A sourced overview of current AI accelerators, as of November 2026

This is a snapshot, dated 12 November 2026, of what the manufacturers themselves say about their AI hardware. It is deliberately not a ranking. Accelerators are bought for different jobs, the vendors quote performance in incompatible units, and the only fair comparison we can make from primary sources is memory: how much, and how fast. We say which figures are manufacturer claims, and we leave out anything we could not read on the maker's own page.

Consumer and desktop

GeForce RTX 5090. NVIDIA's flagship consumer card carries 32 GB of GDDR7 on a 512-bit interface with a total graphics power of 575 W and a recommended 1000 W system supply. NVIDIA rates its AI performance at 3352 AI TOPS. That is a manufacturer figure, and the page does not state the number format or whether sparsity is assumed, so do not compare it with a TOPS number from another vendor. We could not find a memory-bandwidth number on the page, so we do not quote one.

Apple silicon Macs. The Mac Studio page lists the M5 Max with up to 128GB of unified memory at 614GB/s, and the M5 Ultra with 256GB or 512GB at 1.2TB/s. Apple's 480W figure is a maximum continuous power rating for the machine.

Ryzen AI Max+ systems. AMD says its Ryzen AI Max platform provides up to 128GB of unified memory, with up to 96GB available for graphics, plus an XDNA 2 NPU rated at up to 50 TOPS. At CES 2026 AMD also announced the Ryzen AI Halo developer mini-PC with up to 128GB of unified memory, planned for the second quarter of 2026. We have not confirmed that it has shipped.

NPUs in ordinary PCs. Microsoft's Copilot+ PC programme asks for an NPU capable of 40+ TOPS. That sets the baseline for new Windows laptops; it does not make them big-model machines. Our CPU, GPU and NPU guide explains why.

Workstation

RTX PRO 6000 Blackwell. NVIDIA's professional card comes with 96 GB of GDDR7 with error-correcting code, in a 600 W Workstation Edition, a 300 W Max-Q Workstation Edition and a Server Edition rated 400-600 W. For one-card local inference this is the largest memory on a single GPU card in this article outside the data centre.

DGX Spark. A small desktop system built on NVIDIA's GB10 chip, with 128 GB of coherent unified memory, a 273 GB/s interface and a 240 W power supply. NVIDIA claims up to 1 PFLOP of FP4 AI performance, which should be read as a peak figure at a low-precision format. Up to four can be linked over ConnectX-7 networking.

Data centre

Some notes on reading this table.

Per chip versus per system. The DGX B200 row is eight GPUs; dividing its 1,440 GB and 64 TB/s by eight gives a per-GPU picture, but we are quoting NVIDIA's system totals rather than doing the division for you. The GB300 NVL72 figure is a 72-GPU rack. Comparing these lines with a single H200 without that adjustment is the classic mistake.

Units. Google documents the TPU7x in GiB (binary) and GBps, and a GiB is about 7 per cent larger than a GB. Check units before concluding that one chip beats another.

Google. The same documentation lists Ironwood at 4614 peak FP8 TFLOPs per chip and 9,216 chips in a pod, alongside its older v5p and v6e generations. TPUs are available through Google Cloud rather than as a product you buy.

AMD. The MI325X row comes from a June 2024 roadmap announcement in which AMD said the accelerator "will bring" 288GB of HBM3E and 6 terabytes per second, with general availability planned for Q4 2024. The same release said the MI350 series, built on the CDNA 4 architecture and a 3nm process, would offer up to 288GB. At Advancing AI 2025, AMD unveiled the Instinct MI350 Series, consisting of the MI350X and MI355X. The performance multiples in that release rest on AMD's own internal testing and projections, and we do not reproduce them. We could not retrieve a final MI350 specification sheet from AMD's own product pages, which block automated fetches, so there is no MI350 or MI355X memory row.

Not covered. We do not include Intel's Gaudi line or other custom accelerators because we could not verify their current specifications from a first-party page in this research pass.

How to read the performance claims

Every vendor here quotes a peak compute figure, and they are not interchangeable. NVIDIA's DGX Spark number is "up to" 1 PFLOP at FP4, a four-bit floating-point format. Google's Ironwood figure is quoted at FP8 and BF16 separately. NVIDIA's RTX 5090 figure is given in "AI TOPS" without the format stated on the page. Three different units, three different precisions, and each is a theoretical peak rather than a measured result. Before comparing two numbers, ask what format they assume, whether structured sparsity is counted, and who ran the test. If the answer is "the manufacturer, in a lab", treat it as a claim.

What the memory numbers tell you

Put the figures side by side and the stack is easy to see. Desktop cards are at 32GB and 96GB. The unified-memory machines reach 128GB and, in Apple's case, 512GB, but at bandwidths that run from 273GB/s to 1.2TB/s. Data-centre chips run to 141GB to 288GB each, at bandwidths in the several-terabytes-per-second range: 4.8TB/s on the H200, 7380 GBps on Ironwood. Our piece on CPU, GPU and NPU inference shows why that last column matters most for the speed of a single response.

That is also why capacity and bandwidth are traded off differently in each class. Home hardware leans towards capacity; data-centre hardware leans towards bandwidth and, through rack-scale systems, towards pooling many chips' memory over fast links. NVIDIA quotes 130 TB/s of NVLink bandwidth for the GB300 NVL72 rack.

What we cannot say

We have not benchmarked any of this hardware and publish no performance comparisons. Prices are not included, because they change faster than specifications and none were verified. This overview will date quickly: new parts arrive every few months, and anything marked "planned" or "expected" above should be checked against availability before you act on it. A corrections process exists for exactly this reason.