by 
RDNA 4 makes its first appearance in workstations with the 32GB R9700.
Computex AMD is working towards strengthening its advantage over Intel in the high-end desktop (HEDT) and workstation markets by hinting at the upcoming release of their 9000-series Threadripper workstation processors during Computex this week.
When compared to earlier Threadripper models, the 9000 series seems like a rather modest upgrade, primarily benefiting from advancements in manufacturing processes and the shift to AMD’s new Zen 5 architecture.
Released simultaneously with the 9000-series Ryzen Desktop processors last summer, AMD’s Zen 5 architecture showed a 16% improvement in instructions per clock compared to Zen 4. Following this launch, AMD extended this architecture to both its Epyc server lineup and its Threadripper workstations series.
Similar to the previous generation, AMD’s newest Threadripper processors come in both Pro and non-Pro versions, featuring up to 96 cores, 128 PCIe 5.0 lanes, eight DDR5 6400 MT/s memory channels with ECC support, and turbo speeds reaching up to 5.4GHz for all models.
AMD’s workstation-grade TR Pro 9000 components will come with SKUs spanning from 12 to 96 cores, whereas the non-Pro high-end desktop (HEDT) versions will offer a more limited range of configurations, going from 24 to 64 cores.
In history, AMD’s regular Threadripper lineup has included fewer memory channels and PCI Express lanes when compared to their Pro versions. To illustrate this point with an instance from 2023: The consumer-oriented TR 7000 series processors came equipped with only four DDR5 memory channels and offered 48 PCIe 5.0 lanes as opposed to the eight memory channels and 128 PCIe lanes provided by the corresponding Pro models.
We have requested AMD for clarity regarding whether this applies to both the TR and TR Pro 9000 series.
Below is a detailed look at theThreadripper components unveiled today:
AMD has not released retail prices for their newest line of Threadrippers before their Computex keynote later this week. That being said, since Threadripper faces little competition in the high-end desktop segment, AMD can set prices based on what the market will tolerate.
Intel’s most recent lineup of workstation components, introduced during the previous summer, utilizes their somewhat outdated Sapphire Rapids platform from about two years ago. These new offerings max out with up to 60 cores, support for 112 lanes of PCIe 5.0 connectivity, and can handle eight channels of DDR5 memory running at 4800 megatransfers per second (MT/s).
AMD's new 9000 series of Threadripper CPUs are set to launch in July.
Alongside its newly announced Threadripper processors, AMD hinted at a forthcoming workstation GPU equipped with its advanced RDNA 4 graphics architecture and boasting 32GB of video RAM onboard.
The Radeon AI Pro R9700 seems to draw from the chip industry’s earlier announced RX 9070 XT graphics card, incorporating identical 128 AI accelerators. These components deliver approximately 96 teraFLOPs of FP16 performance along with around 1531 TOPS of sparse INT4 capabilities.
As we previously covered, AMD’s RDNA 4 architecture introduces several significant enhancements for AI tasks, such as backing hardware sparsity along with FP8 data types.
At the time of writing, AMD had not yet disclosed the retail price for the R9700.
When compared to the Intel Arc Pro B60 we examined previously, the 300-watt R9700 is expected to provide roughly double the performance for tasks involving dense INT8 workloads, like AI inference. Additionally, if you opt to reduce the precision to 4 bits and sacrifice some image quality, the R9700 could achieve up to four times the performance.
Nevertheless, this chip does not quite measure up to Nvidia's premier workstation equipment. Nvidia’s RTX Pro 6000 workstation GPUs unveiled during their March GTC event feature an impressive 96GB of VRAM along with up to 4 petaFLOPS of sparse INT4 performance. This makes these cards roughly 2.6 times quicker compared to AMD's most recent line of professional graphics solutions.
This might be the reason AMD chose to benchmark the R9700 against Nvidia's consumer-oriented RTX 5080, claiming it delivers as much as 496 percent better performance for executing sizable AI applications.
Although this might seem impressive, the chart is somewhat deceptive and primarily indicates that increased video RAM allows for running larger models without offloading data to the less speedy system memory.
It seems that what AMD has done is select models that are too big to be accommodated within the RTX 5080’s 16GB of VRAM, causing these models to overflow into the computer’s less speedy DRAM instead. This undoubtedly devastates performance for tasks reliant on memory bandwidth, such as AI inference.
If you were to conduct these same tests using a smaller model that fits within the 5080’s vRAM, such as Qwen 3 14B Q4, we anticipate that the outcomes would be noticeably distinct. This difference arises due to the fact that although both GPUs have a 256-bit memory interface, the 5080 utilizes significantly quicker GDDR7 memory modules. Consequently, this allows the 5080 to deliver up to 960GB/s of memory bandwidth.
Although AMD has not provided memory bandwidth numbers for the R9700, we can assume it’s roughly similar to the 9070 XT’s 640GB/s. Based on this assumption, we estimate that the 5080 could potentially offer performance around 1.5 times faster than the 9070 XT.
Regarding AI, having quick memory doesn’t benefit you significantly unless you possess ample storage for your model. Thus, AMD isn’t completely unfounded; there are merely more accurate methods to convey this idea.
The R9700 is set to arrive on store shelves in July, coinciding with the release of AMD's Threadripper CPUs.
No comments:
Post a Comment