Microprocessor Report (MPR) Subscribe

Intel’s Jasper Lake Lifts GPU by 5x

New Atom-Powered Processor Expands Budget-PC Capabilities

March 8, 2021

By Aakash Jani


Intel’s new Jasper Lake brings a 5x GPU-performance increase to the traditionally underserved low-end segment. Relative to its three-year-old predecessor, Gemini Lake, the new chip also raises single-core CPU performance by 32% and multicore performance by 22%. It better suits the high-end Chromebook and budget-PC markets by diversifying its compute blocks, reducing power consumption and system cost.

Jasper Lake comes in six models—two Pentium Silver and four Celeron—as Table 1 shows. Although Intel withheld pricing, we estimate the N6005 and N5105 each list for about $100. These 14nm chips employ either two or four single-thread Tremont cores and Gen11 GPUs. They also integrate the sixth-generation Intel image processor. They come in two TDPs: 10W for low-cost PCs and Chromebooks, and 6W for 2-in-1s and other power-constrained devices.

Table 1. Jasper Lake processors. The Pentium models’ higher execution-unit (EU) count and GPU frequencies yield a 5x gain over previous Gemini Lake chips. (Source: Intel)

In these low-end systems, Jasper Lake competes with AMD’s Athlon series. From the Arm universe, Qualcomm delivers the 5G-powered Snapdragon 8cx for Chromebooks and PCs alike. All three companies hope to gain share in the fast-growing Chromebook market. In 2020, Chrome OS overtook MacOS as the second-most-popular PC operating system.

Tremont Raises CPU Performance by 32%

Gemini Lake employed Intel’s Atom-based Goldmont+ microarchitecture (see MPR 1/1/18, “Gemini Lake Targets Tiny PCs”). The company’s newest price-conscious chips feature the next-generation Tremont CPU. Intel reports the new microarchitecture delivers 30% more per-clock performance than its predecessor by doubling the reorder-buffer size and increasing the number of execution units (EUs) from 8 to 10 (see MPR 11/4/19, “Intel’s Tremont: A Bigger Little Core”). Instead of a single L2 cache, Tremont has an expanded memory subsystem with 1.5MB of shared L2 and 4MB of L3—40% more memory than the previous chip.

Although both the N5105 and its predecessor, Gemini Lake’s J4125, have the same base clock speed of 2.0GHz, the former has a 7% faster boost clock; we estimate it leads in single- and multicore performance by 37% at its peak frequency. Cinebench R20 reveals a 32% single-core lead and 21% multicore lead. These benchmark scores are moderately lower than our estimates, suggesting Intel’s boost mode runs for only a fraction of the test.

Coarse Pixel Shading Raises GPU Throughput

Figure 1 shows Jasper Lake features Intel’s 11th Generation (Gen11) integrated graphics and implements 24 of 32 possible EUs (see MPR 1/21/19, “Intel Turns Graphics up to (Gen) 11”) arranged in four subslices. It delivers a stark upgrade from the Gen9.5-powered Gemini Lake, boasting 2x more EUs and a superior GPU microarchitecture. In the new design, each EU offers 3x more FP32 operations per cycle. In addition to doubling the base GPU frequency to 450MHz and raising the boost frequency by 7%, the chip increases peak graphics performance by 4.8x.

Figure 1. Jasper Lake block diagram. IPU=image processor; GNA=gaussian neural accelerator. The IPU sharpens images when videoconferencing with a single HD camera.

The GPU implements 256KB of shared local memory, 25% more than the Gen9.5 GPU, and 1.3MB of L3 cache, double that of the previous chip. By enlarging the on-chip GPU memory, Jasper Lake saves power by minimizing external-memory reads through its 4x32-bit LPDDR4/DDR4 controller. To increase parallelism, Intel split the L3 cache into four memory banks. The design additionally sports a dedicated voltage regulator (FIVR) for smoother power ramps during intense rendering.

Coarse pixel shading (also called variable-rate shading) debuted in the company’s Gen11 GPUs as a discrete hardware accelerator. It decreases the strain on the shader units by varying the shading rates on the basis of contextual frame information. By proxy, the accelerator increases GPU throughput and reduces power consumption.

 The company also integrated a logic block that performs position-only tile-based rendering. It operates ahead of the shader cores in the rendering pipeline and bins visible tiles. For example, suppose an image of a flower is obscured by an alphanumeric character. The tile-binning accelerator would omit information about the flower and only pass information about the character for rendering. The GPU’s L3 cache then stores the preprocessed tiles for the rendering pipeline to handle. This technology efficiently manages external-memory access by reducing the number of render passes per tile.

Computational Photography for Laptops

The new GPU employs hardware acceleration for low-power media encode/decode and transcoding. It natively supports 8- and 10-bit HEVC for high-definition encoding and decoding. For standard definition, the GPU decodes MPEG-2 at 60 frames per second (fps).

Transcoding requires both decode and encode operations; it serves in photo- and video-editing software. The accelerator decreases power consumption for these processes by employing lossless memory compression to reduce memory operations and a dedicated hardware unit to convert formats.

The chip’s device interface has three pipelines to support two external displays and one device display. Single displays render 4K at a 60Hz refresh rate, and we expect the resolution for each screen in a multidisplay setup to remain at 1080p. The N5105 provides either three DisplayPort (DP) or three HDMI ports through an Embedded DP and MIPI connection.

The SoC implements a less powerful version of Intel’s sixth-generation image processor (IPU), which Tiger Lake debuted (see MPR 9/7/20, Tiger Lake Debuts 10nm SuperFin”). The new design uses MIPI CSI to connect four cameras or image sensors over two data lanes each, but it can handle only three cameras or sensors concurrently.

Jasper Lake devices can capture HD video at 30fps from a single camera or 720p at 30fps from two cameras. The IPU stitches together feeds from both cameras to form high-dynamic-range (HDR) content, a capability that smartphones typically offer.

As more companies employ virtual meetings, chip vendors are beefing up their videoconferencing technologies. Intel worked with Microsoft to create an API that allows bare-metal access to applications such as Microsoft Skype and Teams. Because they bypass the CPU, these applications consume far less power on Intel-based computers relative to other devices.

Scary Lessons From Spectre

Home cooks who frequently wake a laptop to view a recipe may coat the keyboard with floury fingerprints. Intel aids such users with its Wake on Voice technology, which unlocks devices that process wake words in the chip’s audio engine. In Jasper Lake, this engine comprises one Tensilica LX6 controller and one HiFi 3 DSP. The DSP has 64KB of L1 cache and 1MB of L2 cache; the latter stores a wake-word key.

OEMs can opt to use the second DSP, which Intel calls a gaussian neural accelerator (GNA), for noise and echo cancellation. It delivers a modest one billion INT8 operations per milliwatt with a total throughput of 0.038 TOPS. Jasper Lake uses this low-power accelerator for wake-word detection and other simple inference tasks, but it’s too small for more-complex ones such as voice and image recognition, which run on the GPU.

Intel and other processor vendors faced a mountain of scru­tiny after white-hat hackers revealed the Spectre and Melt­down vulnerabilities (see MPR 1/22/18, “Melt­down: Who’s at Fault?”). Intel responded by quickly de­ploy­ing software patches and vowing to bolster security in future products. Jasper Lake is the latest processor to offer its up­graded security suite. The design employs branch-monitor counters and a shadow return-address stack to prevent return-oriented programming attacks. The countermeas­ures also monitor the number of memory reads during an execution window, and the built-in heuristics engine noti­fies the CPU if the count is too high.

Like its predecessor, Jasper Lake’s connectivity suite resides on the south bridge and requires external RF support. The design supports 2x2 Wi-Fi 5 and 2x2 Wi-Fi 6, both featuring 160MHz of bandwidth. OEMs can optionally add an external cellular modem through the multifunction UART lane while maintaining Wi-Fi and LTE coexistence. For headphones, mice, and other wireless peripherals, Jasper Lake implements both the 5.0 and LE versions of Bluetooth.

To handle off-chip communication, the chip has two 12-lane PCIe Gen3 controllers with a total bandwidth of 48GB/s. Its two DDR4/LPDDR4X controllers provide 29.8GB/s of total memory bandwidth.

A Decisive Victory Over AMD

Table 2 matches the Celeron N5105 against the AMD Athlon Gold 3150U and Qualcomm Snapdragon 8cx. The 3150U packs two dual-thread Zen CPUs and a Vega 3 GPU. The 8cx is a laptop-focused derivative of the Snapdragon 855 that targets PCs (see MPR 1/7/19, “Qualcomm Gets Serious About PCs”). We expect both Qualcomm and AMD to release newer and faster Chromebook processors along with low-power Windows processors later this year.

Table 2. Chromebook and budget PC processors. The new Jasper Lake chip bests its AMD counterpart in CPU and GPU performance but falls to Snapdragon in the same categories. *On Cinebench R20 CPU test; †theoretical peak throughput. (Source: vendors)

The N5105 leads the 3150U by 38% in single-core performance and 57% in multicore performance on Cinebench R20 benchmarks. For heavy floating-point workloads such as web browsing and rendering, multicore operations vie for vector-unit access. Even through both chips have the same number of threads, the Intel part tops the AMD part in multicore performance.

We estimate the Cortex-A76 and Tremont architectures yield similar instructions per cycle. By employing a similar frequency, Snapdragon should match the N5105 in single-core throughput. For multicore, we estimate the N5105 lags by about 10%. However, the Intel chip uses 42% more power than the 8cx, which drastically reduces system battery life.

The N5105 falls in the middle for GPU performance. Its theoretical FP32 throughput leads the 3150U’s by more than 2x and lags the 8cx’s by about 50%. Even though none of these chips can handle AAA games, users should be able to enjoy popular low-resolution games such as Among Us and Valheim.

Owing to the wide adoption of noise cancellation, background blurring, and other AI features, chips must be capable of efficient inferencing. Intel’s GNA mainly performs wake-word recognition, whereas Hexagon can inference full speech-recognition models such as MobileBert. For serious AI performance, Jasper Lake inferences on its GPU, but that method is far less power efficient.

Intel’s product stack features a leading Wi-Fi module that implements version 6 of that protocol, while the older 8cx only implements version 5. The company takes Wi-Fi a step further by adding 60GHz WiGig+, which avoids bandwidth competition from other devices but only in peripherals that support this uncommon technology.

The N5105 lacks an integrated LTE modem, unlike the 8cx. But customers remain skeptical of cellular laptops owing to monthly fees, especially for low-end models. But because many students are learning from home, Internet connections may be unstable. Modem-based laptops with cellular plans paid for by the school can address this problem.

Goodbye to Gemini Lake

OEMs and users waited a long time for a Gemini Lake follow-on, but they will enjoy the gains that the new family brings. Jasper Lake provides a 31% CPU-performance boost and dramatically raises GPU performance by almost 5x. While Jasper Lake Chromebooks still struggle with AAA games, the improved GPU delivers stunning visuals for native Chrome OS applications. Most importantly, Intel expanded heterogeneous computing in these chips. The IPU offloads more operations from the CPUs and GPU, reducing power consumption. Yet even with these improvements, the TDP remains unchanged from the last generation.

Jasper Lake serves in both PCs and Chromebooks. In the former category, Intel crushes its x86 competitors. Although the Arm-powered Snapdragon 8cx leads in CPU and GPU performance, its software ecosystem is still maturing. Until the Windows-on-Arm ecosystem improves, Intel will control the low-price-PC market. In Chromebooks, however, native apps are less important, and the Snapdragon chip could gain share due to its superior battery life. As next-generation products from AMD and Qualcomm come to market, this segment will become heavily contested.

Jasper Lake delivers a phenomenal upgrade over Gemini Lake, increasing performance without a power blowout. Intel will benefit from the recent surge of Chromebook users, and we expect the new chip will help expand this already explosive market.

Price and Availability

Intel’s Jasper Lake is in production and should appear in PCs in 2Q21. Although the company withheld pricing, we expect the chips list for $70 to $100 in 1,000-unit volume, depending on the model. Additional information is at ark.intel.com/content/www/us/en/ark/products/codename/128823/jasper-lake.html.

Free Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products

Events

Linley Spring Processor Conference 2022
Conference Dates: April 20-21, 2022
Hyatt Regency Hotel, Santa Clara, CA
Linley Fall Processor Conference 2021
Held October 20-21, 2021
Proceedings available
Linley Spring Processor Conference 2021
April 19 - 23, 2021
Proceedings Available
More Events »