AMD Answers Our Zen 4 Tech Questions, with Robert Hallock

[Displayname1008]

我想知道 ST 性能超过 5950X 的 15% 是否是使用 3DV 缓存的 SKU？如果不是，AMD 是否有可能如前所述，对这些数字非常保守，只引用没有缓存的 SKU？带有 3DV 缓存的 AM5 SKU 会将这些数字推高超过 20-25% 的 ST 增益吗？如果它们在游戏中只匹配 5800X3D 性能，我会坚持更长时间......

no. ST 15% is normal ZEN4 CPU - normal ZEN3 CPU. them all no 3DVcache

 

[btarunr]

turn on IGPU will debase CPU and Mem performance.

This is true. The iGPU eats into the power-budget and that affects CPU boosting headroom. And since iGPU shares system memory, your memory perf naturally drops.

I spent quite some time with the 5700G's iGPU. It's surprisingly capable, but I noticed system performance lowered, so went back to the dGPU.

I wish AMD gave the 5700G a TDP of 105W (140 W PPT). That would've freed up more CPU boosting headroom. Instead they gave it 65W/88W. Maybe the silicon can't take that much power.

 

[konga]

I wonder if that 15% ST performance gain over 5950X is an SKU utilizing 3DV Cache? If not, is it possible that AMD is, as said, being very conservative with these figures and only referencing SKUs without the cache? Would an AM5 SKU with 3DV Cache push these numbers over 20-25% ST gain? If they only match the 5800X3D performance in games, I'll stick with mine longer...

AMD has already confirmed that there won't be 3D V-Cache in the launch lineup. Besides, it would likely reduce the ST performance as it's measured here (in Cinebench), as the 5800X3D does worse in Cinebench than the 5800X. I would not expect a Zen 4 CPU with 3D V-Cache until sometime next year, to be honest.

In his interview with HotHardware, Hallock tried really hard to downplay that 15% ST performance figure, saying "we were conservative on that number in like 4 or 5 unique and important ways," though he did not elaborate on what those ways were. He also repeatedly emphasized that they were working with an engineering sample without a fully optimized frequency curve, and suggested that hitting 5.5 GHz was easy for Zen 4, leaving open the possibility of even higher clocks (whether through OC or higher-end stock configs is unclear)

 

[Crackong]

 

[Minus Infinity]

I wonder if that 15% ST performance gain over 5950X is an SKU utilizing 3DV Cache? If not, is it possible that AMD is, as said, being very conservative with these figures and only referencing SKUs without the cache? Would an AM5 SKU with 3DV Cache push these numbers over 20-25% ST gain? If they only match the 5800X3D performance in games, I'll stick with mine longer...

You don't actually believe this is the actual performance increase right? They are sandbagging. Do you think they would give Intel the specs so far out from release, so they can then go an adjust pricing and performance on Raptor Lake as needed. You will see >> than 15% ST. They are quoting the bare minimums here.

 

[Displayname1008]

This is true. The iGPU eats into the power-budget and that affects CPU boosting headroom. And since iGPU shares system memory, your memory perf naturally drops.

I spent quite some time with the 5700G's iGPU. It's surprisingly capable, but I noticed system performance lowered, so went back to the dGPU.

I wish AMD gave the 5700G a TDP of 105W (140 W PPT). That would've freed up more CPU boosting headroom. Instead they gave it 65W/88W. Maybe the silicon can't take that much power.

on Desktop PC, debase performance not about TDP. becouse You can in BIOS change TDP/PPT or more Power Limit. if you use IGPU, your performance must be lost. only if games support moreGPU function like "Ashes of the Singularity"

 

[konga]

You don't actually believe this is the actual performance increase right? They are sandbagging. Do you think they would give Intel the specs so far out from release, so they can then go an adjust pricing and performance on Raptor Lake as needed. You will see >> than 15% ST. They are quoting the bare minimums here.

I think when a company gives a figure that isn't an obvious highball, we should take them at their word. The onus is on them to prove otherwise. And I think Intel already knows exactly how powerful Zen 4 is going to be. It's very hard to keep secrets in this industry—everyone knows everything.

As a rule, I never give any tech companies the benefit of the doubt. It is better to go in skeptical and be pleasantly surprised than to come up with scenarios where they're "sandbagging" and end up disappointed.

 

[Makaveli]

As a rule, I never give any tech companies the benefit of the doubt. It is better to go in skeptical and be pleasantly surprised than to come up with scenarios where they're "sandbagging" and end up disappointed.

While I do think AMD is being conservative I agree with this approach.

 

[blu3dragon]

Need more info on the "solder the dies to the heatspreader" part. Getting heat out of the tiny ccds is a big issue with zen3, particularly the x3d. Have they improved the thermal interface between the two? Did they build in any extra space on the silicon itself around the hotspots?

 

[zlobby]

It's strange. Robert Hallock was on live with HotHardware presumably after this interview was given, and he was a lot more guarded. He refused to say say anything about the AI acceleration instruction sets, refused to confirm AVX-512, and he wouldn't say if any memory faster than DDR5-6000 would be supported. Yet here he is just casually telling you about their AVX-512 support and DDR5-6400 support. What did you guys do to him to get this info out of him?

They call him a @W1zzard for a reason.

 

[LuxZg]

Am I the only one who thinks that the frase makes one think that AM5 will be short lifed?

I have other fears... That this gives them 5 years until DDR5 reaches DDR4 price parity and that they can wait 2+ years for DDR5 to become abundant :-/ both would be in "AM5 lifetime"

I do hope otherwise, with servers, and even ARM based chips going DDR5 route, I seriously hope DDR5 ramp is actually under way. I want new PC this year

 

[Taraquin]

Fclk will be interesting. I fear it will be 1:4, but hope for 1:2. How far fclk will go is also interesting. Hope we can get 2000+/4000+ functioning on all chips this time, not like Zen 3 where some only did 1833-1866, most did 1900, but some did 2000-2100.

 

[FlanK3r]

good interview

 

[Frank_100]

There was some mention of the zen4 cores being able to function as both P and E cores and maybe shades in-between by controlling power throughout the processor. I hope to hear more about that aspect if it made it into the final design.

 

[Mussels]

How did this end up discussing IGPS again?

In windows, you can choose what GPU does the rendering work and display it on another. It's pretty awesome.
If you dont make any choices, it renders on the GPU its physically connected to. A system is never gunna go oopsies, lets throw cyberpunk onto your APU and ignore your RTX 3090, unless you have your display connected to that IGP

 

[Space Lynx]

You can just disable the igpu in the bios you know

That will depend on the motherboard manufacturer, not all will offer that in their BIOS.

Anyways, I regret cancelling my order my the 5800X3D now, I had my order in on launch day and decided to cancel it at last second because I was uncertain on GPU futures. All I do is game, so I am still hoping to get my hands on a 5800X3D... hopefully at a discounted price this Fall. If not that... then Intel 12700KF

How did this end up discussing IGPS again?

In windows, you can choose what GPU does the rendering work and display it on another. It's pretty awesome.
If you dont make any choices, it renders on the GPU its physically connected to. A system is never gunna go oopsies, lets throw cyberpunk onto your APU and ignore your RTX 3090, unless you have your display connected to that IGP

Where do I find this in Windows? My gtx 1070 laptop has always had problems defaulting to integrated graphics for older games...

 

[TheLostSwede]

I have other fears... That this gives them 5 years until DDR5 reaches DDR4 price parity and that they can wait 2+ years for DDR5 to become abundant :-/ both would be in "AM5 lifetime"

I do hope otherwise, with servers, and even ARM based chips going DDR5 route, I seriously hope DDR5 ramp is actually under way. I want new PC this year

Prices are already coming down. XPG Lancer DDR5 6000 16G*2 CL40 is going for less than US$255 locally here.
Still pricey, sure, but it's half the price it was only a couple of months ago.
Higher-end 16 GB DDR4 modules aren't that much cheaper here.

 

[btk2k2]

AMD has already confirmed that there won't be 3D V-Cache in the launch lineup. Besides, it would likely reduce the ST performance as it's measured here (in Cinebench), as the 5800X3D does worse in Cinebench than the 5800X. I would not expect a Zen 4 CPU with 3D V-Cache until sometime next year, to be honest.

In his interview with HotHardware, Hallock tried really hard to downplay that 15% ST performance figure, saying "we were conservative on that number in like 4 or 5 unique and important ways," though he did not elaborate on what those ways were. He also repeatedly emphasized that they were working with an engineering sample without a fully optimized frequency curve, and suggested that hitting 5.5 GHz was easy for Zen 4, leaving open the possibility of even higher clocks (whether through OC or higher-end stock configs is unclear)

I can hazard a guess as to a few of those unique ways and some speculation on a 3rd.

1) CB R23 ST does not really scale with memory or cache so doubling the L2 and the switch to DDR5 when using this as a benchmark don't add anything. I expect in other workloads (like games) it will be a different story.

2) In the Zen 2 to Zen 3 IPC uplift / performance tests we saw around the place the CB R23 ST bench was on the lower end of the performance increase scale already and there is no real reason to think that won't be the case again given the known changes to the core.

3) This is speculation on my part, however, the slide with that >15% gain claim also had a 5Ghz + claim and we clearly saw that in one game it could hit 5.5Ghz (and in interviews it was claimed this was easy so probably didn't take much if any voltage tweaking or even a super optimised boost algorithm). Even though the test was performed with a 16c engineering sample given the conservative clockspeed shown on the slide it is possible that for the CB R23 ST test they set the max boost to a conservative value to more closely resemble what they expect the lower end of the launch skus to boost to rather than what the gain is when maxing the single core clock speed.

As for your comment a bit later about waiting for AMD to prove it I agree with that entirely. If AMD had come out making claims like upto 30% more ST than Zen 3 and upto 50% more MT than the Zen 3 the chorus would be wait for benchmarks and AMD giving more conservative numbers should not really change that.

 

[asdkj1740]

Prices are already coming down. XPG Lancer DDR5 6000 16G*2 CL40 is going for less than US$255 locally here.
Still pricey, sure, but it's half the price it was only a couple of months ago.
Higher-end 16 GB DDR4 modules aren't that much cheaper here.

kingston fury beast 6000c40 16*2
217usd recently.

 

[Wirko]

There was some mention of the zen4 cores being able to function as both P and E cores and maybe shades in-between by controlling power throughout the processor. I hope to hear more about that aspect if it made it into the final design.

The greatest advantage of Intel's E cores is not performance per watt. It's performance per area. An E core yields about 2/3 the performance in 1/3 the area of a P core (although it's hard to estimate).

 

[TruthPhoenix]

Unless I hear AMD say that the Ryzen 7000 series is optimized for Dragon Age 4, I will stick with my 5950x 4x8gb DDR4 3200 cl14 RTX 3080 12gb...

 

[The red spirit]

That's a bit shortsighted of them to not add 3D cache to all 7000 parts. I could understand 5000 parts, but imo it should have been standard on new gen Ryzens. It's either that or AMD will try to push it as premium feature for bigger profit margins.

 

[Space Lynx]

That's a bit shortsighted of them to not add 3D cache to all 7000 parts. I could understand 5000 parts, but imo it should have been standard on new gen Ryzens. It's either that or AMD will try to push it as premium feature for bigger profit margins.

I just wish they had one launch SKU with 3DCache, I wouldn't expect them to do it on all the SKU's.

 

[Chomiq]

That's a bit shortsighted of them to not add 3D cache to all 7000 parts. I could understand 5000 parts, but imo it should have been standard on new gen Ryzens. It's either that or AMD will try to push it as premium feature for bigger profit margins.

It would also lock them from overclocking and PBO, at least at current stage of implementation.

 

[btk2k2]

That's a bit shortsighted of them to not add 3D cache to all 7000 parts. I could understand 5000 parts, but imo it should have been standard on new gen Ryzens. It's either that or AMD will try to push it as premium feature for bigger profit margins.

Has TSMC even validated 3d stacking a 6nm(7nm?) die on a 5nm base die? Think that is still going through initial testing so probably not ready for large scale production yet.