9000.2607350/post-41248191

View attachment 102734

So this is PPT power and not TDP. Looks to be within 10% on monster and classroom, iso for junkshop vs 7950X 230 PPT. However, the use of unknown curve optimizer offsets renders the result absolutely useless for comparison. Im sure a 7950X could be aggressively CO tuned to very similar scores at 120W PPT.

Its time to just rip the bandaid off, this smoke and mirrors stuff is a major buzzkill.

 

[Kryohi]

The GB6 MT score change bugs me a little. I understand where they're coming from. It's just that people have a mental model of what a MT score usually means (embarassingly parallel workloads), and they hijacked the term and replaced it with something else entirely.

I think Userbenchmark of all places had the right idea: they had separate scores for single-core, quad-core and multicore performance. Too bad the site is garbage otherwise.

IMHO the approach of GB6 is extremely dumb. For actual MT workloads where high performance is needed I've never seen such a bad thread scaling. Even some mixed ST/MT algorithms (UMAP is an example) with a big serial part of the code overall scale better than whatever GB is doing, on desktop CPUs. If they want to keep doing this honestly they should at least create a true MT benchmark and rename the current one.

 

[Det0x]

Where are you @Hail The Brain Slug and @tsamolotoff ? 🙏

 

[Makaveli]

This is how GB6.3 reports 2 dimms being populated on a dual channel memory system. This is full memory bandwidth results.

Why do the other leaks show 4 channels and this one 2? and I see different versions of GB being used also so maybe that is the reason.

 

[Hail The Brain Slug]

Where are you @Hail The Brain Slug and @tsamolotoff ?

I'll run these when I have a chance later. After some discussion in the GB6 thread I'm going to set aside my tuned ram config and just set it to EXPO 6000C30 to get a more accurate representation of a "stock 7950X". I was already running the CPU itself fully stock with no PBO, but the ram and fclk setup may have been inflating some scores.

Problem is, in a PPT limited scenario like this my uncore power will be much less than that Zen 5 sample seems to be running, giving me more power per core to boost higher. Even at my tuned 6400/2133 config with a bunch of voltages jacked my uncore was using less power.

 

[Markfw]

It would seem to me that with all these 9xxx retail results showing up, that chips must be in reviewers hands. See reviews Monday maybe ???

 

[Makaveli]

It would seem to me that with all these 9xxx retail results showing up, that chips must be in reviewers hands. See reviews Monday maybe ???

I don't think we are seeing reviews until july 31

 

[Josh128]

I'll run these when I have a chance later. After some discussion in the GB6 thread I'm going to set aside my tuned ram config and just set it to EXPO 6000C30 to get a more accurate representation of a "stock 7950X". I was already running the CPU itself fully stock with no PBO, but the ram and fclk setup may have been inflating some scores.

Problem is, in a PPT limited scenario like this my uncore power will be much less than that Zen 5 sample seems to be running, giving me more power per core to boost higher. Even at my tuned 6400/2133 config with a bunch of voltages jacked my uncore was using less power.

Could you possibly run JDEC speeds, vs EXPO 6000C30, vs your tuned setup? Honestly, I have no feckin clue the difference this makes in this ridiculously finicky benchmark.

 

[igor_kavinski]

I don't think we are seeing reviews until july 31

ASUS has their Ryzen AI event on 17th so we could see at least Ryzen AI reviews earlier than 31st.

 

[Makaveli]

ASUS has their Ryzen AI event on 17th so we could see at least Ryzen AI reviews earlier than 31st.

yes we will probably see APU reviews sooner but desktop not until the end of the month.

 

[igor_kavinski]

yes we will probably see APU reviews sooner but desktop not until the end of the month.

Geekerwan could snag a retail sample without having to sign an NDA

 

[Hail The Brain Slug]

Could you possibly run JDEC speeds, vs EXPO 6000C30, vs your tuned setup? Honestly, I have no feckin clue the difference this makes in this ridiculously finicky benchmark.

I can run JEDEC 4800 but official support is 5200 and I don't have any ram with that profile. Not really sure what value thing presents, however, since on AM5 I have never seen anyone running JEDEC ram except when populating 4xDR dimms for maximum capacity.

 

[igor_kavinski]

I can run JEDEC 4800 but official support is 5200 and I don't have any ram with that profile. Not really sure what value thing presents, however, since on AM5 I have never seen anyone running JEDEC ram except when populating 4xDR dimms for maximum capacity.

JEDEC Extends DDR5 Memory Specification to 8800 MT/s, Adds Anti-Rowhammer Features

www.anandtech.com

That's the one you will have to do, if you want to try JEDEC. You will need to set voltage to 1.1V

 

[Hail The Brain Slug]

JEDEC Extends DDR5 Memory Specification to 8800 MT/s, Adds Anti-Rowhammer Features

www.anandtech.com

View attachment 102737

That's the one you will have to do, if you want to try JEDEC. You will need to set voltage to 1.1V

That's not enough information to set up every timing accurately to a 5200 profile. I can leave the other 30-some timings on auto but that may not accurately reflect if it was ram with a native jedec 5200 profile.

 

[tsamolotoff]

Where are you @Hail The Brain Slug and @tsamolotoff ? 🙏

Still working, will do the run later (this is like some sort of cyber-edging by our friend eye-gore )120W is though fairly close to what my x3d typically goes up to with MT tasks (~170w max)

 

[Philste]

So this is PPT power and not TDP. Looks to be within 10% on monster and classroom, iso for junkshop vs 7950X 230 PPT.

Yeah but:

Im sure a 7950X could be aggressively CO tuned to very similar scores at 120W PPT.

You don't even need to tuned it: According to launch reviews a 7950X just set to Eco Mode (142W PPT) loses only 5% performance compared to 230W PPT. So if you manually tuned it even further it's probably better than that ZEN5 sample.

Why do the other leaks show 4 channels and this one 2?

Because the correct memory Channel Reporting came with 6.3. Your first screenshot is 6.2.1

 

[tsamolotoff]

Could you possibly run JDEC speeds, vs EXPO 6000C30, vs your tuned setup? Honestly, I have no feckin clue the difference this makes in this ridiculously finicky benchmark.

Probably it has close to zero impact, as mem bw is fairly low during the rendering itself (idk, ~5gbps ish from hwinfo metrics)

 

[tsamolotoff]

Still faster, it seems. Or i'm doing something wrong. Typical in-benchmark clocks on the first screenshot

 

[igor_kavinski]

View attachment 102742

View attachment 102743

Still faster, it seems. Or i'm doing something wrong. Typical in-benchmark clocks on the first screenshot

Disable V-cache and re-run, if you can.

 

[tsamolotoff]

Disable V-cache and re-run, if you can.

impossible, you can't disable v-cache

 

[gaav87]

View attachment 102742

View attachment 102743

Still faster, it seems. Or i'm doing something wrong. Typical in-benchmark clocks on the first screenshot

bclk overclock ?