Tuesday, September 2nd 2025

Intel "Nova Lake-S" Engineering Sample Runs 52 Cores at 4.8 GHz

Intel's upcoming "Nova Lake-S" introduces a completely new microarchitecture and under-the-hood enhancements. However, the appearance of the latest shipping manifests indicates that Intel is ensuring its partners receive early samples to make necessary software/firmware optimizations. And now, it appears that the 52-core top SKU will ship with speeds that may exceed 5.0 GHz. An early engineering sample confirms the appearance of test hardware in partner shipments, which means that Intel is already moving validation pieces to OEMs and cooling vendors so they can start tuning systems well ahead of qualification samples and mass production. The 52-core part is a tiled design that pairs two compute tiles with a single SoC tile. Each compute tile houses eight P-Cores based on Coyote Cove and 16 efficient E-Cores based on Arctic Wolf, while the SoC tile contributes four low-power Arctic Wolf LPE-cores. That gives Nova Lake-S a 16 P-Core, 32 E-Core, and 4 LPE-Core configuration totaling 52 cores for the top SKU, with single-tile variants offering a 28-core option.

Intel will also offer big-cache bLLC variants aimed at workloads that benefit from larger on-die caches. On the platform side, Nova Lake-S will use the LGA-1954 socket with a 45x37.5 mm footprint, which matches LGA-1851 sizing and should preserve broad cooler compatibility even if some vendors need offset plates or other tweaks for optimal contact. The P-cores will ship without SMT, following the same choice made for Arrow Lake, while Intel's cache-first bLLC approach looks intended to counter AMD's stacked 3D V-Cache strategy. For now, the 4.8 GHz reading comes from an engineering sample and Pre-QS shipments; expect clock speeds to be refined upward as QS samples roll out and Intel moves closer to a formal launch. Given Intel's ambition to re-capture a significant part of the desktop market share, we can expect to see boost clock speeds north of 5.0 GHz.
Intel Nova Lake
Sources: NBD, gamma0burst, via Wccftech
Add your own comment

35 Comments on Intel "Nova Lake-S" Engineering Sample Runs 52 Cores at 4.8 GHz

#26
phubar
phanbueywe really just need faster ram/unified memory - x3d is basically a ram bottleneck workaround.
RAM has pretty much always lagged CPU requirements since at least the late 80's.

They're also mostly out of headroom on the RAM process tech. Any improvements are getting real expensive too. Look at the prices of DDR5 8000 vs DDR5 6000 for example. Its a big issue with those Intel systems needing the real expensive $200+ RAM kits that also require several days of tuning + overclocks just to match stock Zen5 performance in many things.

Meanwhile any cheap DDR5 6000 kit is going to be solid with any X3D chip and you can largely skip any tuning so long as it has EXPO support.

That is whey they're looking at stuff like CAMM to replace DIMM. You get a bigger boost in memory performance by changing how it interfaces with the CPU rather than foundry improvements or design changes now.

So the reality is you're going to see AMD, ARM, Intel, Apple etc keep increasing caches and going towards something like a X3D more and more over the years. Or just soldering on RAM to the package. That is another option but its not really a cheap one so they don't like doing it.

And even the best tuned DDR5 is going to lose significantly vs a big L3 cache like X3D. Both in terms of bandwidth and latency.

So long as they keep prices sensible big L3's are a pretty good work around for lower bandwidth + higher latency main system RAM.
Posted on Reply
#27
phanbuey
phubarRAM has pretty much always lagged CPU requirements since at least the late 80's.

They're also mostly out of headroom on the RAM process tech. Any improvements are getting real expensive too. Look at the prices of DDR5 8000 vs DDR5 6000 for example. Its a big issue with those Intel systems needing the real expensive $200+ RAM kits that also require several days of tuning + overclocks just to match stock Zen5 performance in many things.

Meanwhile any cheap DDR5 6000 kit is going to be solid with any X3D chip and you can largely skip any tuning so long as it has EXPO support.

That is whey they're looking at stuff like CAMM to replace DIMM. You get a bigger boost in memory performance by changing how it interfaces with the CPU rather than foundry improvements or design changes now.

So the reality is you're going to see AMD, ARM, Intel, Apple etc keep increasing caches and going towards something like a X3D more and more over the years. Or just soldering on RAM to the package. That is another option but its not really a cheap one so they don't like doing it.

And even the best tuned DDR5 is going to lose significantly vs a big L3 cache like X3D. Both in terms of bandwidth and latency.

So long as they keep prices sensible big L3's are a pretty good work around for lower bandwidth + higher latency main system RAM.
Totally agree about the cost aspect. Apple has unified ram with a much wider bus - so they actually went the other way, they don't have x3d L3 but they do have a 256-bit LPDDR5X-8533 unified memory bus on a chip that only has about 16MB per P core cluster (4 cores) of L2 and no L3 at all (but get ready to pay $$$$ for a M4 with any decent amount of ram).

X3D really works for now, and is affordable and practical, but it can't fully overcome the ram being such a heavy bottleneck, and has moments where even that isn't enough -- overall feels a bit like slapping more and more cache on a spindle harddrive instead of upgrading to an SSD.

Those UE5 asset load stutters, are sysram trying to swap assets in/out of GPU ram - feels super awesome as we get 6090 and 10800x3d and you can now hit 500FPS average on insanely powerful hardware - and here's your 22 FPS .01% low asset stutter like a giant kick in the nuts every 10 seconds.
Posted on Reply
#28
Squared
phanbueyTotally agree about the cost aspect. Apple has unified ram with a much wider bus - so they actually went the other way, they don't have x3d L3 but they do have a 256-bit LPDDR5X-8533 unified memory bus on a chip that only has about 16MB per P core cluster (4 cores) of L2 and no L3 at all (but get ready to pay $$$$ for a M4 with any decent amount of ram).
Apple's 256-bit memory isn't about gaming performance on the CPU. The width together with high frequency allow a lot of bandwidth for the iGPU. AMD did a similar thing with Strix Halo. But for many CPU tasks, only the faster memory speed helps. Usually CPUs have enough memory bandwidth for the game, and latency is the bottleneck. (Faster clock speeds may increase the cycles of latency but since the time each cycle is shorter it still usually means less actual time waiting.) I suspect some games actually run better with single-channel memory, although I haven't seen that tested in a long time.
Posted on Reply
#29
Upgrayedd
From a normie, who are these for?

Workstation users? I thought they bought the 4 and 8 channel stuff?

Will these use the reverse HT patent stuff? And can that patent do more than just 2 cores together? Could you bring 8 e cores together in the single tile die for older games that only use 1 or 2 threads and they'd be faster than the P cores? Or just bring them together to form essentially a 10 or 12 core single tile CPU? Cause as a normie I don't really feel the need for 24+ cores or 32+ threads for gaming at the moment? If GTA6 needs it, alright, but currently, nah.
Posted on Reply
#31
randomdude9182
liveaxelThe rumor is Nova Lake will launch in H2 2026, and seeing engineering samples in the field about a year out makes sense given that timeline.

And yeah, 4.8Ghz boost seems pretty darn high for silicon this far away from release. Maybe NVL will be good!
Its probably super well binned and drawing 350W
Posted on Reply
#32
irisChronomia
phanbueywe really just need faster ram/unified memory - x3d is basically a ram bottleneck workaround.
They have increasingly complex cache designs probably because RAMs can't get faster; we used to have CPUs running at the same internal clock as the SDRAM it is connected to, afterall.
Posted on Reply
#33
phanbuey
irisChronomiaThey have increasingly complex cache designs probably because RAMs can't get faster; we used to have CPUs running at the same internal clock as the SDRAM it is connected to, afterall.
Especially true now that ram is unobtanium
Posted on Reply
#34
irisChronomia
phanbueyEspecially true now that ram is unobtanium
Hah hahahahah yeah. I wonder if more fabs would be allocated to RAM production if this shortage keeps on going.

imho the LLM companies can't actually buy up significant portions of DRAM supply, at least not in the long term - they're not yet turning profits, and money won't just keep falling from the sky forever. This might've been just wishful thinking, lol.
Posted on Reply
#35
Sol_Badguy
irisChronomiaimho the LLM companies can't actually buy up significant portions of DRAM supply, at least not in the long term - they're not yet turning profits, and money won't just keep falling from the sky forever. This might've been just wishful thinking, lol.
OpenAI is buying raw uncut wafers. The purpose is to RAM-block the competition, to ensure OpenAI doesn't lose the race to singularity or whatever the F the goal is.
OpenAI doesn't need to turn in profit it just has to make sure everybody else doesn't. It has the wafers stockpiled and held for ransom, whoever wants RAM has to pay way more than OpenAI paid for the wafers -> the difference is profit, but not from Skynet subscriptions so that every regular Joe can submit a PhD or a Hollywood movie script, nope instead from God tier flipping.
Oh but money will keep falling from the sky forever, the populace is paying their taxes, no?
Posted on Reply
Add your own comment
Nov 15th, 2026 18:50 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts