Tuesday, September 27th 2022
Intel 13th Gen Core "Raptor Lake" Desktop Processors Launched: +15% ST, +41% MT Uplift
Intel today launched its 13th Gen Core "Raptor Lake" desktop processors, and companion 700-series motherboard chipset. These processors are built in the same LGA1700 package as the previous generation "Alder Lake," and are backwards-compatible with 600-series chipset motherboards through a BIOS update. Likewise, 700-series chipset motherboards support older "Alder Lake" processors. With the new 13th Gen Core, Intel is broadly promising an up to 15% uplift in single-threaded performance, which has a bigger bearing on gaming performance; and an up to 41% multi-threaded performance uplift; over the previous-generation, when comparing the top Core i9-13900K with its predecessor, the i9-12900K. Intel also claims to have outclassed the AMD Ryzen 9 5950X in multi-threaded performance, and the Ryzen 7 5800X3D in gaming performance.
Intel's performance claims are backed by some impressive hardware changes despite the company sticking with the same Intel 7 (10 nm Enhanced SuperFin) foundry node as "Alder Lake." To begin with, the single-thread performance uplift comes from the new "Raptor Cove" performance-core, which promises an IPC uplift over the previous-generation "Golden Cove," comes with more dedicated L2 cache of 2 MB per core (compared to 1.25 MB per core in the previous-generation); and significantly higher clock-speeds, going all the way up to 5.80 GHz. "Raptor Lake" has up to 8 P-cores, but the company has put in a lot of work in improving the contribution of E-cores to the processor's overall multi-threaded performance uplift. This is achieved by doubling the E-core count to 16. These are the same "Gracemont" E-cores as previous-generation, but Intel has doubled the L2 cache that's shared in a 4-core Gracemont cluster, from 2 MB per cluster to 4 MB. There are upgrades to even the hardware prefetchers of these cores.
Intel didn't go into the nuts and bolts of what makes up the "Raptor Cove" P-core, but broadly explained that it comes with improved speed paths that enable an up to 600 MHz P-core boost frequency uplift at comparable power to the previous-gen "Golden Cove" while staying on the same process. The Intel 7 node also seems to have got some technological improvements, with the company referring to it as the "3rd generation" of this node (optical 10 nm). This mainly concerns better electrical characteristics from improved channel mobility. Cushioning the P-core with a larger 2 MB dedicated L2 cache also appears to be contributing to the iso-power uplift, as the core spends fewer cycles fetching data from the L3 cache. We will learn more about "Raptor Cove" in the coming days, and will hopefully have a more detailed look at the new core in our reviews of these processors.
The E-core microarchitecture is the very same "Gracemont," but benefits from the node improvements to dial up E-core boost frequencies all the way up to 4.30 GHz. The cores also benefit from the larger 4 MB L2 cache that's shared among four E-cores in a "Gracemont" cluster. "Raptor Lake" has four such clusters, amounting to 16 on the silicon. The E-core clusters have access to the chip's L3 cache, just like the P-cores. As we mentioned earlier, the improved cache, and updated prefetcher algorithm should have an cumulative impact on E-core performance; and when you account for 16 of these, besides the improved 8 P-cores, you begin to see where Intel's 41% generational multi-threaded performance uplift claim is coming from.
Intel also made updates to the processor's uncore components. The L3 cache that's shared among the processor's P-cores and E-core clusters, is now enlarged to 36 MB, from 30 MB in the previous generation. This cache is a continuously addressable block due to the Ringbus interconnect making ring-stops at various physical segments of the cache. Intel has improved the clock-speed of this fabric, which now boosts up to 5.00 GHz, or 900 MHz higher than the previous-gen.
The DDR5+DDR4 memory controllers also receive an update. The processor now natively supports up to DDR5-5600 JEDEC-standard memory speeds, when using 1 DIMM per 80-bit channel (which has two 40-bit sub-channels); or up to DDR5-4400 when using 2 DIMMs per channel (i.e. populating all four memory slots on your motherboard).
Intel also updated the Thread Director middleware that gives the software some degree of awareness of the Hybrid architecture, and attempts to ensure that the right kind of workload is allocated to the right kind of CPU core. Intel has given TD greater thread class awareness through machine-learning techniques (the processor learns over time what the nature of the workload could be). The processor also takes advantage of new scheduling features of Windows 11 22H2 Update, which introduce PID QoS for system background tasks and user-initiated background tasks.
Intel claims that "Raptor Lake" processors will be memory overclocking monsters, capable of speeds as high as DDR5-10000, when pushed to the limit with enthusiast-grade memory. For the P-cores, the company says that 8.00 GHz overclocks are now within reach for enthusiasts. The updated Intel Extreme Tuner Utility (XTU), allows you to set multipliers on a per-core basis, and tune your memory frequency on-the-fly (no reboots involved).
Intel is launching the 13th Gen Core "Raptor Lake" desktop processor family with essentially three processor models—Core i9-13900K, Core i7-13700K, and Core i5-13600K; and their "KF" sub-variants that have disabled iGPUs, and are about $10-20 cheaper, depending on the model.
The Core i9-13900K is the flagship part, with 8 P-cores, and 16 E-cores (8P+16E), with the full 36 MB L3 cache available on the silicon. The P-cores have a base-frequency of 3.00 GHz, and boost up to 5.80 GHz; whereas the E-cores run at 2.20 GHz base, boosting up to 4.30 GHz. The processor base power is rated at 125 W, and the maximum turbo power at 253 W (up from 241 W for the i9-12900K). The i9-13900K comes with an MSRP of USD $589, while the i9-13900KF (which lacks the iGPU), is priced at $564.
The Core i7-13700K is an interesting SKU, as it has the same 8P+8E core-configuration as the previous-gen i9-12900K, but with all the new updates detailed above. Intel carved this SKU out by disabling two of the four E-core clusters on the "Raptor Lake" silicon, and reducing the L3 cache to 30 MB. The P-cores have a base frequency of 3.40 GHz, with a maximum boost frequency of 5.40 GHz; while the E-cores run at 2.50 GHz base, and 4.20 GHz maximum boost. These chips have the same 125 W PBP and 253 W MTP as the i9-13900K. The i7-13700K is priced at $409, and the i7-13700KF at $389.
The Core i5-13600K is an equally interesting processor with which the company hopes to hold on to the mid-range. It now comes with a 6P+8E core-configuration, compared to 6P+4E of the i5-12600K. And of course, you get all the generational improvements detailed above. This SKU is carved out by disabling two P-cores, and two E-core clusters; while also cutting down the L3 cache to 24 MB (which is still higher than the 20 MB of the i5-12600K). The P-cores run at 3.50 GHz base with 5.10 GHz boost; while the E-cores do 2.60 GHz base, with 3.90 GHz boost. While the PBP value is the same 125 W as the higher SKUs, the MTP is reduced to 181 W. Intel is pricing the Core i5-13600K at $319, and the i5-13600KF at $294.
Intel is claiming gaming performance uplifts of up to 18% when comparing the i9-13900K with the previous-gen i9-12900K, across a wide selection of games; while the comparison with the AMD Ryzen 9 5950X "Zen 3" sees gaming performance gains range between 6% to 58%. The gap only widens when you consider the 99th percentile low-water-mark analysis. Although mainly compared with the 5950X, Intel also threw in gaming performance values it tested on the Ryzen 7 5800X3D, which is shown matching the i9-13900K in games where it's beating the i9-12900K, or within 10% of it in games where the i9-13900K gets ahead. This is interesting, as Intel thinks the performance of "Zen 4" Ryzen 7000-series processors should roughly match that of the 5800X3D. In our performance reviews published on September 26, the 5800X3D is 4.5-5% behind the Ryzen 7 7700X, which means "Zen 4" should end up within 5% of the i9-13900K in gaming performance, should these numbers for the 5800X3D from Intel hold up.
The platform I/O of these processors is identical to that of "Alder Lake." You get a 2-channel (4 sub-channel) DDR5 + 2-channel DDR4 memory interface. The processor puts out 28 PCI-Express lanes; 16 of these are Gen 5, and intended for the main x16 PEG slot; while the remaining are Gen 4. The main x4 NVMe interface of the processor is Gen 4, while the DMI chipset bus takes up the remaining 8 lanes (DMI 4.0 x8). You should still find motherboards with Gen 5 M.2 NVMe slots, but these would be eating into the x16 PEG bandwidth. Given that NVIDIA's latest GeForce Ada continues to be PCIe Gen 4, cutting into the bandwidth of the PEG slot to run Gen 5 M.2 SSDs could affect graphics performance (but we'll test this theory in upcoming PCIe-scaling articles with the RTX 4090).
The complete slide-deck for the processor launch event follows.
Intel's performance claims are backed by some impressive hardware changes despite the company sticking with the same Intel 7 (10 nm Enhanced SuperFin) foundry node as "Alder Lake." To begin with, the single-thread performance uplift comes from the new "Raptor Cove" performance-core, which promises an IPC uplift over the previous-generation "Golden Cove," comes with more dedicated L2 cache of 2 MB per core (compared to 1.25 MB per core in the previous-generation); and significantly higher clock-speeds, going all the way up to 5.80 GHz. "Raptor Lake" has up to 8 P-cores, but the company has put in a lot of work in improving the contribution of E-cores to the processor's overall multi-threaded performance uplift. This is achieved by doubling the E-core count to 16. These are the same "Gracemont" E-cores as previous-generation, but Intel has doubled the L2 cache that's shared in a 4-core Gracemont cluster, from 2 MB per cluster to 4 MB. There are upgrades to even the hardware prefetchers of these cores.
The DDR5+DDR4 memory controllers also receive an update. The processor now natively supports up to DDR5-5600 JEDEC-standard memory speeds, when using 1 DIMM per 80-bit channel (which has two 40-bit sub-channels); or up to DDR5-4400 when using 2 DIMMs per channel (i.e. populating all four memory slots on your motherboard).
Intel also updated the Thread Director middleware that gives the software some degree of awareness of the Hybrid architecture, and attempts to ensure that the right kind of workload is allocated to the right kind of CPU core. Intel has given TD greater thread class awareness through machine-learning techniques (the processor learns over time what the nature of the workload could be). The processor also takes advantage of new scheduling features of Windows 11 22H2 Update, which introduce PID QoS for system background tasks and user-initiated background tasks.
The Core i9-13900K is the flagship part, with 8 P-cores, and 16 E-cores (8P+16E), with the full 36 MB L3 cache available on the silicon. The P-cores have a base-frequency of 3.00 GHz, and boost up to 5.80 GHz; whereas the E-cores run at 2.20 GHz base, boosting up to 4.30 GHz. The processor base power is rated at 125 W, and the maximum turbo power at 253 W (up from 241 W for the i9-12900K). The i9-13900K comes with an MSRP of USD $589, while the i9-13900KF (which lacks the iGPU), is priced at $564.
The Core i7-13700K is an interesting SKU, as it has the same 8P+8E core-configuration as the previous-gen i9-12900K, but with all the new updates detailed above. Intel carved this SKU out by disabling two of the four E-core clusters on the "Raptor Lake" silicon, and reducing the L3 cache to 30 MB. The P-cores have a base frequency of 3.40 GHz, with a maximum boost frequency of 5.40 GHz; while the E-cores run at 2.50 GHz base, and 4.20 GHz maximum boost. These chips have the same 125 W PBP and 253 W MTP as the i9-13900K. The i7-13700K is priced at $409, and the i7-13700KF at $389.
The Core i5-13600K is an equally interesting processor with which the company hopes to hold on to the mid-range. It now comes with a 6P+8E core-configuration, compared to 6P+4E of the i5-12600K. And of course, you get all the generational improvements detailed above. This SKU is carved out by disabling two P-cores, and two E-core clusters; while also cutting down the L3 cache to 24 MB (which is still higher than the 20 MB of the i5-12600K). The P-cores run at 3.50 GHz base with 5.10 GHz boost; while the E-cores do 2.60 GHz base, with 3.90 GHz boost. While the PBP value is the same 125 W as the higher SKUs, the MTP is reduced to 181 W. Intel is pricing the Core i5-13600K at $319, and the i5-13600KF at $294.
The complete slide-deck for the processor launch event follows.
224 Comments on Intel 13th Gen Core "Raptor Lake" Desktop Processors Launched: +15% ST, +41% MT Uplift
Further integration the way you suggested it - yeah, I have similar ideas too, don't know if it's on the way to become reality though. A piece of code could be tagged with some metadata, like hints to help the scheduler decide what kind of core is the best fit for it.
[SIZE=3]ASRock X670E PG Lightning in the UK: £330 :([/SIZE]
^absolutely...i think you overlooked the following from my previous post:
"paired with an affordable B-series board"
"I'm sure many (like myself) are waiting to see how B-series boards hold up, possibly small but relevant further trim on DDR5 prices and no doubt some Intel-AMD price war action at the close of 2022"
"Even if the reviews saw the 7600X outpacing the 12600K by a clear mile... i still wouldn't splurge up for a super expensive platform upgrade"
I'm not looking to buy into pompously flaunting X/XE-series boards.... i'm waiting to see how B-series pans out (presumably a better fit for gamers)
Also I don't mind throwing a little extra cash at some high performance DDR5 memory. This is inescapable if going the DDR5 route, whether 13th Gen/Zen 4. For me its simple, wait towards the end of 2022 and re-configure all build possibilities and see where we are with best value configurations. If the AM5 socket isn't far ahead, i wouldn't mind forking out ~$50 (or more) to buy into the AM5 3 year+ support plan.
And you know what the people choose rather? Performance or cost? Cost every single time (most of the time ofc). :laugh:I totally agree with everything you said. And I think that's the problem, Zen 4 processors are not (necessarily) expensive, the whole platform around it is. And that's the real problem.
We need to have cheapest motherboard at 100 or 150$, not 260 or 320 euros in EU (this is the same MB as you listed here). We need RAM at 50 - 70% of its current cost. If you had this, your total cost would drop 50 - 80 for RAM and 110 - 160 for MB for a total of 200 on average. Which would be 4 times the savings compared to the CPU being dropped 50$, so 28% on the 7600X build, and 25% on the 7700X build.
This is BIG.
And this is the Achilles heel of the Zen 4 ecosystem. Platform cost.
Point being: what you're describing doesn't seem realistic. Yes, the worse the cooling, the lower the clocks. Keeping your system reasonably clean is a must - but this is nothing new. This happened on Ryzen 5000 if your thermals exceeded 75°C, so the threshold was lower there, after all. As for "old thermal paste" - I've literally never seen a system I've built suffer thermally from old paste. Buy decent stuff, it'll last for years and years.
Will you see a few hundred MHz lower clocks if you don't clean your system more than, say, once a year? Sure. But that's true for essentially every system, unless your cooling is massively overkill to begin with.
Sad.The thing is, I wouldn't take the "risk" with a new platform (and 12 years from my last upgrade) that might not allowe me to fine tune the heat to noise (fan speed). A very basic things I think.
I imagine most people on air coolers with Ryzen 7000 will set their fan curves to not ramp up until a much higher temp than in the past and simply set a max RPM they find tolerable for 90+ C. Obviously not every air cooler user will find that acceptable.
Any rumers if and when 7900\7950 non X will arrive?
You might request CPU package power monitoring to be added if not already supported in SpeedFan and "Fan Control" software. There could be a motherboard out there that can control fan speed off package power via bios or their software suite. I could see this feature being something more and more people who use air cooling will want with CPUs that hit thermal limits before hitting power limits.
Until the options exist I think a lot of people on air cooling will simply choose a noise level they're ok with and set their final fan curve to hit that speed at 90c or so. It's not like air coolers don't already ramp up at the slightest duration of high demand and ryzen 7000 isn't going to be 95c at idle or watching videos.
More advanced users may undervolt and set power limits like the videos linked in this post.
Of the 15% single-threaded performance, most of that is just a frequency bump from insane 250W PL2 madness on the refreshed process node. 5% actual IPC is nice to have as a bonus but it means that RPL probably isn't going to be significantly better than ADL outside of the flagship i9s and perhaps the i7-13700K.
More pedestrian non-K models like the i5-13400 aren't going to be running at almost 6GHz. Intel will of course artificially cap the turbo frequencies on those to avoid cannibalising ADL i7 inventory as they've done timeless counts in the past.
Zen 4 improvements mostly come from the same thing - clocks and cache. And that is a completely new platform. Alder Lake was so much more impressive (and appealing) a year ago than Zen 4 is now. But this might change with 3D cache and cheaper boards.
Compared to Kaby Lake, Comet Lake and Rocket Lake, this is so much better in every way.
12600K $250 vs <>
12700K $350 vs 13600K $329
12900K $499 vs 13700K $459
I'm not sure I'm seeing the value on 13th gen vs current discounted 12th gen parts.
I don't see performance uplift being significant for 13600K outside of multi-threaded apps, heavy MT I don't care about as long as it is as fast as my current 10850K - which a 12600K is about equal.
www.intel.com/content/www/us/en/products/sku/230500/intel-core-i713700k-processor-30m-cache-up-to-5-40-ghz/specifications.html
Maximum Turbo Power: 253W
I'm not even sure what those numbers are supposed to mean, coming from a 10850K. I have my 28 second average set to 180W and my short power max (7s usually) at 235W while my 240 AIO can handle sustained 220W(ish). In real life though, I never hit those numbers.
I've run HWInfo64 all day at different points since I've owned this 10850K and I know, without doing something to intentionally stress my CPU like a benchmark, typical for me is ~30-35W average, ~150W peak package power. Heck I rarely even hit 150W peak, plenty of times I've seen it more like 100W peak, and the peaks are almost always caused by decompress and patch from steam or Windows update.
Maybe someone who is doing a lot of ray-tracing \ rendering should pay attention, but that's not me.
First of all, major difference is the idle wattage. Holy cow the 12900k drops down to 2 watts. 10900k for me (oced) was always sitting at like 20w+.
The truth is, at the same wattage, the 12900k is a little bit harder to cool, but not by much. It's still a huge die, and the bigger the die - the easier it is cooled.
If you are doing a lot of rendering / ray - tracing, none of that matters, cause no one is going to use any CPU at 250w to do that. You gain 3 to 5% performance going from 150w to 250w, and the same applies to AMD's zen 4. You power limit them and go on with your workloads. It's absurd - and I hope it stops now - listening to the amd crowd arguing about cinebench numbers and wattage, as if anyone in theier right mind will be rendering at 5ghz all core clockspeeds at 250w. It's just borderline silly.
Of course, problems persist for the users of Windows 10, because hybrid CPUs need Windows 11 due to kernel changes (core scheduler/thread manager) fit for them.
I started HW Info yesterday when I posted that and my peak power since has been 105W and my average is 33W. And yes I played a game, Mordor Shadows of War, last night. Otherwise it's email, RDP, web, MS Teams, and a few other normal \ light loads.
Factually speaking, I'm likely using far less energy than I would with a Zen 3 Ryzen simply because their platform idle power draw is typically much higher. And I'm running Windows max performance setting with a 5Ghz all core and power limits raised from stock.
That's real life with an very mild OC 10850K. Current power, min power, max power, average power for the last 20 hours or so.
Clock speeds :
Because you're doing ok on light load's, wow the informed Ness of it all (with 20 hours of doing nothing, turn it off FFS)
Be interesting to see your and fevgatos take on raptor lake post reviews since your clearly one of those who prefer Intel's wares.