Today AMD launches the Threadripper 9000 series in desktop and workstation guises. AMD didn’t say anything that surprised SemiAccurate, just a slew of solid parts.
The lineup is exactly what you would expect from a new generation of Threadrippers, the WX/Pro and the X/Amateur for workstation and high end desktop markets respectively. In addition to the W before the trailing X, you can tell the two chips apart by the last digit, WX ends in 5, X in 0. Oh yeah, WX is 8-channel DDR5 while X is 4-channel but that is minutia, the last digit is what is important for bragging rights at the coffee shop. Yes that is meant to be humor but the slides below showing the lineup are not.
WX means 8-channel Threadripper 9000
X is the ‘small’ 4-channel Threadripper 9000
There are a mercifully small number of SKUs, AMD seems to be doing their part to ease the PTSD brought on by the last few Xeon SKU releases. Overall there are nine Threadripper 9000s, six WX models ranging from 12 to 96 cores and three X parts which go from 24 to 64 cores. All parts have a 350W TDP and have a peak turbo clock of 5.4GHz. Clocks range from 2.5GHz to 4.7GHz which is a really strong indicator that these cores skew pretty high for binning and respond well to voltage. Overclockers should have a field day with these chips.
Two chipsets for Threadripper 9000
Coupled to these are the new chipsets/platforms called WRX90 and TRX50 with 8-channel and 4-channel support. Yes you can put an 8-channel part in a 4-channel device but not the other way around. That said given that likely MUCH higher pricing for the WX parts, AMD noxiously hasn’t released pricing yet, paying more to strangle your high core count WX with low bandwidth doesn’t seem like a good use of money to us. The X parts only have 80 PCIe lanes vs the full 128 for WX, the same as the last gen.
Threadripper 9000 platform changes
The 9000 series platform uses the same socket sTR5 but the new one has a few new tricks, mainly DRAM speed and overclocking. Yes you read that right, AMD allows you to officially overclock the DDR5 on Threadripper 9000s, yay! Standard memory is DDR5/6400 but “Supports EXPO7000+ overclockable memory”. Given how sensitive some of the target workloads are to memory bandwidth, this is much more of a big deal than a simple enthusiast toy. And a tip of the hat to Supermicro, the only OEM at the presentation that officially supports overclocked memory, good job guys. Like all new CPUs, TR9000 only supports 1DPC for a max of 2TB per system. ‘Only’ 2TB,,, /me glances over to the 20MB 5.25″ Seagate RLL drive on his shelf…
Speaking of HDDs, you will be overjoyed to hear that TR9000 supports up to 32 SATA lanes. If you can find a board with SATA ports, it could happen even if we aren’t sure why. Since the PCIe lanes can be bifurcated down to 1x, you could just slap 128 cheap SSDs in the box and get vastly more storage, bandwidth, and save money too. If anyone does this, please send me pictures, I will write it up. On a related note, anyone else craving spaghetti all of a sudden?
There are a bunch of other security improvements and platform improvements too, many of which stem from the changes to the CPU itself. AMD claims a big improvement in I/O bandwidth due to internal plumbing changes, something we think is a valid claim. The 16% IPC increase for the cores however is a bit higher than we recall being the case but I am pretty sure AMD can back that up with a few benchmarks and slides.
AMD Zen 5 CCD Overview
The above image being one of those slides. The CCD itself has massively widened paths around the L2, double in most cases. Why? Mainly because the cores now support AVX-512 at full width so you need that bandwidth to sustain full rate on those units. There are some caveats on the load/store units, basically you can do 4 loads/clock unless you are doing 512b loads in which case you can only do 2. Deep sources tell us this means the data path is now 1024b wide, either that or it was basic math. For stores the number is 2, 1 for 512b transfers. L3 is improved but the size remains at 32MB, ignore the annoying “total system cache” nomenclature, anyone interested in the tech shouldn’t be talked down to like that.
AMD Zen 5 Microarchitecture Overview
The Zen 5 core itself is not new so we will just gloss over some of the highlights. First off is that AMD increased the ALU count from four to six which obviously has big IPC potential. The I- and D-caches have increased bandwidth and the D-cache is enlarged. On top of that, almost every data path is widened, again to support AVX-512 ops without waiting for data. In the CPU world, very small sounding changes can have big ripple effects.
In the aggregate, Threadripper 9000 is a bit better at everything. Software with heavy AVX-512 optimizations will obviously have a big boost, double the throughput vs Zen 4 is technically possible but real world unlikely. Most of the benchmarks supplied seem to have a roughly 20% uplift vs the top of the line Threadripper 7000, quite the worthwhile gain in productivity. Our final opinion on the parts hinges on price, something which AMD declined to reveal. This is stupid, you can’t recommend something with no price, it is only a downside for an otherwise solid part.S|A
