GPU rumors 2021

[smatovic]

[...]

Hehe, back to the 90s with SPARC...IIRC the PTL/FH-Wedel university had an SUN Ultra SPARC running in the data center, with 4 sockets 200 MHz 64-bit, must have been fun, to code a chess engine for that machine back then...
https://www.chessprogramming.org/SPARC#Chess_Programs

AFAIK, Stockfish has no SPARC optimized code, at least not for the SPARC "VIS" SIMD unit in regard of NNUE inference, and I have never seen SF benchmarks for these machines.

As far as I got it, the Fujitsu SPARC64 XII is the last CPU to support Sun Solaris, think of Oracle DB legacy machines for finance and insurance companies:
https://www.fujitsu.com/global/products ... enchmarks/

12 cores, SMT8, 4.25GHz, up to 2 sockets per node with up to 4 nodes:
https://www.fujitsu.com/global/products ... up/m12-2s/

--
Srdja

[smatovic]

AMD announced Zen 5 for release in July, AMD claims 16% IPC increase over Zen 4, now 6 instead of 4 ALUs per core, and AVX-512 seems now to be 512-bit wide (dunno if it will clock down under load), desktop Ryzen from 6 to 16 cores, mobile Ryzen with up to 8 cores with XDNA 2 AI engine (NPU/TPU), server version with up to 128 cores, and Zen 5c (less cache, lower freq.) with up to 192 cores.

Zen 4 introduced AVX-512 instructions. AVX-512 capabilities have been expanded with Zen 5 with a doubling of the floating point pipe width to 512-bit. Additionally, there is greater bfloat16 throughput which is beneficial for AI workloads.

https://en.wikipedia.org/wiki/Zen_5

--
Srdja

[Werewolf]

AMD announced Zen 5 for release in July, AMD claims 16% IPC increase over Zen 4, now 6 instead of 4 ALUs per core, and AVX-512 seems now to be 512-bit wide (dunno if it will clock down under load), desktop Ryzen from 6 to 16 cores, mobile Ryzen with up to 8 cores with XDNA 2 AI engine (NPU/TPU), server version with up to 128 cores, and Zen 5c (less cache, lower freq.) with up to 192 cores.

Zen 4 introduced AVX-512 instructions. AVX-512 capabilities have been expanded with Zen 5 with a doubling of the floating point pipe width to 512-bit. Additionally, there is greater bfloat16 throughput which is beneficial for AI workloads.

https://en.wikipedia.org/wiki/Zen_5

--
Srdja

Slight shame there's no raise from 16 cores, given that Turin Threadripper won't be out in 2024.

[smatovic]

NVLink, InfinityFabric, CXL for GPU-GPU interconnect, now:

Everyone Except Nvidia Forms Ultra Accelerator Link (UALink) Consortium
https://www.hpcwire.com/2024/05/30/ever ... onsortium/

AMD, Broadcom, Cisco, Google, Hewlett Packard Enterprise (HPE), Intel, Meta, and Microsoft announced they have aligned to develop a new industry standard dedicated to advancing high-speed and low-latency communication for scale-up AI Accelerators.

Scale-up via UALink, scale-out via UEC.

I guess we end users will stick on PCIe in our home PCs.

--
Srdja

[smatovic]

[...]
Slight shame there's no raise from 16 cores, given that Turin Threadripper won't be out in 2024.

Hmm, we need to wait for Zen 5 benchmarks with the new AVX-512 unit, but until now, the Zen 3 Ryzen series offer best price/performance ratio (nps per dollar) IMO.

As used machine the Intel Skylake series is still going, it was really a good architecture.

For another non-chess project, I am still evaluating different server setups, scale-up and scale-out, maybe the Ampere Altra as single socket for scale-out, and there are different vendors for scale-up, with multiple sockets, in the market, AMD and ARM offer only max two CPU sockets per node.

--
Srdja

[smatovic]

For Lc0 it looks like the 4090 will still be king for the next 12 months then.

Yep, I think the Turing RTX 2080 TI is still a good shot in regard of FP16 performance.

--
Srdja

Nvidia RTX 5090 and RTX 5080 maybe this fall, for the X-mas shoppers:

We've spoken with some people as well, and the expectation is that we'll see at least the RTX 5090 and RTX 5080 by the time the holiday season kicks off in October or November.

https://www.tomshardware.com/pc-compone ... ng-we-know

AMD RDNA4 maybe fourth quarter of 2024:
https://www.digitaltrends.com/computing ... ce-rumors/

Intel Arc Battlemage maybe this fall:
https://www.tomshardware.com/pc-compone ... -this-fall

The smallest fab-processes are in high demand, might expect 2nm and 1+nm in the next couple years, maybe when the AI chip demand was satisfied we can expect bigger steps in consumer GPUs horse power and also lower prices.

--
Srdja