Wednesday, February 11th 2026

Intel "Nova Lake" Compute Tile Die-sizes Surface

Intel is preparing to launch its next generation Core Ultra Series 4 "Nova Lake" processors in the second half of 2026. We have some hint on what the sizes of the Compute tiles of these chips could measure. "Nova Lake" remains a tile-based disaggregated chip, just like the current "Arrow Lake," with the most advanced foundry node being allocated only to the components that can benefit the most from it, the CPU cores. The CPU complex of the processor is located in the Compute tile, while low-power island E-cores are located on the SoC tile that's built on a slightly older foundry node.

Intel is building the Compute tile of "Nova Lake" on the TSMC N2 (2 nm nanosheet) foundry node. There will be different types of Compute tiles based on the CPU core count and cache sizes. The two most popular ones for the desktop platform will be the mainstream tile with 8P+16E core configuration and a standard L3 cache size shared among the eight P-cores and four E-core clusters. The second, more premium Compute tile type will be 8P+16E with bLLC, or big last-level cache. bLLC is an enlarged L3 cache that's estimated to be 3-4 times the size of the regular L3 cache. This is Intel's answer to AMD's 3D V-Cache. Estimates put the regular Compute tile to measure around 110 mm², while the premium variant with bLLC is estimated to measure over 150 mm². This 36% increase in die-area is entirely from the enlarged last-level cache.
This isn't the entirety of the "Nova Lake" chip, there will be other components, such as the SoC tile that contains the low-power island E-cores, the DDR5 memory controllers, the 74 TOPS NPU, and the PCI-Express Gen 5 root complex, besides the System Agent, Management Engine, and security processor. Besides the SoC tile, are the Graphics tile, which is likely to be built on a more advanced node than the SoC, but less advanced than the Compute tile. Intel has plenty of options for the Graphics tile, including its own Intel 4 foundry node.
Source: HXL (Twitter)
Add your own comment

55 Comments on Intel "Nova Lake" Compute Tile Die-sizes Surface

#51
XeonsNOpterons
YouseethisAlso you missed the news that intel is redirecting much of their cpu capacity to AI thus sending high end/workstation cpu prices to the moon. We should expect price increases by the end of the year, right?
Maybe, but unlikely to affect current workstation CPUs. They already announced their prices and they are much lower than comparable AMD Threadripper. Plus, I meant in the long term, once the AI craze insanity dies down, the renewed workstation competition from Intel will help pull AMD down from its laughing gas infused management charging $11000-13000 for the top end HEDT.
YouseethisZen 6 vs nova lake is going to be interesting because both architectures address the main weaknesses of the previous 4. For zen is the infinity fabric latency and ram speed and for nova is the low p core count and instruction set (maybe?)
On this point there should be no argument. This will be the most exciting showdown in over 5 years. Intel's troubles against AMD have been greatly exaggerated, but nonetheless they have indeed fallen behind in gaming. I personally think that if Intel can keep comparable costs, which they generally have, the larger die sizes is actually an advantage. People forget that how easy it is to cool a CPU/GPU is dependent on the heat power area density, not just the overall heat output. In this regard, especially with Sapphire Rapids-WS, cooling was relatively easy even with an air cooler despite the extreme power draws (still skeptical about AMD's low power draws though).

My bet is also that Diamond Rapids will take the lead over 256-core Venice even with the low end core count estimate of 192-cores.
Posted on Reply
#52
MufasaV
YouseethisOn avx 10.2, the 512 part is optional.
www.phoronix.com/news/Intel-Diamond-Rapids-APX-AVX10

As far as I understand, due to the limitations of the e cores, nova lake will have avx2+many of the instructions of avx512 bit not fp/int 512. The e cores might use 2x 256 units like zen4, who knows.

Also you missed the news that intel is redirecting much of their cpu capacity to AI thus sending high end/workstation cpu prices to the moon. We should expect price increases by the end of the year, right?
www.techpowerup.com/345535/intel-reallocates-pc-production-capacity-to-server-cpus-amid-tight-wafer-supply

Two Nova lake p cores will be as strong as 4e cores allegedly (p cores are clusters of two now with shared l2 like e cores are clusters of four with shared cache) and they do not have ht. We have 16p+32e ≈ 32p cores (+4lpe) without HT VS 24 cores with smt. Intel will have the multithread advantage probably but will not come cheap and the power consumption will be insane. (probably not that hard to cool though because the dies are large)

Zen 6 vs nova lake is going to be interesting because both architectures address the main weaknesses of the previous 4. For zen is the infinity fabric latency and ram speed and for nova is the low p core count and instruction set (maybe?)
Its no longer optional. And the E cores have been rumored since Novalake was revealed as a name to us (before AVX10.2 was a thing), to support full AVX512 by Raichu
Posted on Reply
#53
Youseethis
MufasaVIts no longer optional.
I just found out about it on intel newer revision of the spec. (avx 10.2 v4 and v5 both may 2025). I wonder if it affects nova lake though. Nova also has the lpe cores too. Are they compatible with avx 512 too? If they do not, can they be excluded? We shall see.
Posted on Reply
#54
Evrsr
WirkoNo, not really. In the usual flip-chip design, there's the entire substrate between the transistors and the heatsink. It can be thinned down but I don't know if manufacturers ever do that. However, they do that when stacking dies, which includes building the BSPDN structure.

Here's an illustration from Cadence. First is flip-chip, second is a chip with BSPDN (BSM = backside metal), third is a stack of both.
Si layer is not substrate. BEOL is substrate. Silicon layer is very conductive, somewhat closer to Aluminum, which is why flip chip worked better than not.

What I'm saying is not disputed, 18A has more hotspots:

"Realistic workload-based CPU powermaps reveal that nonuniform dissipation majorly accentuates hotspots in BSPDN based designs. 1 μm -resolution temperature simulations of an 80-core server SoC show ~ 14°C penalties relative to FSPDN caused by interplay of extraneous thermal resis-tance (BEOL + wafer bonding) and BSPDN heat spreading."
Posted on Reply
#55
XeonsNOpterons
MufasaVIts no longer optional. And the E cores have been rumored since Novalake was revealed as a name to us (before AVX10.2 was a thing), to support full AVX512 by Raichu
Awesome stuff, cannot wait to see it!
Posted on Reply
Add your own comment
Nov 17th, 2026 13:49 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts