AMD Ryzen 8000 "Granite Ridge" Zen 5 Processor to Max Out at 16 Cores

[Steevo]

I want lower latency to cache and the ability to turn speculative cache branching predictions on per application and child threads so we can overcome "security" penalties for known applications and regain the performance.

 

[sLowEnd]

E-core was invented by Intel for marketing purposes, so they can slap 16cores tag on their CPUs, E-cores are bad for consumer in every way, with chiplets its even worse.

E-cores are a good way to get more multithreaded performance out of a given area of die space. Intel's E-core takes up about 1/4 the space of a P-core.
I think AMD will eventually follow suit and start incorporating E-cores into their CPUs too in the future. We'll probably get a clearer idea of what they're working towards when we see what Zen 5c is like.

 

[Pepamami]

E-cores are a good way to get more multithreaded performance out of a given area of die space. Intel's E-core takes up about 1/4 the space of a P-core.
I think AMD will eventually follow suit and start incorporating E-cores into their CPUs too in the future. We'll probably get a clearer idea of what they're working towards when we see what Zen 5c is like.

E-core still kinda power hungry, plus I think amd got enough of "FX-experience"

 

[sLowEnd]

E-core still kinda power hungry, plus I think amd got enough of "FX-experience"

They're space-efficient, not power-efficient. IIRC they're actually less efficient than P-cores, perf/watt-wise.

 

[Wolverine2349]

They're space-efficient, not power-efficient. IIRC they're actually less efficient than P-cores, perf/watt-wise.

Since 4 e-cores takes the space of one P core, and they are less power efficient than P cores, does a cluster of 4 e-cores clocked at 4.5GHz take more wattage than 1 P core clocked at 5.6 to 5.7GHHz on Raptor Lake?

 

[Pepamami]

They're space-efficient, not power-efficient. IIRC they're actually less efficient than P-cores, perf/watt-wise.

so they are useless with chiplet way, since u cant use a lot of them, they will just hit power limit.
I think amd can just attach extra 10/14nm cheap zen+/zen2, and call it a day. (cheap and simple)
Adding small "idle" cores to existing CCDs is weird.

 

[ValenOne]

AMD desktop PC = 16 cores forever...

They need to work on their default power settings and make these chips not overvolted to the moon - regular CPU voltages, not SoC problem related.
Both AMD and intel are running their CPU's to their absolute limits, and not spending the time to tune them in at all.

All AMD's best received CPU's have been the ones without the insane balls to the wall attitude at stock.
And they made half of them OEM only after that for some reason.

Look at this for an example of how it feels some days, 63W to 144W with their OC, for what amounted to .03% FPS gains
That's what it felt like running a 5800x with PBO enabled, tons of power for nothing except R23 scores going up.
View attachment 296274View attachment 296275


Give us a solid single CCX design, slap on 3Dcache, but make it power efficient so it doesnt overheat and thermal throttle to sustain that performance without needing high end motherboards and cooling to go with it.

5800x3D smashed things so it's keeping up with the next gen hardware, but simply undervolting it changes it drastically for the better - and they're so over-volted at stock i can run -30 on curve optimiser AND undervolt it with an offset

For gaming on the AM5 platform, lower memory latency tuning has better results when compared to clock speed overclocking.

E-core still kinda power hungry, plus I think amd got enough of "FX-experience"

On a single thread, Intel's E-Cores are close to SkyLake-S. Intel's Skylake S beats AMD's Bulldozer/Piledriver.

For Core i9 13900K, Intel is effectively gluing 8 Zen 4 class CPUs with non-SMT 16 Zen 1 class CPU cores.

Both Zen 1 and E-Cores have multiple 128-bit hardware units that double pump 256-bit AVX-2.


Intel Gracemont's six X86 decoder design is superior when compared to AMD's Bulldozer/Pipeliver.

It's a pretty good little CPU core since the entire hardware is allocated for a single thread. Bulldozer splits its hardware resource into two threads.

Intel Gracemount will beat both AMD Jaguar and Bulldozer designs.

 

[AusWolf]

AMD desktop PC = 16 cores forever...


For gaming on the AM5 platform, lower memory latency tuning has better results when compared to clock speed overclocking.



View attachment 296457
On a single thread, Intel's E-Cores are close to SkyLake-S. Intel's Skylake S beats AMD's Bulldozer/Piledriver.

For Core i9 13900K, Intel is effectively gluing 8 Zen 4 class CPUs with non-SMT 16 Zen 1 class CPU cores.

Both Zen 1 and E-Cores have multiple 128-bit hardware units that double pump 256-bit AVX-2.


Intel Gracemont's six X86 decoder design is superior when compared to AMD's Bulldozer/Pipeliver.
View attachment 296458
It's a pretty good little CPU core since the entire hardware is allocated for a single thread. Bulldozer splits its hardware resource into two threads.

Intel Gracemount will beat both AMD Jaguar and Bulldozer designs.

Is beating Bulldozer/Jaguar with anything still a merit anno 2023?

 

[Wolverine2349]

Is beating Bulldozer/Jaguar with anything still a merit anno 2023?

Yeah exactly. I mean those Bulldozer/Jaguar cores were beyond horrible.

Beating them is not some accomplishment.

 

[ValenOne]

Is beating Bulldozer/Jaguar with anything still a merit anno 2023?

Cinebench R20 single thread
12900K's E-Cores = 410
Ryzen 7 2700X (Zen 1.x) = 411
Intel i5 8400 = 411
Threadripper 1950X (Zen 1.x) = 411

The latest fat X86 CPU cores almost double E-Cores' 410.

Cinebench R20 Scores (Updated Results)

Regularly updated Cinebench R20 Scores List, sortable by multi-core and single-core score with in-depth info on CPU rankings.

www.cgdirector.com

For Core i9 13900K, Intel effectively attached 16 Zen 1.0 cores (Threadripper 1950X) with 8-cores Zen 4.

 

[Wolverine2349]

Cinebench R20 single thread
12900K's E-Cores = 410
Ryzen 7 2700X (Zen 1.x) = 411
Intel i5 8400 = 411
Threadripper 1950X (Zen 1.x) = 411

The latest fat X86 CPU cores almost double E-Cores' 410.

Cinebench R20 Scores (Updated Results)

Regularly updated Cinebench R20 Scores List, sortable by multi-core and single-core score with in-depth info on CPU rankings.

www.cgdirector.com

For Core i9 13900K, Intel effectively attached 16 Zen 1.0 cores (Threadripper 1950X) with 8-cores Zen 4.

.
To be fair Intel P cores are actually faster than AMD Zen 4 cores at same clock speed. LIke 6% better IPC per CInebench R23

And Intel i5 8400 is way better is latency sensitive workloads than Gracemont. Cinebench is not latency sensitive even though IPC is similar.

 

[MWK]

Larger IPC gains I don't feel like amd has improved much to be honest I have a ryzen 9 16 core
3950x and it's amazing... All they will do is raise the L1 cache in zen 5 and then wow the IPC gains are huge. And some minor architectural changes so the zen cores can communicate with the IOD faster and blah blah blah.... And we as fools will fall for this crap

 

[A Computer Guy]

Larger IPC gains I don't feel like amd has improved much to be honest I have a ryzen 9 16 core
3950x and it's amazing... All they will do is raise the L1 cache in zen 5 and then wow the IPC gains are huge. And some minor architectural changes so the zen cores and communicate with the IOD faster and blah blah blah.... And we as fools will fall for this crap

My 3950x still feels amazing but my 5950x does feel more amazing. I can imagine the 7950x must feel ultra amazing.

 

[ValenOne]

.
To be fair Intel P cores are actually faster than AMD Zen 4 cores at same clock speed. LIke 6% better IPC per CInebench R23

And Intel i5 8400 is way better is latency sensitive workloads than Gracemont. Cinebench is not latency sensitive even though IPC is similar.

Nope, Cinebench R23 does not support AVX-512 i.e. refer to Blender.

For games,

AMD's X3D cache only shows the potential for Zen 4 cores when the memory latency issue is mitigated.

Core i9 13900K has 2 MB L2 cache per p-Core while Zen 4 has 1 MB L2 cache per core. Zen 5 has a 2 MB to 3 MB L2 cache per core. https://wccftech.com/amd-next-gen-z...-cache-design-larger-l2-cache-per-core-rumor/

Intel has the higher near 6 Ghz p-Core clock speeds.

Larger IPC gains I don't feel like amd has improved much to be honest I have a ryzen 9 16 core
3950x and it's amazing... All they will do is raise the L1 cache in zen 5 and then wow the IPC gains are huge. And some minor architectural changes so the zen cores can communicate with the IOD faster and blah blah blah.... And we as fools will fall for this crap

I have Ryzen 9 3900X and 7900X, the 3900X is aging.

Zen 2 has two loads and one store.
Zen 3 has three loads and two stores.

From Zen 2 to Zen 4, the reorder order buffer is larger for each release.

Zen 2 CPU core's second thread has store starved situation I.e. Zen 2's SMT is less robust compared to newer x86 cores.

 

[sLowEnd]

Is beating Bulldozer/Jaguar with anything still a merit anno 2023?

If it's done at a low enough power, it is IMO. There's a lot of potential for small, low power devices that don't suck.

 

[R0H1T]

Dozer/Cat cores shouldn't be grouped together! Dozer was horrible right till the end, Jaguar wasn't ~ if it weren't for Intel's contra revenue shens AMD may have gotten more market share with it & launched even better products down the line. Jaguar (Puma?) was almost on par with Intel's efficiency with a node disadvantage IIRC.

 

[AusWolf]

Cinebench R20 single thread
12900K's E-Cores = 410
Ryzen 7 2700X (Zen 1.x) = 411
Intel i5 8400 = 411
Threadripper 1950X (Zen 1.x) = 411

The latest fat X86 CPU cores almost double E-Cores' 410.

Cinebench R20 Scores (Updated Results)

Regularly updated Cinebench R20 Scores List, sortable by multi-core and single-core score with in-depth info on CPU rankings.

www.cgdirector.com

For Core i9 13900K, Intel effectively attached 16 Zen 1.0 cores (Threadripper 1950X) with 8-cores Zen 4.

That's what I mean. There's no need to compare to Bulldozer when current E-cores are already at Zen 1 level.

 

[Wolverine2349]

Nope, Cinebench R23 does not support AVX-512 i.e. refer to Blender.

For games,

AMD's X3D cache only shows the potential for Zen 4 cores when the memory latency issue is mitigated.

Core i9 13900K has 2 MB L2 cache per p-Core while Zen 4 has 1 MB L2 cache per core. Zen 5 has a 2 MB to 3 MB L2 cache per core. https://wccftech.com/amd-next-gen-z...-cache-design-larger-l2-cache-per-core-rumor/

Intel has the higher near 6 Ghz p-Core clock speeds.


I have Ryzen 9 3900X and 7900X, the 3900X is aging.

Zen 2 has two loads and one store.
Zen 3 has three loads and two stores.

From Zen 2 to Zen 4, the reorder order buffer is larger for each release.

Zen 2 CPU core's second thread has store starved situation I.e. Zen 2's SMT is less robust compared to newer x86 cores.

Well AVX512 another matter. Though without AVX512, Intel I think has like 6% better IPC on Raptor Cove than Zen 4.

Zen 5 will be better than Zen 4 in IPC though in all workloads. Zen 5 is supposed to have 20-25% better IPC than Zen 4 so stands to reason will have 14-19% better IPC than Raptor Cove in non AVX512 workloads and AVX512 forget it as Raptor Cove does not support it.

That's what I mean. There's no need to compare to Bulldozer when current E-cores are already at Zen 1 level.

Are current e-cores at Zen 1 level at same clock speed? Cause I have heard they are like Skylake. Isn't Skylake IPC still better than Zen 1?

 

[AusWolf]

Are current e-cores at Zen 1 level at same clock speed? Cause I have heard they are like Skylake. Isn't Skylake IPC still better than Zen 1?

I don't know, and I don't care, to be honest. If one chip is designed to run at 4 GHz, and another is designed to do only 3 GHz, there's zero reason to run them at the same clock speed.

 

[Mussels]

Virtually no use case requires the kind of power available even from the last two or three generations.

My suspicion is the CPU market for home will stagnate for many years.

Servers and workstations is where it will be at.

Only high FPS gaming, at any resolution.
Freesync/Gsync make that redundant when used correctly
Since they can run 2x or 3x FPS boosted up, FPS caps can give you 120FPS at 240Hz, while 121FPS can only be 121Hz
You also get the large framebuffer benefits in that situation, so you get double the length of time for a frame to reach the monitor - so it smooths out microstutter greatly.
I'm eagerly awaitng the next gen 4K displays so i can get back to that level of goodness, vs my current heavily tweaked 60Hz display

E-core was invented by Intel for marketing purposes, so they can slap 16cores tag on their CPUs, E-cores are bad for consumer in every way, with chiplets its even worse.

purely because AMD was beating them in multi threaded benchmarks, they call them E-cores when they're less efficient in every way than the P cores


On the conversation a few posts up:

12th gen E-cores had the efficiency of 10th gen intel/Zen2

I have Ryzen 9 3900X and 7900X, the 3900X is aging.

Agreed, my 3700x seems to be capped around 120FPS in a lot of titles
RAM setup definitely helps them out, running four ranks and low timings seems key since they can rarely clock RAM up high
It all depends if that's enough for the user

 

[tugrul_SIMD]

I have ryzen 7900 all cores 5.3GHz. Tested it on my own mandelbrot set generator benchmark and it run 15% faster than skylake dual avx512 pipelines. Skylake had 7 cycles per pixel, ryzen had 6 cycles per pixel. But only GCC v12 could compile it efficiently. https://github.com/tugrul512bit/VectorizedKernel

I have ryzen 7900 all cores 5.3GHz. Tested it on my own mandelbrot set generator benchmark and it run 15% faster than skylake dual avx512 pipelines. Skylake had 7 cycles per pixel, ryzen had 6 cycles per pixel. But only GCC v12 could compile it efficiently. https://github.com/tugrul512bit/VectorizedKernel

And funny thing is that Intel OpenCL Runtime uses Ryzen CPU better than Amd 's runtime.

 

[Nhonho]

More of the same again, AMD? AMD needs to launch new things, come up with new ideas, it needs to get rid of old thoughts, old ideas, to be able to sell its chips well.

AMD has made a number of blunders in recent years:
- didn't put RDNA1 iGPU in their latest APUs (kept insisting with old, inefficient and power hungry VEGA iGPUs until Ryzen 5000G)
- did not put support for DDR4 memory on Ryzen 7000
- did not put AV1 encoder on the iGPU of Ryzen 7000
- made the recent GPUs of RDNA3 cards in MCM scheme, which greatly increased the latencies and, thus, decreased the performance and, therefore, AMD had to increase the clock and consequently the electrical consumption of these GPUs so that they have competitive performance.




AMD needs to break old thinking and do at least the obvious:

- AMD must "sit down" with software developers (from HandBrake, Avidemux, Adobe, Cyberlink, MAGIX, Blackmagic Design, Apple, etc.) to get its video encoder (from their GPUs and iGPUs) to do the video conversion (in H.264, H.265 and AV1 codecs) in 2 steps. The first major chipmaker (Intel, AMD, Nvidia) to do this will sell chips like water in the desert...
- AMD's video encoder must achieve the same image quality as Nvidia's video encoder.
- Make only single-die GPUs so they have lower latencies and, thus, higher performance.
- All AM5 motherboards had to support BIOS update without the CPU in the socket. So people could buy AM5 motherboards without worrying about having to go through the hassle of building a PC and it not turning on because the motherboard's BIOS doesn't recognize the CPU.
- Need to launch single-die Ryzen CPUs (SoC) so they have higher performance. To launch CPUs with more than 8 cores, it would be enough to put chiplets together, as it already does today.

 

[chrcoluk]

I want lower latency to cache and the ability to turn speculative cache branching predictions on per application and child threads so we can overcome "security" penalties for known applications and regain the performance.

That is actually a really good idea. Could add it to exploit protection settings in windows, with some kind of automatic algorithm which can be overridden, so high risk apps its off, low risk its on.

 

[Pepamami]

purely because AMD was beating them in multi threaded benchmarks, they call them E-cores when they're less efficient in every way than the P cores


On the conversation a few posts up:

12th gen E-cores had the efficiency of 10th gen intel/Zen2

I am expecting E-core to be slower than P-core, but also I am expecting E-core to be alteast(!) equal or less power hungry than P-core, sooon Intel will achieve it, but until then, E-cores are lame.

And my complain was more about the "naming" of cpu, when they call it 12Cores, and not 8P/4E

 

[R0H1T]

so they are useless with chiplet way, since u cant use a lot of them, they will just hit power limit.
I think amd can just attach extra 10/14nm cheap zen+/zen2, and call it a day. (cheap and simple)
Adding small "idle" cores to existing CCDs is weird.

And how will that work? They're made for DDR4, IF is also probably not compatible(?) & they don't have PCIe 5.0 although that's not much to lose. IF is one of major reasons IMO they can't go back to older cores otherwise they'd be selling lots of them right now!