Threadripper 9000 And Radeon AI Pro 9000: AMD’s Workstation Power-Up Unveiled
AMD's HEDT And Pro Workstation Line-Up For 2025 Is Its Most Powerful Yet
We’ve already covered the initial announcements out of Computex and published a couple of deep dives for Zen 5 and RDNA 4, so if you want to learn and the juicy details you’ve got some reading to do. We’ve got some new product details and performance data to share today, however, a bit of review is probably in order. First up, let’s discuss the Ryzen Threadripper 9000 series...
Introducing AMD's Ryzen Threadripper 9000 Series
The new Threadripper 9000 series is all about increased data throughput, performance, and platform bandwidth. We’re entering somewhat of a new era with professional workstations where traditional engineering, scientific, media and entertainment, and design workloads are as prevalent as ever. Nascent AI and machine learning workloads, however, will also benefit from the massive bandwidth and parallel compute capabilities of the Threadripper 9000. With that in mind, Threadripper 9000 is well positioned to power a fresh wave of workstations that will benefit significantly from its massive amounts of compute, bandwidth, and IO. Quad GPUs with full PCIe bandwidth are possible on this platform.
A summary of what Zen 5 offers is outlined in the slide above. Zen 5 offers higher IPC thanks to improved fetch and decode, wider dispatch and execution engines, and increased cache bandwidth. Integer, Floating Point, and Vector math advances have been made as well.
Zen 5 also features better branch prediction with lower latency than Zen 4. The architecture features dual decode pipelines that utilize more sophisticated and accurate TAGE (TAgged GEometric) prediction algorithms. Instruction cache latency and bandwidth have been improved too.
The architecture features an 8-wide dispatch / retire engine -- versus 6-wide in Zen 4 -- with 6 ALUs (3 Multiples) and a new unified ALU scheduler design, whereas older Zen architectures featured a unique scheduler per ALU. As a result, Zen 5 has a 40% larger execution window that can support up to 448 outstanding operations. The architecture can essentially handle more instructions and better recover from misses.
AMD also enlarged the Zen 5 data cache by 50% (32K vs. 48KB) and doubled the maximum bandwidth to the L1 cache and FPU. Zen 5 features a 48KB 12-way L1 data cache, but maintains a 4-cycle load to keep latency low, despite the higher capacity. The bandwidth from L2 to L1 has also been doubled versus Zen 4.
In Zen 4, AVX-512 is implemented using double-pumped 256-bit data chunks, so 512-bit AVX-512 workloads were broken down into two 256-bit chunks and sent through the engine. Zen 5, however, features a full 512-bit wide data path. The FP/vector math unit in Zen 5 features six pipelines with two-cycle latency for FADD instructions, which is a reduction from 3-cycles in Zen 4. They can also handle a larger number of in-flight FP instructions.
As we’ve mentioned a number of times in the past, Zen 5 cores offer an average 16% IPC uplift versus Zen 4. That obviously translates to better single-thread performance at similar clocks, but in aggregate, multi-thread performance is improved as well, once that uplift is scaled across many cores.
Threadripper 9000 Works In Existing Motherboards
Threadripper 9000 series processors are drop in compatible with existing WRX90 and TRX50 platforms, which both features the sTR5 socket. There will probably be some new motherboards coming alongside the processors, but they’re technically not a requirement – existing boards will work after a BIOS / UEFI update.
When used in WRX90 motherboards, Threadripper Pro 9000WX processors can be configured with up to 8-channel memory and they feature 148 total PCIe lanes, 144 of which are usable, and 128 of which are PCIe Gen 5. The TRX50 HEDT platform will max out with 4-channel memory and a total of 92 PCIe lanes (88 usable, 48 PCIe 5). Both platforms require DDR5 RDIMMs, just like the Threadripper 7000 series, but faster speeds will be officially supported.
In terms of their memory capabilities, Threadripper Pro 9000WX processors will support up to 2TB of system memory at data rates up to DDR5-6400, for peak bandwidth of up to 410GB/s in an 8-channel configuration. For non-Pro HEDT Threadripper 9000s, peak bandwidth is halved, due to its maximum of 4 memory channels. Higher speeds are available through overclocking, however, and the platform will support AMD EXPO memory profiles for quick and easy configuration.
The Threadripper Pro 9000WX series also differs from its standard HEDT counterparts in that it will offer AMD’s “Pro” security and manageability technologies. The actual build and packaging is different as well. As you can see in the images below, the Threadripper Pro 9000WX have four additional compute dies. The HEDT parts aren't the same parts with cores disabled -- they actually lack the compute dies for those cores.
These latest Threadrippers scale to 96 cores and 192 threads, just like the previous generation line-up; the main difference is that the Threadripper 9000 series is based on AMD's Zen 5 architecture.
With the HEDT parts, AMD has introduced a trio of models. Sitting at the top is the Ryzen Threadripper 9980X, a 64-core/128-thread chip with a 3.2GHz base clock, up to a 5.4GHz boost frequency, and a generous heaping of total cache at 320MB, including 256MB of L3 cache. Next is the Ryzen Threadripper 9970X with 32 cores, 64 threads. It features a 4GHz base clock, a 5.4GHz max boost clock, 128MB of L3 cache, and 160MB of total cache. Finally, there's the Ryzen Threadripper 9960X with 24 cores, 48 threads, a 4.2GHz base clock, up to a 5.4GHz max boost clock, 128MB of L3 cache, and 152MB of total cache.
While the HEDT models top out at 64 cores and 128 threads, the Pro 9000 WX series scales much higher. The flagship is the Ryzen Threadripper Pro 9995WX with 96 cores, 192 threads, a 2.5GHz base clock, up to a 5.4GHz boost frequency, a massive 384MB of L3 cache, and 480MB of total cache. On the other end of the spectrum, the Ryzen Threadripper Pro 9945WX features a tamer configuration with 12 cores, 24 threads, a 4.2GHz base clock, up to a 5.4GHz boost clock, 64MB of L3 cache, and 152MB of total cache.
Threadripper 9000 Expected Performance
In terms of their expected performance levels, AMD presented quite a bit of data versus both previous-gen Threadrippers and Intel's Xeon W workstation processors.
Note that this comparison pits the 64-core HEDT Threadripper 9980X against Intel's 60-core Xeon w9-3595X. As you can see, the gains vary but can be massive and the Threadripper can occasionally offer more than double the performance.
And now for some Threadripper Pro comparisons. Versus the previous-gen Threadripper Pro 7995WX, which has a similar core count, the upcoming Threadripper Pro 9995WX offers up to 26% higher performance in multi-threaded workloads. We expect improvements in responsiveness and single-thread workloads as well.
When you compare a 96-core Threadripper Pro 9995WX to Intel's top Xeon W processor, the Threadripper Pro pulls even further ahead. Whether its for design work, media production or AI / ML workloads, according to AMD's numbers, the new Threadripper Pro makes minced meat of the current top-end Xeon W. We'll try to verify some of these numbers, assuming we get to review the new Threadrippers at some point in the future.
AMD Radeon AI Pro 9700: RDNA 4 For Engineers And Pro-Vis
Which brings us to the Radeon AI Pro 9700. At its core, the Radeon AI Pro 9700 is similar to the consumer-class Radeon RX 9070 XT, but it features a totally different board design with double the memory, which makes it much better suited for Pro-Visual and AI workloads.
The Radeon AI Pro 9700 is built around the RDNA 4-based Navi 48 GPU. The GPU is arranged into 4 shader engines, each with 8 workgroup processors, which are equipped with two compute units each. This is a similar structure to previous gen, RDNA 3 GPUs, but there are many architectural enhancements, which improve performance and efficiency and add new features.
Within each CU there are updated 3rd generation Ray Accelerators that offer 2x the peak throughput of RDNA 3. RDNA 4’s Ray Accelerators offer double the ray intersection rates, with improved BVH compression, accelerated ray traversal and shading, and add support for a feature called Oriented Bounding Boxes. 3rd Generation Matrix Accelerators are also present, which offer improved performance, along with support for 8-bit float data types, with structured sparsity support as well. RDNA 4 is better suited to AI / ML workloads and it supports additional data types.
The GPU features 2MB of aggregate CU cache, along with 8MB of L2 cache, and 64MB of 3rd generation Infinity Cache situated out at the edges, between the memory controllers and shader engines. The L2 cache surrounds an updated, centrally located Command Processor, with enhanced packet acceleration. The updated Command Processor also supports some new instructions and it features improved branch prediction, to better predict data fetches and keep caches better utilized.
On the Radeon AI Pro 9700, the GPU connects to 32GB of GDDR6 memory operating at an effective data rate of 20Gbps, over a 256-bit interface (four, 4x16 memory controllers). The memory controllers also feature enhanced memory compression technology, to make better use of available bandwidth.
The Radeon AI Pro 9700 features the full implementation of the Navi 48 GPU, with 64 total CUs, 64 Ray Accelerators, and 128 AI (Matrix) Accelerators. It offers up to 1,531 peak TOPs (Int4 with Sparsity) with a boost clock of about 2.9GHz and a total board power of 300 watts. That’s slightly lower than the Radeon RX 9070 XT, for those keeping track.
Dual media engines and an updated Radiance Display Engine are present on the GPU as well. There’s actually a number of updates with the media and display engines, and we’ve got much more detail regarding gaming / graphics related updates – but you’re going to have to read our RDNA 4 deep dive for the full scoop. There’s too much to ground to cover for our analysis today.
In a major departure from previous Radeon Pro cards, AMD has decided to allow partners to bring Radeon AI Pro 9700 cards to market. AMD made this move to broaden availability and to bring its Pro hardware and software to a wider range of customers. AMD also hopes to work with its partners to develop new business on ROCm and Radeon.
Radeon AI Pro 9700 Expected Performance
With its faster memory and newer architecture, the Radeon AI Pro 9700 should offer some significant performance improvements over previous-gen Radeon Pro cards as well.
The GPU offers up to 4X more TOPs with some data types, but in terms of FP32 performance the Radeon AI Pro 9700 is only marginally more performant than the Radeon Pro W7800.
Even with similar amounts of memory, however, the upcoming Radeon AI Pro 9700 can offer hige performance improvements over the RDNA 3-based Radeon Pro in terms of tokens per second. And as we saw in our Radeon RX 9070 XT review, the Navi 48 GPU offers a big boost for rendering and gaming versus previous-gen Radeons as well.
AMD Threadripper 9000 And Radeon AI Pro For Workstations: Key Take-Aways
All of the things that have made Zen 5 and RDNA 4 successful in the consumer space are coming to professional workstations, and we expect AMD's going to enjoy plenty of success. We'll obviously hold final judgement until we get to actually test the products, but things are looking quite good at this point.
