AMD Announces Ryzen 7000 Series "Zen 4" Desktop CPUs - Linux Benchmarks To Come

[yump]

TDP (Thermal Design Power) is not about power draw, it is about the capacity of the cooler needed to cool the cpu....

The capacity of a cooler is not measured in watts, but in watts per kelvin (or degree Celsius).

Power draw and heat output are the the same down to a fraction of fraction of a percent. This is widely known, and historically the power draw of a chip has been less than or equal to it's thermal design power in almost any workload. But recently (well, semi-recently), Intel and AMD have come up with convoluted arguments that allow them to lie like Jedi about how efficient their CPUs are.

Intel's scheme is that there is a long term power limit that will maintain the average power less than or equal to the TDP over tens of seconds. But most motherboard manufacturers disabled that limit by default or strongly suggested disabling it in the UI. In the last generation even that fig leaf is abandoned, and Intel's guidance is for motherboards to run the 12900K at 241 W indefinitely.

AMD's scheme is that if you use a cooler with the capacity specified, at the ambient temperature specified, the chip's thermal throttling mechanisms will cause the power draw to settle in at about equal to the TDP over the long term. The "capacity specified" is goosed such that the TDP is a factor of 1.35 less than the actual power the chip will use when sufficiently cooled that the thermal throttling mechanisms do not engage.

So when AMD says the 7950X has a TDP of "170 W", it is reasonably safe to assume that with a good cooling system it will pull the full 230 W the AM5 socket is rated for.

[Linuxxx]

I'm excited that AMD is actually pushing AVX-512 on regular desktops even if it's not a full-width core. Finally a reason to upgrade from Haswell!

True, as it seems that Intel's 13th gen. Raptor Lake will still be missing out on AVX-512 because of their E-cores:

Raptor Lake's E-Cores still do not support AVX-512, so we expect that Intel will keep the feature disabled, which is odd given that AMD's Ryzen 7000 will fully support the extensions.

FWIW, hopefully they follow AMD's lead here and bring it back as the same 2 * 256-bits width wide registers, else going with future AMD CPUs becomes a no-brainer...

[coder]

hopefully they follow AMD's lead here and bring it back as the same 2 * 256-bits width wide registers,

I predict that will happen in Meteor Lake, which is allegedly due out by late 2023.

[CochainComplex]

I guess with Clear Linux it might be possible to bench the possibilities. AVX512 is quite well implemented where it is useful and possible. Hopefully Clear Linux does correct dispatching by recognizeing the AMD as AVX512 capable

[ayumu]

I predict that will happen in Meteor Lake, which is allegedly due out by late 2023.

Intel better have something decent by then. They'll be competing with Zen5, which is supposedly the largest redesign since the first Zen.

[birdie]

I'm excited except for DDR5 prices which means I'm not upgrading yet unless I can buy DDR5 6400 CL30 32GB modules for under $250 and then I live in a country which is under heavy sanctions, so it's a pipe dream anyways.

[coder]

I'm excited except for DDR5 prices which means I'm not upgrading yet

I'm still waiting for ECC DDR5 UDIMMs even to be available.

If they were, I might've upgraded to Alder Lake, but now I'm likely to wait and see how Raptor Lake matches up against Ryzen 7000.

[birdie]

I'm still waiting for ECC DDR5 UDIMMs even to be available.

If they were, I might've upgraded to Alder Lake, but now I'm likely to wait and see how Raptor Lake matches up against Ryzen 7000.

ECC is certainly nice to have but I don't have anything mission critical, so I don't care much. Secondly, from what I've seen with previous generation RAM standards, all ECC RAM is 1) clocked significantly lower 2) with a lot higher latency 3) priced a lot higher which makes this RAM very unappealing for a normal user who just needs the most performance for the least amount of money.

[coder]

ECC is certainly nice to have but I don't have anything mission critical, so I don't care much. Secondly, from what I've seen with previous generation RAM standards, all ECC RAM is 1) clocked significantly lower 2) with a lot higher latency 3) priced a lot higher which makes this RAM very unappealing for a normal user who just needs the most performance for the least amount of money.

AFAIK, there's no fundamental reason for any of that. ECC UDIMMs just have an extra RAM chip (or 2, for DDR5) per rank. All of the work to check & generate the parity information is handled inside the CPU's memory controller.

I think the main reason ECC UDIMMs tend to be lower-spec is just because they cater to a more conservative customer base who doesn't/can't overclock (because most boards supporting ECC don't allow it) and is willing to trade a bit of performance for more stability.

[numacross]

AFAIK, there's no fundamental reason for any of that. ECC UDIMMs just have an extra RAM chip (or 2, for DDR5) per rank. All of the work to check & generate the parity information is handled inside the CPU's memory controller.

I think the main reason ECC UDIMMs tend to be lower-spec is just because they cater to a more conservative customer base who doesn't/can't overclock (because most boards supporting ECC don't allow it) and is willing to trade a bit of performance for more stability.

Does Alder/Raptor Lake even support dedicated ECC? I was under the impression that mainstream Intel parts (excluding some Celeron/Pentium/i3) do not support it. DDR5 has in-chip ECC, but that's a different beast.

The reason for ECC UDIMMs having strict JEDEC speed and timings is platform validation. Both CPU and motherboard vendors usually only specify/test with JEDEC specs.