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

[DisEnchantment]

Speculate at will

 

[Hail The Brain Slug]

https://videocardz.com/pixel/msi-x870e-motherboards-now-support-up-to-192gb-of-ddr5-memory-at-6400-mt-s

I have a 6400MT/s 2x48GB EXPO kit that's on my motherboard QVL and when I enable EXPO it sets it 6400MT/s, 2000MHz FCLK, and 2:1 UCLK. This is actually stable, indicating the ram and motherboard are fine at 6400. If I change to 1:1, it spams memory errors from virtually every address, indicating the IMC can't handle it. This isn't a motherboard issue.

Until I see some independent verification that demonstrates MSIs boards will actually do 1:1 UCLK with 2+ dual rank dimms, I will assume it is at 2:1. Considering this is silicon lottery for your IMC territory, I am skeptical this is even possible.

 

[CP5670]

Apparently 2:1 is supposed to hurt performance a bit and it's actually faster to do 6000 at 1:1.

 

[lightmanek]

I have a 6400MT/s 2x48GB EXPO kit that's on my motherboard QVL and when I enable EXPO it sets it 6400MT/s, 2000MHz FCLK, and 2:1 UCLK. This is actually stable, indicating the ram and motherboard are fine at 6400. If I change to 1:1, it spams memory errors from virtually every address, indicating the IMC can't handle it. This isn't a motherboard issue.

Until I see some independent verification that demonstrates MSIs boards will actually do 1:1 UCLK with 2+ dual rank dimms, I will assume it is at 2:1. Considering this is silicon lottery for your IMC territory, I am skeptical this is even possible.

Yes, but you're missing the point ... 4x48GB not 2x48GB, a massive difference!
Currently most boards default to amazingly depressing 3600MT/s speeds when running 4x48GB on Zen 4/5 and you need a lot of effort to make it semi stable at 5200MT/s

 

[Win2012R2]

Yes, but you're missing the point ... 4x48GB not 2x48GB, a massive difference!

Could be very good for a cheap server, if only that memory was supporting ECC and if AMD gifts us dual chiplet 3D cache...

 

[CakeMonster]

I'm just waiting for 64GB sticks, hopefully it won't be too long.

 

[igor_kavinski]

I'm just waiting for 64GB sticks, hopefully it won't be too long.

Last I heard it was 25H1.

 

[eek2121]

I have a 6400MT/s 2x48GB EXPO kit that's on my motherboard QVL and when I enable EXPO it sets it 6400MT/s, 2000MHz FCLK, and 2:1 UCLK. This is actually stable, indicating the ram and motherboard are fine at 6400. If I change to 1:1, it spams memory errors from virtually every address, indicating the IMC can't handle it. This isn't a motherboard issue.

Until I see some independent verification that demonstrates MSIs boards will actually do 1:1 UCLK with 2+ dual rank dimms, I will assume it is at 2:1. Considering this is silicon lottery for your IMC territory, I am skeptical this is even possible.

I have 64gb in 1:1 mode I run at 6200 with a 2100 mhz IF clock. The system also boots at 6400, but I haven’t had time to get it stable. I’ve seen people hit 6400 1:1 and 8200 with 1:2 with their 96gb kits and a 2200 MHz IF clock.

You just have to bump up voltages, loosen timings, and make sure the DIMMs stay cool.

EDIT: Added screenshot of ZenTimings. Oh and I'm not done optimizing timings either.

 

[511]

Strix Halo

https://twitter.com/x/status/1868701872401362956

 

[FlameTail]

Single core performance closer to Strix Point, not Granite Ridge? I was hoping for the latter.

Perhaps it's because this is an ES, and the score will be higher in retail devices.

 

[CouncilorIrissa]

It does not look like there is any meaningful difference between STX and STX-H in 1t.
https://browser.geekbench.com/v6/cpu/compare/9452051?baseline=9460276, only SIMD-heavy Background Blur and Object Remover subtests see a minor increase.

Reported clocks are locked @ 5.1

 

[Hail The Brain Slug]

Single core performance closer to Strix Point, not Granite Ridge? I was hoping for the latter.

Perhaps it's because this is an ES, and the score will be higher in retail devices.

.... it's the same architecture at the same clockspeed, why would strix halo at 5.1 GHz score meaningfully different from strix point at 5.1 GHz?

 

[GTracing]

Looks like the ram is only running at a quarter the speed it should be? 1994MT/s? Or is that just reporting shenanigans?

Isn't Strix Halo supposed to have a 100W TDP? Seems hot for a 13" tablet. But apparently the previous Flow Z13 could draw 100W under load, so I guess that tracks.

 

[FlameTail]

.... it's the same architecture at the same clockspeed, why would strix halo at 5.1 GHz score meaningfully different from strix point at 5.1 GHz?

I thought Strix Halo would clock higher than Strix Point.

And it also has twice the L3 cache per Zen5 CCX.

 

[igor_kavinski]

Looks like the ram is only running at a quarter the speed it should be? 1994MT/s? Or is that just reporting shenanigans?

Seems ok. Multiply by four and it's roughly 8000 MT/s.

ASUSTeK COMPUTER INC. ROG Flow Z13 GZ302EA_GZ302EA_031572486 vs ASUS System Product Name - Geekbench

browser.geekbench.com

Look at the Max GO in MT!

8 wins against a higher TDP desktop chip!

Even beats Det0x's souped up 9950X in the photo filter MT test! https://browser.geekbench.com/v6/cpu/compare/9460276?baseline=9430627

I knew Ryzen was badly bandwidth starved. This confirms it.

And the real kicker is that GB6 can't isolate EVERY performance aspect of any given chip!

So there's gotta be more wins in store for the Max in other benchmarks!

 

[igor_kavinski]

Google caught the Belarussians red-handed trying to sell Strix Halo before announcement!



Price starting at around $3100.

 

[MS_AT]

.... it's the same architecture at the same clockspeed, why would strix halo at 5.1 GHz score meaningfully different from strix point at 5.1 GHz?

Are they the same? Wasn't Halo supposed to be using desktop like chiplets with full SIMD and twice L3 cache?

 

[CouncilorIrissa]

I thought Strix Halo would clock higher than Strix Point.

And it also has twice the L3 cache per Zen5 CCX.

Twice the L3 cache at twice the cores, so same amount of L3 per core

Looks like the ram is only running at a quarter the speed it should be? 1994MT/s? Or is that just reporting shenanigans?

Isn't Strix Halo supposed to have a 100W TDP? Seems hot for a 13" tablet. But apparently the previous Flow Z13 could draw 100W under load, so I guess that tracks.

It's reporting. 7500 MT/s LPDDR5 is reported as 1867.

.... it's the same architecture at the same clockspeed, why would strix halo at 5.1 GHz score meaningfully different from strix point at 5.1 GHz?

It's presumably using the GNR CCD (meaning it should have Zen 5 cores with the full FPU rather than the STX chop) and a different IOD that should remedy the IOD <-> CCD throughput bottleneck caused by the single 32B/clk GMI3 link.

 

[Tuna-Fish]

Single core performance closer to Strix Point, not Granite Ridge? I was hoping for the latter.

Perhaps it's because this is an ES, and the score will be higher in retail devices.

It will lose some for using LPDDR5, memory latency is going to be higher. Hopefully not this much, though.

 

[poke01]

Seems ok. Multiply by four and it's roughly 8000 MT/s.

ASUSTeK COMPUTER INC. ROG Flow Z13 GZ302EA_GZ302EA_031572486 vs ASUS System Product Name - Geekbench

browser.geekbench.com

Look at the Max GO in MT!

8 wins against a higher TDP desktop chip!

Even beats Det0x's souped up 9950X in the photo filter MT test! https://browser.geekbench.com/v6/cpu/compare/9460276?baseline=9430627

I knew Ryzen was badly bandwidth starved. This confirms it.

And the real kicker is that GB6 can't isolate EVERY performance aspect of any given chip!

So there's gotta be more wins in store for the Max in other benchmarks!

This is the future of laptops. Exciting isn’t it.

 

[LightningZ71]

Doesn't this chip have MALL cache? Shouldn't that help with memory latency a bit?

 

[Saylick]

Doesn't this chip have MALL cache? Shouldn't that help with memory latency a bit?

STH does. 32 MB worth, if I remember correctly.

 

[Josh128]

Dont expect compute miracles out of Halo. I believe power management/efficiency for the speed it packs is going to be its biggest advantage outside of the GPU aspect.

 

[inquiss]

Doesn't this chip have MALL cache? Shouldn't that help with memory latency a bit?

Think it helps GPU more than CPU

 

[fastandfurious6]

ASUSTeK COMPUTER INC. ROG Flow Z13 GZ302EA_GZ302EA_031572486 vs ASUS System Product Name - Geekbench

browser.geekbench.com

Look at the Max GO in MT!

8 wins against a higher TDP desktop chip!

Even beats Det0x's souped up 9950X in the photo filter MT test! https://browser.geekbench.com/v6/cpu/compare/9460276?baseline=9430627

I knew Ryzen was badly bandwidth starved. This confirms it.

And the real kicker is that GB6 can't isolate EVERY performance aspect of any given chip!

So there's gotta be more wins in store for the Max in other benchmarks!

Yes............ as expected

but how is Ray Tracer MT 63% explained? Almost half score

and why ST -15%?

also Flow Z13 is tiny and probably throttles in full load

max halo in 13inch tiny ultralightweight, weighs only ~1.2kg!

 

[igor_kavinski]

but how is Ray Tracer MT 63% explained? Almost half score

and why ST -15%?

also Flow Z13 is tiny and probably throttles in full load

There's your answer. I bet we may see a lot of performance unlocked at 150W TDP and better cooling in workstation laptops like Thinkpads, Dell Precisions and HP Zbooks.

The iGPU is also unable to stretch its legs properly and losing to a desktop RX 7600 in OpenCL and Vulkan GB6 tests.