Two AMD Ryzen 7000 Series Processors Based on Zen 4 Core Appear: 16-Core and 8-Core SKUs

[eidairaman1]

I doubt 3D cache will even make it to the 7000 series.

The fact its only being released on the gaming focused 5800X, suggests its something that AMD is only doing to overtake(?) the 12600K, 12700K & 12900K in gaming benchmarks, until they're ready to launch the new 7000 series in Q3 or Q4.

AMD will not like Intel being able to say they have the fastest gaming CPU in their marketing for most of 2022...It just gives them a free pass to gain market share, if the 3D cache 5800X can match (or beat) Alder Lake for gaming, 3D cache will have served its purpose.

How many sockets did intel force on end users? Oh right, too many.

AMD is only just now considering AM5 sockets.

 

[stimpy88]

hot and inefficient? It is not out yet and people already know how it will work and what it offers? difficult to get linear scaling?
Can I ask where are you getting this stuff from?

I got it from an interview with AMDs Robert Hallock on PCWorld. He never said that the chip ran hot, but said that the chip was bound by TDP limits, and was clocked (down) accordingly, because they wanted to "ease on the thermals". He also directly said there were performance penalties from using 2 CCDs with 3Dstacking cache, due to the fact that Windows and games do not simply allocate cores inside only one CCD. He did hint that the chip should be overclockable to some extent though.

And my last quote from him was that the cache had no benefits outside of gaming, and not all games benefit. They (AMD) had not found an application that measurably benefited from the extra cache. So expect this new CPU to actually perform worse in some cases with Windows/office and other applications, due to the lower clocks. That's why it's only marketed as a gaming CPU.

Some people like to dig for information, and don't just read the PR blurb.

 

[efikkan]

Yes that's unlikely right now or at least till IPC gains go back to single digit crawl just like half a decade back.

It's important to keep in mind that adding a bunch of L3 cache like that is going to be a "one time" thing, it's not going to be one of those improvements where you can keep doing it and it will add another level of performance.

Also, while cache improvements are appreciated, it's important to remember that they don't add to the CPUs maximum theoretical performance, they just help keep the CPU saturated. So cache improvements alone is not going to be exciting in the long run.

But I'm not worried about improvements slowing down, at least not as long as the production nodes can keep it up. It's very likely that we are going to see >= 50% IPC gains and significant SIMD gains in the next ~5 years, as there are many good improvements lined up.

 

[lexluthermiester]

they just help keep the CPU saturated.

That's the point. When a CPU has a ton of memory space local to it's compute units, it can do more work faster and slow down to fetch data from system memory less frequently. For any application that is sensitive to memory access times, this extra L2 & L3 cache will improve processing dramatically.

 

[efikkan]

That's the point. When a CPU has a ton of memory space local to it's compute units, it can do more work faster and slow down to fetch data from system memory less frequently. For any application that is sensitive to memory access times, this extra L2 & L3 cache will improve processing dramatically.

Sure, but you have to understand there are diminishing returns to just adding cache, especially L3 which is a spillover cache (meaning only evicted cache lines from L2).
Adding L3 will mainly reduce cache misses instruction cache lines, so we will probably see many computationally heavy workloads with minimal gains from loads of L3.
A little more L2 though, that's way more useful.

 

[lexluthermiester]

Sure, but you have to understand there are diminishing returns to just adding cache, especially L3 which is a spillover cache (meaning only evicted cache lines from L2).
Adding L3 will mainly reduce cache misses instruction cache lines, so we will probably see many computationally heavy workloads with minimal gains from loads of L3.
A little more L2 though, that's way more useful.

While true, 64MB of L3 cache is an amount that will make a difference.

 

[ratirt]

I got it from an interview with AMDs Robert Hallock on PCWorld. He never said that the chip ran hot, but said that the chip was bound by TDP limits, and was clocked (down) accordingly, because they wanted to "ease on the thermals". He also directly said there were performance penalties from using 2 CCDs with 3Dstacking cache, due to the fact that Windows and games do not simply allocate cores inside only one CCD. He did hint that the chip should be overclockable to some extent though.

What you got from the interview about the temp was kinda wrong. Just because it is bound by the TDP doesn't mean it runs hot or is inefficient. Every chip is bound to a TDP limits. You can't go over certain TDP since the chip would not handle it. It surely does not mean it will be hot and inefficient.
Games are very sensitive to low latency. 3Dcache is for that. Obviously it depends on the workload you do. In terms of linear scale, I thought the Vcache does not scale linearly. Performance with the use of 3dVcache depends on the workload. 3dcache scales linearly if you add more you get better performance to some extent of course. That's why new Epycs supposed to have so much of the cache.

 

[stimpy88]

What you got from the interview about the temp was kinda wrong. Just because it is bound by the TDP doesn't mean it runs hot or is inefficient. Every chip is bound to a TDP limits. You can't go over certain TDP since the chip would not handle it. It surely does not mean it will be hot and inefficient.
Games are very sensitive to low latency. 3Dcache is for that. Obviously it depends on the workload you do. In terms of linear scale, I thought the Vcache does not scale linearly. Performance with the use of 3dVcache depends on the workload. 3dcache scales linearly if you add more you get better performance to some extent of course. That's why new Epycs supposed to have so much of the cache.

That's what the guy said when he was asked why it was clocked lower than the regular 5800x. Don't know how I could have got that wrong. He said it was to reduce the thermal output of the chip, nothing about power. Anyway, you put more power through a chip, it will run hotter, which circles back to his original statement, doesn't it?