Discussion Intel current and future Lakes & Rapids thread

[DrMrLordX]

Max boost 3.1 GHz?

wat

 

[Ajay]

Max boost 3.1 GHz?

wat

Looks like a geekbench problem - since max boost == base clock.

 

[A///]

Its the Xeon W5 3435X So 16 cores spread across 4 tiles, wich is 4 cores per tile. If you dont care about AVX512 but want 8 Channel and Lots of PCI express lanes then the 5955WX is just as fast as the 3435X and should be cheaper too.

Thanks I've no real reason to buy either as I prefer higher clocks for all around use which I why I've stuck to mainstream systems. I did run Xeons and Opterons in the back past but that and having to own a mainstream system or two was too much.

the last such system I owned was when disc drives were still common and used until just as they were beginning to be rapidly phased out of even prebuilts. they have their purpose, but not to mine.

I will of course be open to donations of working, brand new 64 or 96 core Threadrippers for Zen 4 or high end Xeon workstations if anyone is feeling particularily charitable.

 

[nicalandia]

Max boost 3.1 GHz?

wat

Geekbench Browser

 

[nicalandia]

That must be some High End Water Cooler...

https://twitter.com/x/status/1636513303148249088

Water Cooled Threadripper can do the same with only 691 Watts, thats 234 Watts less

 

[A///]

larger surface area = easier to cool. noctua's xeon hsf cools up to 750 watts. the size isn't standard. standardizes hsfs for servers have always sucked and water cooling at high end fancy shmancy data centers is the default.

 

[DrMrLordX]

Looks like a geekbench problem - since max boost == base clock.

Yeah looks like it. According to @nicalandia 's reply, it looks like it gets closer to 4.4 GHz.

 

[soresu]

Water Cooled Threadripper can do the same with only 691 Watts, thats 234 Watts less

Lol Zen4 TR is just going to rip that a new back passage when it lands 😆🔥

 

[DrMrLordX]

Lol Zen4 TR is just going to rip that a new back passage when it lands 😆🔥

That's cool and all, but given AMD's track record, that's gonna take awhile.

 

[nicalandia]

That's cool and all, but given AMD's track record, that's gonna take awhile.

And Intel is not doing its part to preasure AMD with Sapphire Rapids Lacklustre performance.

 

[A///]

That's cool and all, but given AMD's track record, that's gonna take awhile.

Should be early 2h 2023. June to september.

 

[DrMrLordX]

Should be early 2h 2023. June to september.

Interesting, and surprising if true. A bit off-topic tho. Sadly the Sapphire Rapids Xeon W products appear to be stillborn.

 

[nicalandia]

Interesting, and surprising if true. A bit off-topic tho. Sadly the Sapphire Rapids Xeon W products appear to be stillborn.

They were supposed to go against Threadripper 3000 series.

 

[nicalandia]

That's on Huge Expensive Die.

 

[Abwx]

Water Cooled Threadripper can do the same with only 691 Watts, thats 234 Watts less

View attachment 79541

At worst 48 Zen 4 cores should score 108k@450W and 64 cores will get to 120k@350W, expect desperate marketing from Intel as their offering is just disastrous in comparison of the competition.

 

[Geddagod]

That must be some High End Water Cooler...

View attachment 79539

View attachment 79540

https://twitter.com/x/status/1636513303148249088

Water Cooled Threadripper can do the same with only 691 Watts, thats 234 Watts less

View attachment 79541

Is Intel's chiplet cost really that high?
A monolithic 24C SPR vs 24C Zen 3 is like 15% more efficient in CB23, but a 56C SPR uses >30% more energy for the same score as a 64C Zen 3 TR? Yes the TR has more cores, but it still shouldn't be this bad.

 

[Geddagod]

That's on Huge Expensive Die.

View attachment 79560


View attachment 79561

Icelake or SPR MCC?

 

[nicalandia]

Is Intel's chiplet cost really that high?
A monolithic 24C SPR vs 24C Zen 3 is like 15% more efficient in CB23, but a 56C SPR uses >30% more energy for the same score as a 64C Zen 3 TR? Yes the TR has more cores, but it still shouldn't be this bad.

Intel AIB interconnect does consume more energy than monolithic

 

[nicalandia]

Icelake or SPR MCC?

SPR MCC

 

[BorisTheBlade82]

Is Intel's chiplet cost really that high?
A monolithic 24C SPR vs 24C Zen 3 is like 15% more efficient in CB23, but a 56C SPR uses >30% more energy for the same score as a 64C Zen 3 TR? Yes the TR has more cores, but it still shouldn't be this bad.

Intel AIB interconnect does consume more energy than monolithic

Yes, it is of course more than monolithic but there is more to it than might meet the eye at first sight:
Intel had this idea of a unified L3 cache across several tiles. That is why there are huge amounts of traffic across EMIB for next to no benefit. AMD OTOH has much less bandwidth via IFoP at much higher power per bit transferred to begin with. But because their L3 is exclusive per CCD, there is only a tiny bit of coherency traffic going on. And this is why they are much more efficient.
Of course there might be edge cases - but it is more than questionable why Intel took that design direction in the grand scheme.

 

[igor_kavinski]

Intel had this idea of a unified L3 cache across several tiles.

Performance improvement for DB workloads? Maybe also some special confidential workloads by their large customers that depend on this design decision?

AMD will attack this problem with their 1152 MB of 3D cache on Genoa-X so maybe Intel had no choice but to go the way they did otherwise it would have meant losing their important customers to AMD.

 

[BorisTheBlade82]

Performance improvement for DB workloads? Maybe also some special confidential workloads by their large customers that depend on this design decision?

Yes, who knows - at least not me, I promise.
But all the reviews show that for common workloads their design simply seems to be inferior.

 

[igor_kavinski]

But all the reviews show that for common workloads their design simply seems to be inferior.

Have you seen any HammerDB server benchmarks recently? HammerDB is relatively easy to use to find out the OLTP and OLAP performance of a server. If they kept the storage array similar between the servers, it would provide a good comparison point. Another probably better way is to create a huge 100GB RAMDRIVE and run HammerDB on that and you get an idea of the complete performance profile of the CPU+Memory subsystem of a particular server platform.

 

[Shmee]

Is Sapphire Rapids available yet? I thought they were supposed to be out by now...I am interested in them or a TR competitor due to more IO.

 

[Exist50]

Yes, it is of course more than monolithic but there is more to it than might meet the eye at first sight:
Intel had this idea of a unified L3 cache across several tiles. That is why there are huge amounts of traffic across EMIB for next to no benefit. AMD OTOH has much less bandwidth via IFoP at much higher power per bit transferred to begin with. But because their L3 is exclusive per CCD, there is only a tiny bit of coherency traffic going on. And this is why they are much more efficient.
Of course there might be edge cases - but it is more than questionable why Intel took that design direction in the grand scheme.

I wish we had some more data on SNC4 vs SNC-OFF modes. Would give a lot more insight on what workloads, if any, benefit.