One of the strange parts of this year is that we are having a virtual Computex 2021. That puts many of the keynotes on a holiday weekend in the US and in the evenings when generally PR folks would consider posting news they wanted to not be seen. Still, AMD is forging ahead after last evening’s Intel Computex 2021 Keynote Coverage and NVIDIA will be next. Here, Dr. Lisa Su, CEO of AMD will have a number of announcements. We are covering this live so please excuse typos.
AMD Computex 2021 Keynote Coverage
Something exciting here, AMD is starting with the data center!
AMD says it will have over 100 EPYC servers and over 400 EPYC-based cloud instances available this year. The company says it has twice the solutions compared to Rome. AMD showed a Java benchmark versus the Intel Xeon Platinum 8380 and said it was ~50% faster with just over 50% more cores. We did not see a lot behind this.
With the current challenge of the GPU supply, AMD is launching new G series APUs that have integrated graphics. The AMD Ryzen 5700G has 8 cores and 8 graphics cores. The AMD Ryzen 5 5600G has 6 cores and 7 graphics cores at $259.
As someone who drives a Tesla Model S (albeit NVIDIA powered), this is exciting. The new Model S and Model X utilize AMD RDNA 2 for their infotainment systems. Hopefully, this is the new Plaid generation and will be in my future Cybertruck that is on order.
This was previously announced, but AMD is also going to be in next-gen Samsung processors.
Next, AMD is discussing mobile gaming which is important these days with mobile workforces.
AMD has some bigger GPUs including now with 12GB of GDDR6 in its mobile part. The Radeon RX 6700M will have a similar clock speed but with 10GB of GDDR6. The AMD Radeon 6600M 2177MHz clock and 8GB GDDR6.
Something cool is that AMD is discussing FSR. FidelityFX Super Resolution. Here, AMD is using upscaling to get more performance while still getting solid visual fidelity.
AMD says it will be supported on its RX 6000, RX 5000, RX 500, Vega, and also Ryzen products with onboard graphics. That could be fun for something like the Lenovo ThinkCentre M75q Gen2 Tiny. AMD says it will work even on NVIDIA GPUs. The first set of games AMD says will be available by June 22, 2021.
AMD seems to be going down the path Intel has been working on with its Evo brand and is starting to brand frameworks for notebooks. AMD Smart Access Memory will be coming to AMD Advantage notebooks with the new Radeon 6000 series notebooks.
AMD is ending the keynote with advanced technology. AMD said that Zen4 will be available in 2022.
The company is now fighting back against Intel’s packaging messaging. Specifically, the company is looking at using 3D chiplets to start stacking memory with the logic using TSV technology.
In the Ryzen 5000 series test chip, it is stacking a 64MB L3 cache die atop the compute die using a copper-to-copper bond that it has been working on with TSMC.
Here is a shot of the test chip with the 64MB die stacked.
Effectively, this shows we can expect AMD to continue down the path of offering large on-chip caches. Already AMD’s CPUs can offer up to 32MB of L3 cache per core. With this, we would expect that figure to increase. The other side is that we could also see less compute die area spent on cache with it instead moved to a larger stacked chip. This gives the area to build chips with higher density. It also allows AMD to, for example, make a large cache using 7nm while migrating its compute cores to 5nm just as it decoupled the I/O die from the compute die with the AMD EPYC Rome and Milan series. Expect TSMC to use AMD as one of its showcase customers as Intel pushes its next-gen packaging. While some may see Intel’s packaging marketing as directed to AMD, it is also being directed at TSMC.
3D chiplet prototype gives 12% faster FPS on the Ryzen 5000 series platform with a bigger cache. The company says in gaming this is giving around a 15% average increase in performance. This packaging will start going into production later this year.
Overall some cool bits in this demo. From AMD’s methodology on developing Zen-based cores and then leveraging across desktop and server, that 3D packaging is likely destined for the EPYC space. Eight dies with an additional 64MB each means that just based on this tech demo that would be an additional 512MB of cache on a Milan style chip using these 7nm SRAMs.
Get excited for 2022. More cache, bigger chips, CXL, and more are coming.