cancel
Showing results for 
Search instead for 
Did you mean: 

Microarchitect Analysis " AMD ZEN 5 " VS " INTEL LION COVE "

restsugavan
Level 14

ZEN5MICROARCHITECT.jpg

AMD ZEN 5 MICROARCHITECT DIAGRAM FROM HOTCHIP2023

New Intel Lion Cove Microarchitect @ Computex Taipei 4 June 2024 Pat Gelsinger Annouced.New Intel Lion Cove Microarchitect @ Computex Taipei 4 June 2024 Pat Gelsinger Annouced.

 

INTEL LION COVE MICROARCHITECT 48KB L0 192KB L1 Data with 3 MB per Core L3 Cache all 512-bit wide

Not a long time from here Computex 2024 Taipei will starting. This year there are battle around 2 camps microarchitect product.

AMD serving us with ZEN 5 microarchitect the first fully AVX512 implement platform for it products. By the way Intel going to send

3 Microarchitect to challenge with AMD. The first was Raptor Cove one , second is Redwood Cove and Lion Cove.

Look like AMD ZEN 5 microarchitect can be braking Intel 3 microarchitect at once.

What did your opinion about Intel for these situation ? 

W11 24H2 26252.5000 Core i9 7980XE 02007108 MCE ME 11.12.96.2535 R6E Modified BIOS 4001 SAMSUNG OG9 FW 1019.0 SSD 970 EVO PLUS 1 TB x 3 NVIDIA RTX 4090 GAME READY 560.70 64GB GSKILL DDR4 3200MHz JBL 9.1 Sound Bar DTS-X
488 Views
7 REPLIES 7

Int8bldr
Level 12

Thanks for the architecture charts. Very informative!
In my view, Intel is far ahead of AMD in terms of AI and capable vector instructions and AMD is trying to catch up (very late in the AI game).
I think Zen5 is somewhere between Sunny Cove and Golden Cove (from a vector instruction / AI instruction capability perspective). Zen5 supports BF16 (which Sunny Cove does not) but Zen5 does not support FP16 and AMX (BF16 and INT8 matrix multiplications = core in AI!!) which Golden Cove (used in Saphire Rapids CPUs) does. With the Redwood Cove micro architecture which sits inside the Intel Granite Rapids CPUs (coming this year), the gap is even bigger with support even for vectors of complex FP16 (in the D model) etc.
If intel can ship 64 up to 128 core Granite Rapids CPUs in volume this year, then Intel has a winner otherwise AMD ZEN5 may still beat Intel by having more cores (even if each core is less capable) since AMD will likely have 128 core ZEN5 CPUs.

just my personal speculation. I think intel did the right bet on AI and was early with vector instructions to accelerate those critical workloads but somehow, they got persuaded by the market to deemphasize AVX-512 and VNNI in favor for more cores and higher clock frequency to compete with AMD's more core approach. But then the game shifted to AI and then the competition is NVIDIA, not AMD... so they were squeezed...

I have high hopes for Redwood Cove inside Intel Granite Rapids.

Int8bldr
Level 12

Also, the fact that ZEN5 is now, in 2024 embracing AVX-512 the same way Intel did with Skylake in 2017, validates that Intel was right all along to bet in 512-bit wide data paths and 512-bit wide vector instructions in the form of AVX-512. Intel should have stuck to that and not deprecated it for consumer desktops the way they did in Alder Lake onwards... I have a Tiger Lake laptop that runs AVX-512 enabled s/w super fast at high clocks while staying cool.

@Int8bldr  Hi friend.

It look like Intel was sit-in seriously scenario now. ( Look at Pat's face on this video ) and would you please see the 2nd  video below

Why did Pat's interview by Bloomberg so seriously ? This video below may concern for him on Microsoft's CEO Sattaya Nadella speaking next to first 15 minute of the folks.

 

W11 24H2 26252.5000 Core i9 7980XE 02007108 MCE ME 11.12.96.2535 R6E Modified BIOS 4001 SAMSUNG OG9 FW 1019.0 SSD 970 EVO PLUS 1 TB x 3 NVIDIA RTX 4090 GAME READY 560.70 64GB GSKILL DDR4 3200MHz JBL 9.1 Sound Bar DTS-X

restsugavan
Level 14

Deep Drive to Lion Cove and Skymont Microarchitect

W11 24H2 26252.5000 Core i9 7980XE 02007108 MCE ME 11.12.96.2535 R6E Modified BIOS 4001 SAMSUNG OG9 FW 1019.0 SSD 970 EVO PLUS 1 TB x 3 NVIDIA RTX 4090 GAME READY 560.70 64GB GSKILL DDR4 3200MHz JBL 9.1 Sound Bar DTS-X

Int8bldr
Level 12

Thank you! Very interesting!

With Lunar Lake, it looks like Intel is very much influenced by what Apple is doing with their CPUs, integrating a bunch of different types of specialized function blocks and memory on silicon, e.g. like in Apple M3 and the Snapdragon X Elite.

Looks like they can get 120-140 int 8 TOPS out of the Lunar Lake - very narrow area of TOPS but still 120 INT8 TOPS. works for inference in some AI workloads but not so much for training and general AI. Still pretty good for a modern consumer desktop/laptop top the next say few years. Interesting to see how the NPU are designed! Very capable! and the new Xe matrix extension engine (XMX) running INT8, INT4 and INT2. Thats how they get to the 120 TOPS. but INT4/2 is not FP32 and therefore not as versatile and general. I rather have 100 FP32 TOPS honestly... (10980XE can already do 40 INT8 TOPS and 10 FP32 TOPS). 

I also saw they launched the Xeon 6700E “Sierra Forest” CPU series with up to 144 cores. But those are e-cores so not interesting for me on a desktop/WS... still 144 cores is pretty nice to see. 

I'm waiting for the Granite Rapids P cores (Redwood cove) or maybe even Lion cove P-core based CPU with say 64 cores - I think they will be called Diamond Rapids! That's where the real general performance deal is going to be! AMD is far behind. I'm hoping we can get 200-400 INT8/FP16/BF16 and 100 FP32 out of a new Redwood cove P core Granite Rapids 64 core HEDT CPU - I hope this will be launched H2 2024 or H1 2025? or maybe hold out for Diamond Rapids...

@Int8bldr Totally agreed 👍 

We'll need to wait. If Arrow Lake did a great job like Core 2 Duo. Perhaps LGA 4710 will be the next " Rampage 7 Extreme " indeed. New technology such as CAMM.2 will be replaced our DIMM.2 like new LGA 1851 practice as new standards. Diamond Rapid X with at least 128 Core Lion Cove Cores will be our new rigs mate . Can't wait for Diamond Rapid X Series LGA 4710 many kind of features are here.

2024-06-04_8-51-42.jpgL0 Cache 2 x 512-bit wide bus 48KB Size L1 D-Cache 2x 512-bit wide Upto 192KB Size L2 Cache 2x512-bit wide Upto 3072 KB Size per Core   


2024-06-04_8-51-26.jpg

 6 x 256-bit wide ALU 3 x 256-bit JMP 3 x 256-bit MUL 

2024-06-04_8-51-25.jpg

 Mega overhaul microarchitect on execution unit config

2024-06-04_8-51-20.jpg

 Easier overclocking with new 16.67 MHz interval tuning support 

2024-06-04_8-51-12.jpg

 Lion Cove P-Core natively support Intel Hyperthreading next generation bring +30% per core performance uplift

2024-06-04_8-51-18.jpg

 Intel 18A Node Process bring Intel back in games of performance efficienty.

 

W11 24H2 26252.5000 Core i9 7980XE 02007108 MCE ME 11.12.96.2535 R6E Modified BIOS 4001 SAMSUNG OG9 FW 1019.0 SSD 970 EVO PLUS 1 TB x 3 NVIDIA RTX 4090 GAME READY 560.70 64GB GSKILL DDR4 3200MHz JBL 9.1 Sound Bar DTS-X

Int8bldr
Level 12

I agree, Lion Cove looks extremely impressive! That's the one to go for IF we can get a HEDT version it in 2025. Massive improvements at all levels: the Lion Cove micro architecture, the 18A node size -> 128 core should be possible in a HEDT format. Could be the Rampage VII extreme.