Is it worth purchasing the R6EE etc or holding for the next CPU range?

[N00b1nat0r]

Hi all,
I'm looking at upgrading from my R4F hex core to the R6EE with a 10/12 core, I'm not looking for PCIe gen4 as I will move over my 2x 1070's for cost reasons. I will be running 2x M2 drives minimal and with a couple of sata SSD's.

The reason I am looking at this board is for the larger memory capacity, m2 drive slots and with the x299 CPU's the pcie lanes to cover both dual SLI and the number of M2 drives.

My question is regarding with new gen CPU'S coming out is it better to try holding off going to an EOL board CPU or go for it?

I also have been looking against the QVL for RAM and most or all of the G.Skill memory isnt available anymore except for different latency times sticks. Does anyone know if these work with the R6EE board?

F4-4133C17Q-32GTZR
F4-4000C18Q-128GTZR
F4-4000C15Q2-64GTZR

[Int8bldr]

following the Intel news the last couple of days, I think it's clear that right now it's best to wait anther year for Sapphire rapids. if you can not wait, it's hard, but maybe you can find an used x299 board, e.g. R6EE and a core i9 on ebay or similar if budget is tight.

With regards to Alderlake, I think Intel is making a big mistake killing of AVX512 for consumer (they are fusing of AVX512 silicon in the performance cores effectively knee capping the Golden Cove arch). AVX512 just made it into consumer grade CPUs with icelake and tigerlake (even in laptops), and after 5 years (since launch of AVX512) there are now tuned libraries that take advantage of AVX512 to efficiently accelerates many workloads. So as far as I am concerned, I will avoid Alderlake altogether. I am actually quite pissed about this: Due to the pandemic, many corporation have re-platformed their employees on laptops with less and less stationary PCs (people working from home or flexing drives laptop trend) -> leads to that corporate s/w have to run on the laptops. With tigerlake they got AVX512 on laptop so s/w libraries tuned with AVX512 could run there and dev could target that h/w stack. but now with Alderlake it's gone (likely also on core iX Alderlake based laptops).
So I'm not upgrading anything to Alderlake period. And Sapphire rapids is the one to hold out for - specs look great. Hope there will be a 58 or even a 96 core HEDT version. hope there will be a workstation laptop CPU that keeps AVX512. Let's see what happens.

[Braegnok]

N00b1nat0r wrote:
Hi all,
I'm looking at upgrading from my R4F hex core to the R6EE with a 10/12 core, I'm not looking for PCIe gen4 as I will move over my 2x 1070's for cost reasons. I will be running 2x M2 drives minimal and with a couple of sata SSD's.

The reason I am looking at this board is for the larger memory capacity, m2 drive slots and with the x299 CPU's the pcie lanes to cover both dual SLI and the number of M2 drives.

My question is regarding with new gen CPU'S coming out is it better to try holding off going to an EOL board CPU or go for it?

I also have been looking against the QVL for RAM and most or all of the G.Skill memory isnt available anymore except for different latency times sticks. Does anyone know if these work with the R6EE board?

F4-4133C17Q-32GTZR
F4-4000C18Q-128GTZR
F4-4000C15Q2-64GTZR

You might consider AMD X570 Motherboard, 12-Core 5900X Vermeer CPU.

Asus Crosshair VIII Extreme. https://rog.asus.com/motherboards/rog-crosshair/rog-crosshair-viii-extreme-model/spec

AMD Ryzen 9 5900X reviews. https://www.techspot.com/products/processors/amd-ryzen-9-5900x-37ghz-socket-am4.225647/

Ryzen 5000 CPUs have a total of 24 PCIe Gen4 lanes. Four out of the twenty-four are used for the interconnect to the motherboard chipset, leaving 20-lanes Gen4 4.0 for other utilization. 16 lanes (PCIe x16) are intended for graphics cards that are connected as x16 or two at x8.

X570 PCH includes sixteen PCIe lanes, in addition to the 24 PCIe lanes on the CPU, with a total of 4 (four) 4-lane physical interfaces (4x PCIe 4.0 x4 PHY), fully configurable in PCIe x16, x8, x4, x2, x1, and SATA modes. Also the X570 chipset offers 8x USB 3.2 Gen 2, 4x USB 2.0, 4x SATA, and 8 separate PCI-Express 4.0 lanes. That leaves eight flexible lanes that can be used for PCI-Express and SATA.

The DMI link on X299 being limited to 4GB/s does make it a bit harder to utilize cards such as RAID adapters or high speed network adapters, but the boards have better assortment of x8 and x16 wired slots. The X570's 8GB/s PCH interconnect gives it more overhead for things like additional high speed NVMe M.2 drives, but the slot configuration will typically make it much less well suited for plugging in extra devices.

Raw bandwidth per lane is a nice thing to have, but typically expansion-happy workstations live and die by lane count. X570 makes a mean platform, but anyone working with compute/AI, VMs or servers will often need X299/C621 or X399/TRX40/Epyc. You can do monster X570 builds, but they require a bit of planning to account for the lane limitations.

[restsugavan]

Heart of Intel Sapphire Rapid the Golden Cove Microarchitecture.




Golden Cove Microarchitecture the first Intel 12 port Execution.

[Int8bldr]

restsugavan wrote:

[
Thank you for this useful info.
Yeah, Sapphire rapids is the one to hold out for. i am convinced about that. it looks very good!

On another note:
The more I read about Alderwood the less I like it.
My take:
They have knee capped the Golden Cove cores and added some sort of next gen atom cores (similar to sky lake) also know as crap cores to increase the package core count just to be able to say they have the same cores count as AMD. Market department driven product.
In addition, for the whole thing to actually work efficiently, you need a new scheduler => windows 11, which is unsurprisingly is launched around that same time.

Alderwood feel like a lap top product where this performance/efficient core model might be useful running off batteries, but on a desktop/workstation - I think not.

Especially considering AMD's use of quality cores across all cores at a high core count!

No Sapphire Rapids is the way to go for HEDT and workstation.

[restsugavan]

@int8bldr

I totally agree with you mate.



Sapphire Rapid is the way to go



Just waiting ASUS build Rampage 7 Extreme.:o

[Int8bldr]

Sapphire Rapids news from Hot Chips.
honestly a rehash of what was released on Intel Architecture Days the last week.
but here is a taste of what was presented at theHot Chips conference this week.

Comment:
Last slide on AMX vs AVX512.
8x speed up on INT 8 is not as impressive as I would have liked to see. this means that the AMX TMUL instruction takes double even triple digit number of clock cycles to execute because theoretically, it should be a speed up of something like 256x if using the same amount of clock cycles as AVX-512 corresponding VFMA instructions but we "only" get 8x. 32x is lost somewhere - probably by an expansion of number of clock cycles used for TMUL (4x32 = approx 128 cycles - assuming no memory latency impact - and that is a false assumption 😉 )

for the AVX-512 FP32 vs AMX BF16 comparison there is a 16x speed up but half of that comes from that BF16 is half size of FP32 so => also 8x speed up. we'll have to wait for more details.

overall the CPU looks super impressive in Data, IO and core improvements.

[restsugavan]

Hi @Int8bldr 😮

According to Golden Cove architect itself. Personally I think the AVX512 level that Golden Cove support must be more improve from
Willow Cove and Sunny Cove a lot. Its extend execution ports support 3 x 256-bit Load or 2 x 512-bit Load / 2 x 512-bit Store
which allow 512-bit data flying all architect better than our Skylake X ( 2 x 512-bit load / 1 x 512-bit store ).




These will help 2 x AVX512 instructions can be executing on the same time on it dual 512-bit FMA Unit of each core
Thus the AVX512 instruction will be more quickly process and double throughput from Skylake X.


Another jump on performance side consider from the Golden Cove




Intel add 5th ALU with LEA 1-cycle all 5 ports latency



Intel add FP16 FMA3 format support for AVX512
( These will reduce memory footprint require from L2 cache and lower latency executing )

from those keys element itself. Golden Cove will be big improvement IPC over Skylake X a lot.

[Int8bldr]

Thanks @restsugavan!
This is helpful explaining that the Golden Cove AVX-512 base line (in intel's comparison) has improved over Skylake AVX-512. So that explains some of the lost x32 improvement I was looking for. I was comparing Skylake AVX 512 to Golden Cove AMX.

Also, I dug a bit deeper into the Golden Cove AVX-512 base line., looking at the actual benchmark description off intel site. and there is addition explanation. they are comparing the AVX-512 VNNI instruction (a horizontal fused multiply and add instruction) vs AMX TMUL, just pure multiplication-add throughput per clock cycle AND only using a 8x8 INT8 tile size matrices. Wtih this info I can see why it's only 8x improvement with INT8 it's just a consequence of that AMX have 8x8 mult-adders executed simultaneously while VNNI has only 8 INT8 mult-adders. So in the benchmark AMX is engaging 8x more INT8 mult-adders. simple: therefore you get an 8x speed up (8x8/8 = 8).

However, this is maybe underselling the AMX since T-regs can be 1KByte large say 16x64 = 1024 INT8 in a tile. I wonder why they did not use the full 1024 Byte T-reg size in their demo, using the full power of AMX .

Speculating here, but maybe it's because they only have 8x8 =64 INT8 mult-adders in the Golden Cove core and while the architecture can have 1024 INT8 (16x64) values, the TMUL instruction can only operate groups of 8x8 INT8 at a time and then it moves on to the next group until the full tile 16x64 is executed? Still with this approach we should see significate more speedup than 8 over AVX-512 VNNI. so I am still very curious and cannot explain how AMX is architectured and why they picked the benchmark they did. Need to wait for more details...

Meanwhile I'm interested in what AMD will come up with rumor is that have implemented some sort of AVX-512 in the next Zen gen....

[restsugavan]

According to many source flying around internet.

AMD was already implement their AVX512 on ZEN 4 micro architect comparable with Sunny Cove level.














However the power utilized must be investigate compare to Intel the master of AVX512.
This was interesting how AMD doing to solve power requirement of AVX512 instruction used.






For more information about VIA Centaur AVX512 processing method see the link below.

Centaur Unveils Its New Server-Class x86 Core: CNS; Adds AVX-512 – Page 2 – WikiChip Fuse

By the way VIA (Cyrix for classic name) Centaur do AVX512 by design it as addition co-processor method. This method utilized power when the CPU need only. It was design to reduce power hungry when processing more complex AVX512 instructions. Although VIA Centaur only have a half throughput when compare to Intel Skylake X SKUs but it did not suffer from downing clock speed same as Intel CPUs at all.

AMD may choose this idea or not ? Time will tell us all. :cool:

[MirceaForce]

N00b1nat0r wrote:
Hi all,
I'm looking at upgrading from my R4F hex core to the R6EE with a 10/12 core, I'm not looking for PCIe gen4 as I will move over my 2x 1070's for cost reasons. I will be running 2x M2 drives minimal and with a couple of sata SSD's.

The reason I am looking at this board is for the larger memory capacity, m2 drive slots and with the x299 CPU's the pcie lanes to cover both dual SLI and the number of M2 drives.

My question is regarding with new gen CPU'S coming out is it better to try holding off going to an EOL board CPU or go for it?

I also have been looking against the QVL for RAM and most or all of the G.Skill memory isnt available anymore except for different latency times sticks. Does anyone know if these work with the R6EE board?

F4-4133C17Q-32GTZR
F4-4000C18Q-128GTZR
F4-4000C15Q2-64GTZR

if you will go for x299 I recommend the ROG RAMPAGE VI EXTREME OMEGA