AMD "Zen 4" Dies, Transistor-Counts, Cache Sizes and Latencies Detailed

[Dirt Chip]

58% transistors increase for 13% IPC increase. They are having a hard time increasing IPC.

IGP and AVX takes it tall.
It will get better in zen4+ and zen5 for sure

 

[Slesreth]

So integrated graphics is confirmed? I didn't hear anyone from AMD mention it during the presentation..

Yes it is, it's on the AMD spec pages.

The same specs are listed on all four processors.

​

Graphics Capabilities​

Graphics Model: AMD Radeon™ Graphics​

Graphics Core Count: 2​

Graphics Frequency: 2200 MHz​

GPU Base: 400 MHz​

 

[defaultluser]

The same specs are listed on all four processors.

​

Graphics Capabilities​

Graphics Model: AMD Radeon™ Graphics​

Graphics Core Count: 2​

Graphics Frequency: 2200 MHz​

GPU Base: 400 MHz​

does anyone have a comparison versus 5700g yet? will that ALSO HAVE TO WAIT UNTIL THE END OF THE MONTH?

 

[Count von Schwalbe]

I read somewhere that APU's (monolithic?) Will still be a separate part of the product stack. Also, the IGP of the main series is essentially for display output, not 3D performance.

Based in the info above, it should have around 560 GFLOPS of compute.

Vega 11, in the 5700G, has 2.048 TFLOPS. Expect around 1/4 the performance.

Obviously, the move to DDR5 could offset the reduction slightly, but it is best to wait for the APU's.

 

[Minus Infinity]

Zen 4c is Epyc only. Not meant for gaming at all. Its for cloud instances. Not everything made by humans is for gaming.

Possibly, but Zen 5c is reportedly coming to the new hybrid Zen 5 cpu's as the .little cores. I see no reason why AMD couldn't offer Zen 4c on desktop. It's basically a stripped down Zen 4 core, less cache etc and in insiders say it's performance is around 10-30% less.

 

[Sabotaged_Enigma]

Official website says it's 71 mm² Zen 4 chiplet and 122 mm² I/O die.

 

[HD64G]

If Zen5 has both Zen5+Zen4c cores in the same CPU, it could be a great combo for workstations (both desktops and notebooks) but not of any use to the average PC user. So, let's focus to Zen4 atm.

 

[LuxZg]

IGP and AVX takes it tall.
It will get better in zen4+ and zen5 for sure

IGP is not part of compute die, it's in I/O die.

 

[Wirko]

So, let's focus to Zen4 atm.

That will be harder than it seems. In this thread I can see that many TPU members have a sharp focus on distant future.

 

[TheoneandonlyMrK]

does anyone have a comparison versus 5700g yet? will that ALSO HAVE TO WAIT UNTIL THE END OF THE MONTH?

This 2 unit igp is for office use, and work offload, the 5700G would trounce it, that's like comparing a 6400X gpu to a 5600G

 

[marios15]

People arguing about 58% transistor increase on a 8(EIGHT CORE) die, then complaining about that translating into 13% IPC, which is measured on 1(ONE) core.
EDIT: logic fail
If you take 58% divide by 8 gives 7,25%
So 7% transistor increase per core for 13% IPC per core increase?
That's 186% gain per transistor increase.

Didn't they say ~50% multi-core perf increase over last gen at same power? Sounds pretty reasonable if you include the clockspeed improvements, double AVX execution, increased cache.

 

[LuxZg]

People arguing about 58% transistor increase on a 8(EIGHT CORE) die, then complaining about that translating into 13% IPC, which is measured on 1(ONE) core.
If you take 58% divide by 8 gives 7,25%
So 7% transistor increase per core for 13% IPC per core increase?
That's 186% gain per transistor increase.

Didn't they say ~50% multi-core perf increase over last gen at same power? Sounds pretty reasonable if you include the clockspeed improvements, double AVX execution, increased cache.

Your math and logic skills are lacking. If 8 core vs 8 core uses 58% more transistors, then 1 vs 1 core still uses 58% more transistors.

But you are somewhat correct that we shouldn't look at IPC vs transistors, because transistors also enable higher clocks, and possibly better multithreaded perf, so we should look at that ~50% overall performance boost.

That's still just half the truth because it's not on same production node, so part of frequency uptake is also due to 5nm.

So 58% transistors probably equate to ~30% performance.

But it certainly doesn't work the way you suggest, 58/8=7.25 is correct, but doing so with % is incorrect.

 

[marios15]

Your math and logic skills are lacking. If 8 core vs 8 core uses 58% more transistors, then 1 vs 1 core still uses 58% more transistors.

But you are somewhat correct that we shouldn't look at IPC vs transistors, because transistors also enable higher clocks, and possibly better multithreaded perf, so we should look at that ~50% overall performance boost.

That's still just half the truth because it's not on same production node, so part of frequency uptake is also due to 5nm.

So 58% transistors probably equate to ~30% performance.

But it certainly doesn't work the way you suggest, 58/8=7.25 is correct, but doing so with % is incorrect.

Yeah I missed a logic step there.

The process node helps but iirc the design needs to be changed to actually take advantage of the new node VF curve, otherwise you only gain density improvements

 

[DemonicRyzen666]

Slow avx 512, dual load 256 avx, again amd didn't learn their lesson with see4.1 and avx on bulldozer. There is a ton of problems with that implementation.

I see I was right
It is slow in CinebenchR23 because of this desgin choice. Although I remeber a lontime ago Intel's AVX has the ablity to push AVX's up from the old to highest newest version in the cpu to improve speed, athough it was also in that same pdf I read that it would take SSE4.1 to AVX. Mysterly after I asked a question abou that to a fameous intel Engineer the PDF disappered off from view.

 

[Oberon]

I don't know if you know this (actually, I'm quite confident that you don't), but Cinebench doesn't make use of AVX-512, so its implementation in Zen 4 has no direct impact on the architecture's performance in that benchmark.

 

[marios15]

I see I was right
It is slow in CinebenchR23 because of this desgin choice. Although I remeber a lontime ago Intel's AVX has the ablity to push AVX's up from the old to highest newest version in the cpu to improve speed, athough it was also in that same pdf I read that it would take SSE4.1 to AVX. Mysterly after I asked a question abou that to a fameous intel Engineer the PDF disappered off from view.

It's not completely slower...

Zen4's AVX512 Teardown - mersenneforum.org

Embargo has lifted! So here's that teardown I've promised! I won't get into how this happened, but AMD graciously sent me two test setups this year. A retail Zen3 setup back in January, and an engineering Zen4 setup in August. The motivation here is to let me optimize y-cruncher for these...

www.mersenneforum.org

 

[DemonicRyzen666]

It's not completely slower...

Zen4's AVX512 Teardown - mersenneforum.org

Embargo has lifted! So here's that teardown I've promised! I won't get into how this happened, but AMD graciously sent me two test setups this year. A retail Zen3 setup back in January, and an engineering Zen4 setup in August. The motivation here is to let me optimize y-cruncher for these...

www.mersenneforum.org

it's still non-native AVX512 whole reason for the high clockspeeds
a Native AVX512 would use more power be hotter & clock lower