News: 1736324648

  ARM Give a man a fire and he's warm for a day, but set fire to him and he's warm for the rest of his life (Terry Pratchett, Jingo)

Additional Microprocessors Decoded: Quick guide to what AMD is flinging out next for AI PCs, gamers, business

(2025/01/08)


CES AMD unleashed a flood of CPUs at the annual CES super-conference this week in Las Vegas – chips that span the gamut from high-end desktop processors to those aimed at gaming notebooks, handhelds and, yes, so-called AI PCs.

While Qualcomm chases the lower-end of the AI PC market, AMD is targeting the upper-end with its newly unveiled [1]Ryzen AI Max 300 series, code-named Strix Halo, alongside additional Ryzen AI 300 series parts.

Just like the [2]original batch of Strix Point Ryzen AI 300 chips [3]launched last summer, AMD's Max chips feature the same 50 [4]TOPS XDNA 2 NPU as before, as well as the usual Zen 5 CPU cores. However, that's not to say the latest chips don't offer anything more in the AI department.

[5]

In addition to the NPU, AMD's top-specced Strix Halo components feature up to 40 RDNA 3.5 GPU cores delivering a combined total of 126 TOPS. Much of that extra grunt presumably comes courtesy of the GPU. And while NPUs are closely associated with AI PCs, not all AI workloads are optimized to take advantage of them.

[6]

[7]

While we're told it's possible to run LLMs on NPUs, the memory requirements associated with them means more often than not they end up running on the GPU.

With a 256-bit bus feeding the beefier GPU, the top-specced Strix Halo parts should deliver 256GB/s of bandwidth, which in theory should enable them to run large language models, such as Llama 3.2 3B locally at speeds of over 100 tokens per sec.

[8]

But while the new chips feature a more powerful GPU, they take a step back in memory capacity. Whereas Strix Point maxed out at 256GB of DDR5, Strix Halo only supports 128GB of LPDDR5x — no SODIMM support here — with 96GB directly accessible to the graphics pipeline.

AMD will also offer its Strix Halo processors with so-called Pro features, such as fleet management, enterprise security, and cloud recovery services, when they hit the market later this quarter. Here's the line up:

Model

Cores / Threads

Boost / Base Clock

Total Cache

GPU

TDP

NPU TOPS

GPU Cores

AMD Ryzen AI Max+ 395

16C/32T

5.1 / 3.0 GHz

80MB

Radeon 8060S

45-120W

50

40

AMD Ryzen AI Max 390

12C/24T

5.0 / 3.2 GHz

76MB

Radeon 8050S

45-120W

50

32

AMD Ryzen AI Max 385

8C/16T

5.0 / 3.6 GHz

40MB

Radeon 8050S

45-120W

50

32

AMD Ryzen AI Max+ Pro 395

16C/32T

5.1 / 3.0 GHz

80MB

Radeon 8060S

45-120W

50

40

AMD Ryzen AI Max Pro 390

12C/24T

5.0 / 3.2 GHz

76MB

Radeon 8050S

45-120W

50

32

AMD Ryzen AI Max Pro 385

8C/16T

5.0 / 3.6 GHz

40MB

Radeon 8050S

45-120W

50

32

AMD Ryzen AI Max 380

6C/12T

4.9 / 3.6 GHz

22MB

Radeon 8040S

45-120W

50

16

Alongside its Ryzen AI Max parts, AMD is also expanding its AI 300 family of products to include four lower-core count SKUs with and without AMD Pro features enabled:

Model

Cores / Threads

Boost / Base Clock

Total Cache

GPU

TDP

NPU Tops

AMD Ryzen AI 7 350

8C/16T

5.0 / 2.0 GHz

24MB

Radeon 860M

15-54W

50

AMD Ryzen AI 5 340

6C/12T

4.8 / 2.0 GHz

22MB

Radeon 840M

15-54W

50

AMD Ryzen AI 7 Pro 350

8C/16T

5.0 / 2.0 GHz

24MB

Radeon 860M

15-54W

50

AMD Ryzen AI 5 Pro 340

6C/12T

4.8 / 2.0 GHz

22MB

Radeon 840M

15-54W

50

X3D returns to mobile

Among the more unusual SKUs announced at the annual consumer electronics showcase is the unassumingly named 9955HX3D, which sees the return of AMD's 3D V-cache packaging tech to the notebook arena.

Looking at the spec sheet, the chip looks like just another Ryzen mobile processor, with 16 CPU cores, 32 hardware threads, and clock speeds that max out at 5.4 GHz. However, compared to the nearly identical 9955HX, the X3D version packs an additional 64MB of cache for a total of 144MB.

If you're not familiar with AMD's 3D V-cache tech, it was [9]introduced with the Milan-X Eypcs and Ryzen 7 5800X3D in early 2022, and expands the processor's L3 cache using an auxiliary SRAM tile. For cache-sensitive applications, including gaming, this can have a big positive impact on performance. This isn't actually the first time that AMD has brought this tech to mobile either, having made its debut in notebooks with the launch of the 7945HX3D in mid 2023.

[10]

In the case of the 9955HX3D, the part uses 3D packaging to stack the SRAM tile underneath the compute tile. According to AMD this helps to mitigate performance, clock, and temperature penalties observed with its first-gen implementation, which placed the SRAM on top of the compute dies.

However, this being a high-end mobile part, it doesn't get that hot with a configurable TDP — presumably this will be set by the notebook maker — of 55 to 75W.

AMD's X3D mobile chip launches alongside two new non-X3D parts including the the 9955HX, which as we mentioned earlier is essentially the same chip just without the SRAM tile, as well as the Ryzen 9 9850HX which features 12 CPU cores and 24 threads clocking to 5.2 GHz.

All three come equipped with Radeon 610 graphics on board albeit with just two GPU cores; these chips are clearly aimed at machines with dedicated GPUs onboard.

Model

Cores / Threads

Boost / Base Clock

Total Cache

GPU

GPU cores

TDP

AMD Ryzen 9 9955HX3D

16C/32T

5.4 / 2.5 GHz

144MB

Radeon 610M

2

55-75W

AMD Ryzen 9 9955HX

16C/32T

5.4 / 2.5 GHz

80MB

Radeon 610M

2

55-75W

AMD Ryzen 9 9950HX

12C/24T

5.2/ 3.0 GHz

76MB

Radeon 610M

2

55-75W

The parts are due to hit the market in the first half of 2025.

AMD fills out its X3D desktop line up

Along with new X3D mobile chips, AMD also [11]filled out its lineup of 3D V-cache-packed desktop processors with the 16-CPU-core 9950X3D and 12-CPU-core 9900X3D.

Much like the eight-core 9800X3D we put through [12]its paces in November, AMD's latest chips cram 144MB of cache on the 9950X3D, and 140MB on the 9990X3D parts, up from 80MB and 76MB, respectively, on its non-X3D siblings.

And just like 9850HX3D mobile part we discussed earlier, the 16-core desktop chip doesn't suffer any frequency penalties clocking to the same 5.7GHz boost and 4.3GHz boost as the bog-standard 9950X launched this summer.

The same can't be said of the lower spec Ryzen 9 part debuting this week. The 9900X3D tops out at 5.5GHz, a full 100MHz less than the regular old 9900X, though base clocks remain unchanged at 4.4GHz.

Despite boasting double the L3 cache and nearly identical clock speeds, AMD's stated TDP remains unchanged at 170W and 120W for the 16-and 12-core parts respectively. Max operating temperatures, which have traditionally been lower on X3D parts are also unchanged at 95C.

Model

Cores / Threads

Boost / Base Clock

Total Cache

PCIe

TDP

AMD Ryzen 9 9950X3D

16C/32T

5.7 / 4.3 GHz

144MB

Gen 5

170W

AMD Ryzen 9 9900X3D

12C/24T

5.5 / 4.4 GHz

140MB

Gen 5

120W

The latest crop of X3D desktop parts are due to arrive later this quarter.

AMD hasn't given up on handheld gaming

AMD has made a name for itself building custom SoCs for Microsoft XBox and Sony PlayStation lines, and most recently, Valve's Steam Deck. That last one has seen a wave of competitors arrive from the likes of Asus, MSI, and others, looking to get in on the handheld gaming market.

Following the success of the SteamDeck, which features a custom Ryzen processor, AMD rolled out its Z-series of processors aimed at handheld gaming.

This week, AMD unveiled a trio of Z2 chips which feature four to eight of AMD's Zen 5 CPU cores, 10-24MB of cache, and 12-16 RDNA graphics cores onboard.

Model

Cores / Threads

Boost / Base Clock

Total Cache

GPU Arch

GPU Cores

TDP

AMD Ryzen Z2 Extreme

8C/16T

5.0 / 2.0 GHz

24MB

RDNA 3.5

16

15-35W

AMD Ryzen Z2

4C/8T

5.1 / 3.3 GHz

24MB

RNDA 3

12

15-30W

AMD Ryzen Z2 Go

4C/8T

4.3 / 3.0 GHz

10MB

RDNA 2

12

15-30W

But while all three chips slide into the 15-30W TDP range (up to 35W on the Extreme), the GPUs can vary wildly based on the SKU. The base model Z2 Go features 12 RDNA 2 graphics cores on board. As a refresher, RDNA 2 made its debut in 2020. The standard Z2 packs the same number of cores, but steps up to AMD's more modern RDNA 3 tech. However, if you want AMD's latest mobile GPU architecture, you'll have to opt for the Z2 Extreme which packs 16 RDNA 3.5 cores.

How big a difference this makes given the highly constrained power and thermals of a handheld device isn't clear, but it's still worth keeping in mind for those cross-shopping handheld consoles based on the chips when they arrive later this quarter.

Zen 4 lives on

True to tradition, AMD's Zen 5 CPU cores may be the new hotness, but it hasn't given up on Zen 4 just yet.

At CES this week, AMD recycled its now two-year-old cores into 11 new Ryzen 200 and Ryzen Pro 200 processors. The parts include chips with four, six, or eight cores with boost clocks ranging from 4.7GHz to 5.1GHz at the high end, and support for AMD's older 16 TOPS XDNA NPUs on select SKUs.

Model

Cores / Threads

Boost / Base Clock

Total Cache

GPU

TDP

NPU TOPS

AMD Ryzen 9 270

8C/16T

5.2 / 4.0 GHz

24MB

Radeon 780M

35-54W

16

AMD Ryzen 7 260

8C/16T

5.1 / 3.8 GHz

24MB

Radeon 780M

35-54W

16

AMD Ryzen 7 250

8C/16T

5.1 / 3.3 GHz

24MB

Radeon 780M

15-30W

16

AMD Ryzen 5 240

6C/12T

5.0 / 4.3 GHz

22MB

Radeon 760M

35-54W

16

AMD Ryzen 5 230

6C/12T

4.9 / 3.5 GHz

22MB

Radeon 760M

15-30W

16

AMD Ryzen 5 220

6C/12T

4.9 / 3.2 GHz

22MB

Radeon 740M

15-30W

NA

AMD Ryzen 3 210

4C/8T

4.7 / 3.0 GHz

12MB

Radeon 740M

15-30W

NA

AMD Ryzen 7 Pro 250

8C/16T

5.1 / 3.3 GHz

24MB

Radeon 780M

15-30W

16

AMD Ryzen 5 Pro 230

6C/12T

4.9 / 3.5 GHz

22MB

Radeon 760M

15-30W

16

AMD Ryzen 5 Pro 220

6C/12T

4.9 / 3.2 GHz

22MB

Radeon 740M

15-30W

NA

AMD Ryzen 3 210

4C/8T

4.7 / 3.0 GHz

12MB

Radeon 740M

15-30W

NA

Notebooks featuring AMD's Ryzen 200-series processors are slated to hit the market beginning in the second quarter. ®

And over in Nvidia world...

Nvidia has [13]squeezed a Grace-Blackwell Superchip into a $3K mini PC.

It also, as expected, [14]announced its [15]Blackwell -based $1,999 RTX 5090 and $999 RTX 5080 graphics cards, due out at the end of the month, plus a $749 RTX 5070 Ti and a $549 RTX 5070 for February.

The GPU giant openly [16]released a collection of AI models called [17]Cosmos , which can generate physics-aware videos of situations from written and visual prompts.

Similarly, it [18]released its open Llama Nemotron language models and Cosmos Nemotron vision language models.

An Omniverse Blueprint dubbed Mega, for "developing, testing and optimizing physical AI and robot fleets at scale," was [19]shown off .

Get our [20]Tech Resources



[1] https://www.amd.com/en/newsroom/press-releases/2025-1-6-amd-announces-expanded-consumer-and-commercial-ai-.html

[2] https://www.amd.com/en/newsroom/press-releases/2024-6-2-amd-unveils-next-gen-zen-5-ryzen-processors-to-p.html

[3] https://www.theregister.com/2024/07/15/amd_ai_pc_goal/

[4] https://www.theregister.com/Tag/TOPS

[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Z35a1v9jyF4FcyWCI7X_7gAAAEs&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0

[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z35a1v9jyF4FcyWCI7X_7gAAAEs&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0

[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z35a1v9jyF4FcyWCI7X_7gAAAEs&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0

[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z35a1v9jyF4FcyWCI7X_7gAAAEs&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0

[9] https://www.theregister.com/2022/03/15/amd_ryzen_7_5800x3d_launch/

[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z35a1v9jyF4FcyWCI7X_7gAAAEs&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0

[11] https://www.amd.com/en/newsroom/press-releases/2025-1-6-amd-announces-new-graphics-and-gaming-products-for.html

[12] https://www.theregister.com/2024/11/07/amd_ryzen_9800x3d_review/

[13] https://www.theregister.com/2025/01/07/nvidia_project_digits_mini_pc/

[14] https://blogs.nvidia.com/blog/generative-ai-studio-ces-geforce-rtx-50-series/

[15] https://www.nextplatform.com/2024/03/18/with-blackwell-gpus-ai-gets-cheaper-and-easier-competing-with-nvidia-gets-harder/

[16] https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-world-foundation-model-platform-to-accelerate-physical-ai-development

[17] https://github.com/NVIDIA/Cosmos

[18] https://blogs.nvidia.com/blog/nemotron-model-families/

[19] https://blogs.nvidia.com/blog/mega-omniverse-blueprint/

[20] https://whitepapers.theregister.com/



Love is what you've been through with somebody.
-- James Thurber