Intel's Arrow Lake-S & Arrow Lake-HX CPUs will feature a slightly different cache structure than Lunar Lake CPUs & much higher AI TOPs than Raptor Lake.
Intel Arrow Lake-S & Arrow Lake-HX High-End CPUs Bring 37 AI TOPs, Up To 76 MB Combined Cache
More information regarding Intel's Arrow Lake-S & Arrow Lake-HX CPUs has been shared by leaker, @Jaykihn, who now discusses the cache configuration for the high-end chips along with their latest improvements over Raptor Lake-S and Raptor Lake-HX CPUs.
According to the latest data, it looks like Intel's Arrow Lake-S and Arrow Lake-HX mostly feature the same cache structure as Lunar Lake CPUs since they are based on the same Lion Cove P-Cores and Skymont E-Cores but the difference is that Lunar Lake integrates the slightly power-tuned version of Lion Cove P-Cores while the E-cores are low-power in design. The L0d, L1d, and L1i cache are going to remain the same but the differences come in the L2 and L3 cache.
Intel's Arrow Lake-S and Arrow Lake-HX CPUs will feature 3 MB (12-Way) L2 cache per P-Core and 4 MB (16-Way) L2 cache per E-Core cluster (4 cores). On Lunar Lake, the L2 cache is 2.5 MB (10-way) per Lion Cove P-Core while the configuration remains unchanged for the E-Cores. For L3, the Intel Arrow Lake-S & Arrow Lake-HX CPUs will feature a 3 MB L3 cache per Lion Cove P-Core and 3 MB of L3 cache per Skymont E-Core cluster. Lunar Lake features the same 3 MB cache for the P-Cores but doesn't feature any dedicated L3 cache for the Skymont LP-E cores.
- Arrow Lake P-Core Cache (Lion Cove): 3 MB L2 / 3 MB L3
- Arrow Lake E-Core Cache (Skymont): 4 MB L2 / 3 MB L3
- Lunar Lake P-Core Cache (Lion Cove): 2.5 MB L2 / 3 MB L3
- Lunar Lake E-Core Cache (Skymont): 4 MB L2 / 0 MB L3
- Raptor Lake P-Core Cache (Raptor Cove): 2 MB L2 / 3 MB L3
- Raptor Lake E-Core Cache (Gracemont): 4 MB L2 / 3 MB L3
Intel Arrow Lake-S/HX vs Lunar Lake Cache Configuration:
| Cache | Lunar Lake (P-Core) | Lunar Lake (E-Core) | Arrow Lake (P-Core) | Arrow Lake (E-Core) |
|---|---|---|---|---|
| L0 Data | 48 KB 12-Way | None | 48 KB 12-Way | None |
| L1 Data | 192 KB 12-Way | 32 KB 8-Way | 192 KB 12-Way | 32 KB 8-Way |
| L1 Instruction | 64 KB 16-Way | 64 KB 16-Way | 64 KB 16-Way | 64 KB 16-Way |
| L2 | 2.5 MB 10-Way | 4 MB 16-Way (Cluster) | 3 MB 12-Way | 4 MB 16-Way (Cluster) |
| L3 | 3 MB | None | 3 MB | 3 MB (Cluster) |
We know that Intel's Arrow Lake CPUs will include two main dies, an 8+16 and a 6+8 SKU. The CPUs will come in various configurations based on these two dies. If we were to do the math, the top die would feature 24 MB of L3 cache for the 8 P-Cores and 12 MB of L3 cache for the four E-Core clusters. That gives us 36 MB of L3 cache which matches the current Raptor Lake-S CPUs. For the L2 cache, the same die will feature 24 MB of L2 cache for the 8 P-Cores and 16 MB of L2 cache for the E-Cores which gives us 40 MB of total L2 cache. This gives us a combined 76 MB cache pool, or 12% more cache than Raptor Lake-S/HX CPUs which maxed out at 68 MB of combined cache.
Intel Arrow Lake-S vs Raptor Lake-S Die SKU Cache Configurations (Credits: @Jaykihn):
| CPU | Arrow Lake-S (8+16) | Arrow Lake-S (6+8) | Raptor Lake-S (8+16) | Raptor Lake-S (6+8) |
|---|---|---|---|---|
| L2 Cache (P-Core) | 24 MB | 18 MB | 16 MB | 12 MB |
| L2 Cache (E-Core) | 16 MB | 8 MB | 16 MB | 8 MB |
| L3 Cache (P-Core) | 24 MB | 18 MB | 24 MB | 18 MB |
| L3 Cache (E-Core) | 12 MB | 6 MB | 12 MB | 6 MB |
| L2 Cache (Total) | 40 MB | 26 MB | 32 MB | 20 MB |
| L3 Cache (Total) | 36 MB | 24 MB | 36 MB | 24 MB |
| Combined Cache (Total) | 76 MB | 48 MB | 66 MB | 44 MB |
We can do the same for the 6+8 configuration for the Intel Arrow Lake-S and Arrow Lake-HX SKUs which should give us an 18 MB L3 cache for the P-Cores and 6 MB L3 cache for the E-Cores while the L2 cache is 18 MB for the P-Cores and 8 MB for the E-Cores. That's a combined L3 cache of 24 MB, a combined L2 cache of 26 MB, and a total pool of 50 MB. That's around 13.5% more combined cache versus the 44 MB featured on the Raptor Lake-S 6+8 configuration.
In a previous report, it was suggested that Intel's Arrow Lake-S & Arrow Lake-HX CPUs will employ an NPU with an upgraded version going into a possible Arrow Lake Refresh lineup. The leaker also shares some information in this regard with some small but noticeable improvements being made on the new lineup versus Raptor Lake-S CPUs.
ARL-S full lineup will require more time to compile, apologies.
For now, I can confirm that the 285K, 265K, and 245K monikers are official.
Moreover, all Core Ultra 9 SKUs will ship with integrated graphics. So, no Core Ultra 9 F SKUs.
— Jaykihn (@jaykihn0) July 7, 2024
In terms of AI capabilities, the GPU TOPS will see a 3x gain from just 3 TOPS on Raptor Lake to 9 TOPS on Arrow Lake with the newer Xe architecture. The NPU incorporated on these chips will be rated at 13 TOPS while the CPU TOPS see an 87.5% uplift with 15 TOPS. This will give a combined 37 AI TOPS which puts it close to the Ryzen 8000G "Hawk Point" APUs which have a combined TOPS of 39.
Other improvements include support for PCIe Hot-Plug, more ESPI device support, the addition of I3C interfaces, increased DMA Audio stream inputs, HD audio bi-directional streams, channels per PDM streams, PCIe clock output, and SRC clock requests.
Intel Arrow Lake-S vs Raptor Lake-S Changes (Credits: @Jaykihn):
| Changes | Raptor Lake-S | Arrow Lake-S |
|---|---|---|
| CPU TOPS | 8 | 15 |
| NPU TOPS | 0 | 13 |
| GPU TOPS | 3 | 9 |
| Total TOPS | 11 | 37 |
| PCIe Hot-Plug | No | Yes |
| eSPI Device # Support | 2 | 4 |
| I2C/I3C Interfaces | 8/0 | 6/2 |
| DMA Audio Stream Input Support | 7 | 10 |
| HD Audio Bidirectional Streams | 6 | 8 |
| Channels Per PDM Stream | 2 | 8 |
| PCIe Clock Output | 7 | 14 |
| Imaging Clock | 4 | 0 |
| SRC Clock Request | 10 | 14 |
Overall, the Intel Arrow Lake-S and Arrow Lake-HX CPUs should be a good addition to a fresh new platform in the form of LGA 1851 (800-series motherboards) and high-end laptops. It was already reported by us that the Arrow Lake-S lineup (K-Series) along with the first Z890 boards will be out in October and we recently got to see some performance figures based on ES2 samples.
News Sources: @Jaykihn #1, #2
