Intel yesterday introduced new desktop processors of the Arrow Lake series, and compared to the boring Core 14th generation, which were only clocked higher than the predecessor, this time no stone was left unturned. Changes have occurred on all fronts.
Let’s start with the name. They have a new model designation in the format Core Ultra + series + model. The best of the chips is called Core Ultra 9 285Kunder the earlier nomenclature it would be Core i9 15900K. For notebooks, it already switched to new names last year with Meteor Lake (Core X 100) chips.
Intel also used a chiplet/tiled design instead of monolithic silicon for the first time in a desktop like competitor AMD. And for the first time ever, he does not make the processor cores himself in the desktop, but outsourced the production to TSMC.
The company uses a total of six tiles, but two do not carry logic circuits. The main one is the Compute tile with processor cores that came from the 3nm process (N3B), i.e. the most advanced one, which until now only Apple used. Intel made the GPU using the 5nm process (N5P). The SOC tiles (contains NPU, memory controllers, video processor, Wi-Fi…) and I/O (PCIe, Thunderbolt, USB) are based on TSMC’s 6nm process. Next to these four, there is one more empty tile (the so-called filler tile) so that the heat distributor has something to lie on, because there is space left in the corner. And all of these tiles use Foveros 3D encapsulation, so they sit on an active interposer that Intel is the only one to manufacture in-house (22nm technology).
Processors without HyperThreading
The fact that Intel continues to combine powerful and economical processor cores does not change. However, he changed the architecture for both. The powerful ones called Lion Cove promise a 9% higher IPC (power per cycle) compared to the previous Raptor Cove. Part goes for a higher L2 cache per core (3 instead of 2 MB), part for changed units. On the other hand, Intel removed HyperThreading support years later, so even powerful cores can only process one thread at a time.
The company compensates for this in two ways. Thanks to a better manufacturing process, the running frequencies of all cores have been noticeably increased (depending on the model, an extra 500 to 700 MHz). And the new economical Skymont cores will help in multi-threaded applications. Compared to the previous Gracemonts, they promise a 32% increase in IPC. Partly also because of the microarchitecture change, partly because they now have access to the shared L3 cache. Intel also deployed the third generation Thread Director, a hardware scheduler that takes care of the efficient allocation of tasks to individual cores.
The graphics accelerator then finally switched to the more modern Xe architecture, but in the desktop it will only serve as an office display. More interesting than the performance increase is that it supports the modern DisplayPort 2.1. The video processor has also been improved, which can now decode videos in 8K60 resolution with 10-bit colors and HDR in HEVC (H.265), AVC (H.264), AV1 and VP9 formats, and can then encode at twice the frame rate.
24
(8+16)
3,7–5,7 GHz
3,2–4,6 GHz
40+36 MB
Intel Graphics
13 TOPS
125 W (250 W)
i9-14900K
24
(8+16)/32
3,2–6,0 GHz
2,4–4,4 GHz
32+36 MB
UHD 770
–
125 W (253 W)
20
(8+12)
3,9–5,5 GHz
3,3–4,6 GHz
36+30 MB
Intel Graphics
13 TOPS
125 W (250 W)
i7-14700K
20
(8+12)/28
3,4–5,6 GHz
2,5–4,3 GHz
28+33 MB
UHD 770
–
125 W (253 W)
16
(8+8)
3,9–5,5 GHz
3,3–4,6 GHz
36+30 MB
–
13 TOPS
125 W (250 W)
i7-14700KF
20
(8+12)/28
3,4–5,6 GHz
2,5–4,3 GHz
28+33 MB
–
–
125 W (253 W)
16
(8+8)
4,2–5,2 GHz
3,6–4,6 GHz
26+24 MB
Intel Graphics
13 TOPS
125 W (159 W)
i5-14600K
14 (6+8)/20
3,5–5,3 GHz
2,6–4,0 GHz
20+24 MB
UHD 770
–
125 W (181 W)
14
(6+8)
4,2–5,2 GHz
3,6–4,6 GHz
26+24 MB
–
13 TOPS
125 W (159 W)
i5-14600KF
14
(6+8)/20
3,5–5,3 GHz
2,6–4,0 GHz
20+24 MB
–
–
125 W (181 W)
Arrow Lake then brings NPU to the desktop for the first time. But Intel used an older generation neural coprocessor with a performance of 13 TOPS, which is not enough for Microsoft Copilot+ certification (40 TOPS) and the related functionality of AI gadgets in Windows 11.
The processors are aimed at the new LGA 1851 platform, for which Intel will also introduce 800 series chipsets. The I/O options have been expanded, which is especially interesting considering that 20 PCIe 5.0 lines are now directly leading from the processor, not just 16 as with the Raptor Lacquer. The graphics card and the main SSD will not have to be divided. In addition, there will be four more direct PCIe 4.0 lanes for SSDs. The new motherboards are supposed to offer up to 10 USB 3.2 Gen 2 connectors (10 Gb/s), two Thunderbolt 4, Wi-Fi 6E, Bluetooth 5.3 and 1 Gb/s Ethernet. But board manufacturers can also connect a faster interface externally.
Arrow Lake supports dual-channel DDR5 memory, this time with a higher frequency of 6400 MT/s (instead of 5600 MT/s). However, the company itself says that even those clocked above the standard at 8000 MT/s will run without problems. The capacity is up to 192 GB.
Performance and consumption
The two previous generations of desktop processors made only minor changes at the microarchitecture level, while Intel failed to switch to a more advanced manufacturing process. As a result, he drove the performance up only with a higher clock rate, which resulted in an absurdly high consumption of chips, which ideally had to be cooled with water. Arrow Lake has a new architecture and at least two generations better production process. What follows from this?
Intel promises that the Ultra 200 processors will offer up to 6% higher single-threaded performance and 14% higher multi-threaded performance compared to the 14th generation Core when factoring in IPC and clocks. This is not a very interesting intergenerational increase, but the company claims that it has significantly reduced consumption. In normal use it should be up to 58% lower, while gaming it should be reduced by up to 165 watts. The new processors will go on sale on October 24, when independent tests will be out to confirm or deny this. The TDP values are not fundamentally different, but in operation the new chips can really be more economical.
The firm also published several comparisons with competing chips from AMD. In synthetic benchmarks such as SPECrate 2017, Cinebench 2024 or 3D Mark, it promises up to 4% higher single-threaded and 13% higher multi-threaded performance. Specifically, it put the Ultra 9 285K and the Ryzen 9950X side by side. Game performance should be plus or minus at the level of the Ryzen 7950X3D, in some ways better, in some cases worse. But why Intel does not compare directly with the best gaming processor Ryzen 7800X3D is a mystery. In addition, the 9800X3D is literally around the corner, which should increase the performance by another 10+ percent.