Is an APU the same as an SoC?

An APU is a kind of system on a chip. AMD’s technical article describes the Ryzen AI Max+ 395 as an SoC that sets a new standard for an APU in memory capacity and generative AI.

What does Strix Halo mean?

Strix Halo is AMD’s codename for the Ryzen AI Max family. AMD’s page for the 395 lists it as the former codename.

Can an APU run a 70-billion-parameter model?

AMD says the Ryzen AI Max+ 395 with 128 GB of unified memory can run 70-billion-parameter models on the device. Speed depends mainly on memory bandwidth, so ask for measured results at your precision and context length.

APU or discrete GPU for an office AI?

An APU fits larger models in one small box, while a discrete card reads its memory faster. Which suits you depends on model size and how many people share it, which our Strix Halo or RTX question covers.

ANVE: the private AI office, on the record. See what that means

ANVE

Private AI glossary · updated 2026-10-16

APU
One chip with CPU, GPU and NPU, and one pool of memory for all three.

Short answer. An APU (accelerated processing unit) is a single chip that combines CPU cores and a GPU, often with an NPU, sharing one pool of memory. AMD’s Ryzen AI Max+ 395, codenamed Strix Halo, has 16 Zen 5 CPU cores, a Radeon 8060S GPU with 40 graphics cores, an NPU rated up to 50 TOPS and 256-bit LPDDR5x memory up to 128 GB. For AI the point is memory: AMD says the GPU can use up to 112 GB of that pool, far more than the 8 to 24 GB of dedicated memory typical of a discrete graphics card.

What is on the chip

AMD’s specification page for the Ryzen AI Max+ 395 lists the following.

Part	AMD’s listed specification
CPU	16 Zen 5 cores, 32 threads, up to 5.1 GHz boost
GPU	Radeon 8060S, 40 graphics cores, 2,900 MHz
NPU	Up to 50 TOPS (the page also lists up to 126 TOPS overall)
Memory	256-bit LPDDR5x, up to 128 GB, up to LPDDR5x-8000
Power	Default TDP 55 W, configurable from 45 to 120 W
PCIe	PCIe 4.0, 16 lanes
Former codename	Strix Halo

Why one memory pool matters

On a discrete card the GPU’s memory is separate. AMD’s ROCm blog puts it plainly: a GPU with 24 GB of VRAM can only hold models that fit in those 24 GB, whatever system memory exists. On an APU the CPU and GPU share the same physical pool. AMD says the Ryzen AI Max+ 395 can have 128 GB of unified memory, with up to 112 GB allocatable to the GPU, and that sharing lets data move between processors without expensive memory copies.

The share the GPU gets is a setting. In AMD’s own test system, 64 GB of a 128 GB pool was assigned to the GPU through the BIOS, and a 76 GB model had to be split between GPU and CPU.

Trade-offs against a discrete card

Point	APU with shared memory	Discrete GPU card
Largest model that fits	Set by the pool: up to 128 GB on the 395, with up to 112 GB for the GPU (AMD)	Set by the card: 24 GB on the RTX PRO 4000 Blackwell (NVIDIA)
Memory bandwidth	256 GB/s, calculated from a 256-bit bus at 8,000 MT/s	672 GB/s listed for the RTX PRO 4000 Blackwell
ECC	Not listed on AMD’s page for the 395	ECC GDDR7 listed by NVIDIA
Power	45 to 120 W configurable for the whole chip	145 W maximum for the card alone
Software	AMD’s ROCm matrix lists the 395 (gfx1151)	NVIDIA’s CUDA

Questions to ask a vendor

How much of the memory can the GPU address, and who sets it?
What are the memory speed and bus width? Ask for the bandwidth in GB/s.
What power limit does this machine run the chip at, and can it hold it under a long job?
Which operating system and runtime versions does the maker support, for example ROCm?
Is ECC memory available, or not?

Where ANVE stands

ANVE One is specified around an AMD Strix Halo class chip with 64 GB of unified memory; ANVE Office uses a separate NVIDIA graphics card instead. The software runs today on a development machine. No sealed unit exists yet, and ANVE has published no speed figures.

Questions people ask

Is an APU the same as an SoC?: An APU is a kind of system on a chip. AMD’s technical article describes the Ryzen AI Max+ 395 as an SoC that sets a new standard for an APU in memory capacity and generative AI.
What does Strix Halo mean?: Strix Halo is AMD’s codename for the Ryzen AI Max family. AMD’s page for the 395 lists it as the former codename.
Can an APU run a 70-billion-parameter model?: AMD says the Ryzen AI Max+ 395 with 128 GB of unified memory can run 70-billion-parameter models on the device. Speed depends mainly on memory bandwidth, so ask for measured results at your precision and context length.
APU or discrete GPU for an office AI?: An APU fits larger models in one small box, while a discrete card reads its memory faster. Which suits you depends on model size and how many people share it, which our Strix Halo or RTX question covers.

What is not real yet

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.

APUOne chip with CPU, GPU and NPU, and one pool of memory for all three.