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

[DisEnchantment]

Speculate at will

 

[fastandfurious6]

smells like FUD

 

[igor_kavinski]

Found two more threads:

https://www.reddit.com/r/ASRock/comments/1ikxrvf/two_dead_9800x3ds

https://www.reddit.com/r/ASRock/comments/1iegh2w/brand_new_9800x3d_dead

Seems like someone at ASROCK messed up pretty bad.

 

[leoneazzurro]

xFrom the dead 9800X3D reddit thread:

View attachment 116881

Quite frankly two occurrences which at least one appears to be the user's fault out of tens of thousands CPU sold are statistically insignificant. In any case, there's Buildzoid's take on it:

 

[Hitman928]

We can rule that out I think. Wouldn't have lasted 20 days unless there's some process that causes unconductive crap to turn into conductive crap through heat over a period of three weeks. Most likely ASROCK BIOS defaults or EXPO voltage handling are to blame.

If there's a bent pin, it may not cause issues at first, but after many thermal cycles the pin continues to deform until it gets too close to a neighboring pad/pin and then the problem becomes spectacularly apparent.

 

[leoneazzurro]

If there's a bent pin, it may not cause issues at first, but after many thermal cycles the pin continues to deform until it gets too close to a neighboring pad/pin and then the problem becomes spectacularly apparent.

Or it stops making proper contact and at that point the consequences are similar.

 

[Schmide]

Bring back the pin grid array

 

[Thibsie]

Bring back the pin grid array

In the classroom, kids doing PCs unbuild/rebuild break pins half of the time.if I asked them to remove CPU. Thanks LGA

 

[Jan Olšan]

smells like FUD

Well, it is conceivable Nvidia for example could try try to divert attention due to their melting issue (or at least give fans ammo for whataboutisms cope), but I *really* doubt it. Such conspiracy explanations are unlikely. Maybe more to be expected from a single silly fan (think Userbenchmark guy), but in that case the dude would have to sacrifice a lot of money.

It's likely just normal failures. Hidden manufacturing fault, board somehow frying the CPU (wouldn't be the first time on AMD or Intel platforms - remember V-Cache is voltage sensitive)...

 

[Thibsie]

Well, it is conceivable Nvidia for example could try try to divert attention due to their melting issue (or at least give fans ammo for whataboutisms cope), but I *really* doubt it. Such conspiracy explanations are unlikely. Maybe more to be expected from a single silly fan (think Userbenchmark guy), but in that case the dude would have to sacrifice a lot of money.

It's likely just normal failures. Hidden manufacturing fault, board somehow frying the CPU (wouldn't be the first time on AMD or Intel platforms - remember V-Cache is voltage sensitive)...

I suggest sacrifying the dude instead

 

[igor_kavinski]

https://www.reddit.com/r/LocalLLaMA/comments/1ip8mtm/amd_ryzen_ai_max_395_strix_halo_mini_pc_tested

Read bandwidth seems like a huge disappointment

 

[MS_AT]

Read bandwidth seems like a huge disappointment

What is the significance of the yellow hypervisor text above the Save button. Could the test be running inside a VM?

 

[igor_kavinski]

What is the significance of the yellow hypervisor text above the Save button. Could the test be running inside a VM?

Seems AIDA64 has detected sandbox or virtualized environment. Hmmm...

 

[Schmide]

https://www.reddit.com/r/LocalLLaMA/comments/1ip8mtm/amd_ryzen_ai_max_395_strix_halo_mini_pc_tested

View attachment 117230

Read bandwidth seems like a huge disappointment

I think from the chips and cheese interview, most is provisioned to the graphics. They kept saying something like all the fabric goes there.

 

[MS_AT]

I think from the chips and cheese interview, most is provisioned to the graphics. They kept saying something like all the fabric goes there.

They said in the interview that Halo CCD has the same BW (actually it's higher as Granite Ridge has 32B/c read and 16B/c write BW between CCD and IOD, and Halo is supposed to have 32B/c BW for both R/W) that desktop parts, so both CCDs together should be able to reach 200GB/s read. It makes little sense for write to be able to reach that and read not being able to do so, therefore I would wait for tests with final bioses before reaching any conclusions.

 

[therealmongo]

If there's a bent pin, it may not cause issues at first, but after many thermal cycles the pin continues to deform until it gets too close to a neighboring pad/pin and then the problem becomes spectacularly apparent.

Ooouh, now you got me worried, not 9800x3d but 7800x3d, a friend gave me an Asus x670e Gene that they had damaged several pins. I managed to straighten them and am using the system as my daily. Its been functioning fine since I repaired it and thats around 6 months ago.

 

[TimCh]

We already knew the CPU read bandwidth would be limited by the fabric, the low score isn’t really a surprise.

Edit: Yes, it is strange that read is so far below the write bandwidth.

 

[igor_kavinski]

I managed to straighten them and am using the system as my daily. Its been functioning fine since I repaired it and thats around 6 months ago.

Just don't watch anything on it that God doesn't approve of

 

[eek2121]

https://www.reddit.com/r/LocalLLaMA/comments/1ip8mtm/amd_ryzen_ai_max_395_strix_halo_mini_pc_tested

View attachment 117230

Read bandwidth seems like a huge disappointment

No, that latency is a disappointment. The R/W speeds are fine, IMO.

 

[igor_kavinski]

No, that latency is a disappointment.

That's what you get with LPDDR any version. Only way to avoid that is to not use LPDDR. Or come up with some sort of FrankenDDR tech that uses both DDR4 and LPDDR5 together in such a way that low latency is provided by the former and high bandwidth by the latter.

 

[DavidC1]

That's what you get with LPDDR any version. Only way to avoid that is to not use LPDDR. Or come up with some sort of FrankenDDR tech that uses both DDR4 and LPDDR5 together in such a way that low latency is provided by the former and high bandwidth by the latter.

140nS is really high if that's accurate. Most mobile systems are in the 100nS range. It makes Arrowlake look very good.

 

[Schmide]

It's lpddr5x which should have high latency high bandwidth, quad or octo channel. I think the system can do weird subchannel dual quad or something which is probably what were seeing here. Video gets a dedicated quad channel plus the MALL cache. CPU and GPU share the other quad channel.

 

[Tuna-Fish]

It's lpddr5x which should have high latency high bandwidth, quad or octo channel. I think the system can do weird subchannel dual quad or something which is probably what were seeing here. Video gets a dedicated quad channel plus the MALL cache. CPU and GPU share the other quad channel.

Doesn't work like that, the channels are not dedicated.

 

[igor_kavinski]

Should be a good day tomorrow (new hardware reviews to read!): https://videocardz.com/newz/amd-ryzen-ai-max-strix-halo-reviews-coming-soon-asus-teases-launch-event

 

[igor_kavinski]

Available for sale in Dubai: https://uaetechdubai.ae/product/asu...m-1tb-ssd-amd-strix-halo-graphics-win-11-pro/

I don't like "fat tablet" form factor. Why couldn't ASUS have a normal Creator laptop too?

 

[Geddagod]

7ish hours until this, I hope slides get released online soonish.
Very curious about what claims they have about the CCX power vs Zen 4, V/F curve, and CaC numbers they have about Zen 5. I don't think AMD has yet to release a gen since zen where even at low power their next gen core hasn't had a perf/watt increase vs their past gen, but this doesn't seem to be the case with Zen 5.