AMD held a briefing last week on the AMD Ryzen Threadripper 9000 series parts as well as the AMD Radeon AI Pro R9700 series GPUs. We can now share some of the updates on the upcoming parts that we expect to arrive this summer.
New AMD Ryzen Threadripper 9000 Series
The AMD Ryzen Threadripper 9000 series and AMD Ryzen Threadripper Pro 9000 series extend AMD’s workstation lead over Intel’s solutions by updating platforms to Zen 5 with higher memory frequencies. AMD had many slides in its presentation, but here is the slide that has those key points with one other important fact: it is a same-socket (sTR5) upgrade in existing Threadripper 7000 series platforms.
Going back to the design objectives, we get a similar maximum core count and 350W TDP, but we get the updated AVX-512 datapath, Zen 5 generational improvements and more. Remember, the DDR5-6400 is the base speed, but we fully expect faster ECC RDIMM modules to come online for these platforms.
Speaking of platforms, here is the platorm summary which should look familiar.
There are two versions of the platforms. First, there is the higher-end 8-channel DDR5, 128 PCIe lane WRX90 platform. Second, there is the TRX50 with 4-channel DDR5 and 80 lanes of PCIe.
On the AMD Ryzen Threadripper Pro 9000WX line, there are six new SKUs but still a maximum of 350W TDP.
The AMD Ryzen Threadripper 9000 series is the high-end desktop (HEDT) part that only goes in the TRX50 platforms with 4-channel memory.
For those wondering, the reason for having the TRX50 is simply to decrease the cost of motherboards. Standard server motherboards now cost well over twice what they did eight years ago and a lot of that is driven by more and faster I/O in terms of PCIe and DDR5. Fewer memory channels and PCIe lanes is a way to lower the cost of the systems by requiring the motherboards to be less complex.
AMD had a bunch of other slides, but if you are very familiar with the Threadripper 7000 series, then this takes AMD’s leadership in workstation processors and increases the gap by usually 15-25% without increasing core counts.
New AMD Radeon Pro AI R9700 32GB GPU
On the AMD GPU side we had new slides on the AMD Radeon Pro AI R9700. This new GPU uses RDNA 4 graphics and GDDR6 memory. AMD told us that 32GB capacity is a bus limit which is why it is not trying to match the NVIDIA RTX Pro 6000 series 96GB on a single GPU.
Still RDNA 4 has an updated compute engine to provide more AI performance.
Here are the key speeds and feeds of the new R9700 GPU.
The big driver is that having more raw compute FLOPS and TOPS and support for things like FP8, AMD is adding a lot more AI performance in this generation.
In this segment, having a lot of memory and decent AI compute is a big deal. Also, with ROCm 7, AMD is (finally) committing to the notion that it needs to get developers to use ROCm from their mobile and desktop workstations all the way to the larger HBM 8-GPU platforms like the MI325X and MI355X.
Final Words
AMD has a strong leadership position in the workstation market. Intel is still selling Sapphire Rapids-based Xeon W workstation CPUs. Or another way to look at it is that Intel is selling derivatives of almost 10 quarter old/ two generation old server CPUs into the workstation market while AMD is updating to the Zen 5 architecture that powers its current generation of server and desktop CPUs. Xeon W is in trouble just like the Xeon 6300 series due to Intel’s low pace of putting out newer generation parts.
On the GPU side, the R9700 is a neat idea, but it will be interesting to see how it is adopted. AMD is not competing just with NVIDIA here. We suspect that AMD is also competing with its own Ryzen AI Max+ 395 here. While there are many reasons the R9700 is superior, the Ryzen AI Max allows one to configure up to 96GB of the 128GB LPDDR5X memory in the maximum configurations as GPU memory. In other words, if you are more GPU memory capacity bound than anything else, the notebook Ryzen AI processor might offer a tempting soluttion.
Hopefully we get to try these platforms out soon. We may have some new high-speed memory modules waiting for them.
Threadripper needs integrated graphics like AI Max+ 395. Chiplet assembly to the rescue.
KGB
Agree with that, but the AI Max+ puts the AI NPU/GPU in the SOC, which makes sense since it’s highly MEM IO dependent, but doesn’t lend itself to replacing the Epyc/TR SOC. The epyc/TR SOC is chock full of pcie lanes and infinity fabric for io. I don’t think you can directly replace this with a strix halo soc without limiting mem and io which is a non-starter in the workstation market.
I want to see a threadripper with on-package hbm and half the chiplets replaced with an RDNA4+ GPU. Instead of getting 96C TR, you’d only get 48C (6x classic zen5 or maybe zen5c or x3d?) but you end up with 6x rdna chiplets (they don’t make these as chiplets yet) with HBM3/4 for on-die faster memory. Maybe that’s 3x 36GB hbm stacks and 3x rdna chiplets on the package? Having a 100GB GPU along with 48 cores of zen5 in a package like that would scream. Socket compatibility would be easy enough as the interconnect is already in package.
Intel V series chips get the HBM part right, Apple gets it more right with larger memory but they aren’t modular.
@KGB, you could use a Threadripper MB with IPMI.
I don’t understand why they didn’t make these WRX90/TRX50 parts capable of 12/6 memory channels, instead of 8/4. Now there is no choice but to buy EPYC and lose clock frequency if you need the memory bandwidth.