Tag:

CPUs

AMD says Instinct MI400X GPU is 10X faster than MI300X, will power Helios rack-scale system with EPYC ‘Venice’ CPUs

by admin June 13, 2025

AMD gave a preview of its first in-house designed rack-scale solution called Helios at its Advancing AI event on Thursday. The system is set to be based on the company’s next-generation EPYC ‘Venice’ processors, will use its Instinct MI400-series accelerator, and will rely on network connections featuring the upcoming Pensando network cards. Overall, the company says that the flagship MI400X is 10 times more powerful than the MI300X, which is a remarkable progress given that the MI400X will be released about three years after the MI300X.

When it comes to rack-scale solutions for AI, AMD clearly trails behind Nvidia. This is going to change a bit this year as cloud service providers (such as Oracle OCI), OEMs, and ODMs will build and deploy rack-scale solutions based on the Instinct MI350X-series GPUs, but those systems will not be designed by AMD, and they will have to interconnect each 8-way system using Ethernet, not low-latency high-bandwidth interconnects like NVLink.

Swipe to scroll horizontally

Year

2025

2026

2024

2025

2026

2027

Density

128

NVL72

NVL144

NVL576

GPU Architecture

CDNA 4

CDNA 5

Blackwell

Blackwell Ultra

Rubin

Rubin Ultra

GPU/GPU+CPU

MI355X

MI400X

GB200

GB300

VR200

VR300

Compute Chiplets

256

144

576

GPU Packages

128

144

FP4 PFLOPs (Dense)

1280

1440

720

1080

3600

14400

HBM Capacity

36 TB

51 TB

14 TB

21 TB

147 TB

HBM Bandwidth

1024 TB/s

1,400 TB/s

576 TB/s

936 TB/s

4,608 TB/s

CPU

EPYC ‘Turin’

EPYC ‘Venice’

72-core Grace

88-core Vera

NVSWitch/UALink/IF

–

UALink/IF

NVSwitch 5.0

NVSwitch 6.0

NVSwitch 7.0

NVSwitch Bandwidth

3600 GB/s

7200 GB/s

14400 GB/s

Scale-Out

800G, copper

1600G, optics

Form-Factor Name

OEM/ODM proprietary

Helios

Oberon

Kyber

The real change will occur next year with the first AMD-designed rack-scale system called Helios, which will use Zen 6-powered EPYC ‘Venice’ CPUs, CDNA ‘Next’-based Instinct MI400-series GPUs, and Pensando ‘Vulcano’ network interface cards (NICs) that are rumored to increase the maximum scale-up world size to beyond eight GPUs, which will greatly enhance their capabilities for training and inference. The system will adhere to OCP standards and enable next-generation interconnects such as Ultra Ethernet and Ultra Accelerator Link, supporting demanding AI workloads.

You may like

”So let me introduce the Helios AI rack,” said Andrew Dieckman, corporate VP and general manager of AMD’s data center GPU business. “Helios is one of the system solutions that we are working on based on the Instinct MI400-series GPU, so it is a fully integrated AI rack with EPYC CPUs, Instinct MI400-series GPUs, Pensando NICs, and then our ROCm stack. It is a unified architecture designed for both frontier model training as well as massive scale inference [that] delivers leadership compute density, memory bandwidth, scale out interconnect, all built in an open OCP-compliant standard supporting Ultra Ethernet and UALink.”

From a performance point of view, AMD’s flagship Instinct MI400-series AI GPU (we will refer to it as to Instinct MI400X, though this is not the official name, and we will also call the CDNA Next as CDNA 5) doubles performance from the Instinct MI355X and increases memory capacity by 50% and bandwidth by more than 100%. While the MI355X delivers 10 dense FP4 PFLOPS, the MI400X is projected to hit 20 dense FP4 PFLOPS.

Overall, the company says that the flagship MI400X is 10 times more powerful than the MI300X, which is a remarkable progress given that the MI400X will be released about three years after the MI300X.

“When you look at our product roadmap and how we continue to accelerate, with MI355X, we have taken a major leap forward [compared to the MI300X]: we are delivering 3X the amount of performance on a broad set of models and workloads, and that is a significant uptick from the previous trajectory we were on from the MI300X with the MI325X,” said Dieckman. “Now, with the Instinct MI400X and Helios, we bend that curve even further, and Helios is designed to deliver up to 10X more AI performance on the the most advanced frontier models in the high end.”

Swipe to scroll horizontally

Year

2024

2025

2024

2025

2026

2027

Architecture

CDNA 4

CDNA 5

Blackwell

Blackwell Ultra

Rubin

GPU

MI355X

MI400X

B200

B300 (Ultra)

VR200

VR300 (Ultra)

Process Technology

N3P

4NP

N3P (3NP?)

Physical Configuration

2 x Reticle Sized GPU

2 x Reticle Sized GPUs

2 x Reticle Sized GPUs, 2x I/O chiplets

4 x Reticle Sized GPUs, 2x I/O chiplets

Packaging

CoWoS-S

CoWoS-L

FP4 PFLOPs (per Package)

100

FP8/INT6 PFLOPs (per Package)

5/-

10/?

4.5

INT8 PFLOPS (per Package)

4.5

0.319

BF16 PFLOPs (per Package)

2.5

2.25

TF32 PFLOPs (per Package)

1.12

2.5

FP32 PFLOPs (per Package)

153.7

1.12

0.083

FP64/FP64 Tensor TFLOPs (per Package)

78.6

1.39

Memory

288 GB HBM3E

432 GB HBM4

192 GB HBM3E

288 GB HBM3E

288 GB HBM4

1 TB HBM4E

Memory Bandwidth

8 TB/s

almost’ 20 GB/s

8 TB/s

4 TB/s

13 TB/s

32 TB/s

HBM Stacks

NVLink/UALink

Infinity Fabric

UALink, Infinity Fabric

NVLink 5.0, 200 GT/s

NVLink 6.0

NVLink 7.0

SerDes speed (Gb/s unidirectional)

224G

GPU TDP

1400 W

1600 W (?)

1200 W

1400 W

1800 W

3600 W

CPU

128-core EPYC ‘Turin’

EPYC ‘Venice’

72-core Grace

88-core Vera

The new MI400X accelerator will also surpass Nvidia’s Blackwell Ultra, which is currently ramping up. However, when it comes to comparison with Nvidia’s next-generation Rubin R200 that delivers 50 dense FP4 PFLOPS, AMD’s MI400X will be around 2.5 times slower. Still, AMD will have an ace up its sleeve, which is memory bandwidth and capacity (see tables for details). Similarly, Helios will outperform Nvidia’s Blackwell Ultra-based NVL72 and Rubin-based NVL144.

However, it remains to be seen how Helios will stack against NVL144 in real-world applications. Also, it will be extremely hard to beat Nvidia’s NVL576 both in terms of compute performance and memory bandwidth in 2027, though by that time, AMD will likely roll out something new.

At least, this is what AMD communicated at its Advancing AI event this week: the company plans to continue evolving its integrated AI platforms with next-generation GPUs, CPUs, and networking technology, extending its roadmap well into 2027 and beyond.

Follow Tom’s Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.

Source link

June 13, 2025 0 comments

Product Reviews

ASE adopts AMD CPUs, begins evaluating Instinct MI300-series GPUs for AI

by admin June 6, 2025

AMD this week said in a blog post that ASE Technology, the world’s largest outsourced semiconductor assembly and test (OSAT) provider, has transitioned to EPYC and Ryzen processors across its data centers and client systems, respectively. The transition has resulted in significant performance improvements and energy efficiency gains. However, perhaps more important is that ASE is now evaluating AMD’s Instinct MI300-series processors for AI workloads.

By adopting AMD’s EPYC processors for servers and Ryzen CPUs for client desktop and laptop PCs, ASE achieved a 50% boost in system performance and a 6.5% reduction in power consumption compared to its previous infrastructure, which resulted in a 30% decrease in total cost of ownership, delivering both operational and financial benefits. AMD’s blog does not disclose which processors ASE used before adopting AMD-based solutions, nor does it indicate whether all systems in ASE’s fleet now use EPYC or Ryzen processors. However, the mention of operational and financial benefits points to a substantial adoption of AMD-based systems.

“We need to handle a big volume of data analysis, including leading-edge technology for AI applications and our smart factories,” said Jekyll Chen, Director of IT Infrastructure for ASE. “We work for many semiconductor companies. Our challenges are the need for high performance, low latency, and high core count, in alignment with ASE’s ESG policy. Stability and scalability are two primary goals for us.”

You may like

(Image credit: ASE Technology)

ASE Technology Holdings is the world’s largest outsourced semiconductor assembly and test provider, with packaging facilities in China, Japan, Korea, Malaysia, Singapore, and Taiwan. The company has worked with AMD on advanced 2.5D packaging since 2007, and this largely resulted in the invention of high-bandwidth memory (HBM). However, while ASE does provide packaging services for AMD these days, we are unsure whether ASE packages AMD’s AI GPUs, as Instinct processors utilize TSMC’s CoWoS technology.

AMD says that many companies are adopting or evaluating its Instinct processors for on-prem AI inference, though ASE is probably the first company of this scale to confirm the evaluation of these accelerators. In fact, the confirmation may indicate that ASE is close to adopting Instinct MI300-series GPUs for its internal AI workloads.

“We must perform data processing, run AI algorithms, and make sure everything operates smoothly, efficiently, and with the flexibility needed in our smart factories,” Chen said. “For client PCs, we need to make sure that they meet the needs of engineering design and the high-performance objectives of digital transformation. We also evaluated the performance, stability, core count, efficiency, total cost of ownership, AI speed, and multi-tasking capabilities of the new servers.”

Follow Tom’s Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.

Source link

June 6, 2025 0 comments

Product Reviews

Graphene thermal pad for AMD CPUs promises 17X better conductivity than thermal paste, 2X improvement over Thermal Grizzly

by admin June 5, 2025

Chinese manufacturer Coracer has introduced the GPE-01 graphene thermal pads for AMD’s AM5 processors, which are currently some of the best CPUs you can buy. Previously, the GPE-01 was compatible only with Intel’s LGA1851 and LGA1700 chips.

Traditional thermal paste has existed for decades. While it may not be the most exciting component, it effectively fulfills its purpose. In recent years, however, enthusiasts have developed a preference for alternative materials to fill the gap between the processor and CPU cooler. As a result, options like liquid metal and thermal pads have become more common in the market and on our page of the best thermal paste and TIM we’ve tested.

Segotep introduced the GPE-01 (21 x 44mm) at the end of last year, tailored for the elongated designs of Intel’s LGA1851 and LGA1700 processors. While Segotep is a well-established brand in the Chinese market, we were unaware of Coracer. A Google search brought no results, leaving us uncertain whether Coracer is a new sub-brand of Segotep or if the original manufacturer licensed the GPE-01 to another brand.

You may like

Regardless, the AM5 variant of the GPE-01 measures 32 x 32mm and aligns perfectly with the processor’s integrated heat spreader (IHS). While AM5 chips feature an unconventional design, the GPE-01 adopts a square shape that covers the entire IHS but doesn’t overreach to the corners.

According to Coracer, the GPE-01 utilizes a combination of graphene and silicon to achieve an eye-popping thermal conductivity up to 130 W/m·K. The thermal pad is encased in an isolating material that stops the graphene from creating a short circuit with the processor. Graphene is known for its excellent conductivity. As a result, manufacturers increasingly utilize this material in thermal pads and paste.

GPE-01 Graphene Thermal Pad Specifications

Swipe to scroll horizontally

Brand

Thermal Conductivity

Pricing

GPE-01

130 W/m·K

Thermal Grizzly Conductonaut

73 W/m·K

$12.43

Honeywell PTM7950

8.5 W/m·K

$19.99

Thermal Grizzly KryoSheet

7.5 W/m·K

$18.13

Arctic MX-6

7.5 W/m·K

$7.99

The GPE-01 boasts an impressive thermal conductivity of 130 W/m·K. This figure is nearly 2X higher than Thermal Grizzly’s Conductonaut liquid metal thermal paste. Additionally, it exceeds the thermal conductivity of Thermal Grizzly KryoSheet graphene thermal pads and Arctic MX-6 carbon filler-based thermal paste by more than 17X, and the Honeywell PTM7950 phase change thermal pad by 15X.

As with any claims regarding thermal conductivity, we advise being skeptical of figures provided by vendors. There have been instances where manufacturers inflate the thermal conductivity values to promote their products. While we are not alleging that this is the situation with the GPE-01, it is wise to approach the numbers with caution. Research has shown that graphene is a thermally conductive material that can reach levels up to 4,000 W/m·K, so there may be some credence to the GPE-01’s high conductivity figures.

In addition to its astronomical conductivity values, Coracer states that the GPE-01 can last up to 10 years. This assertion is more credible, given that the performance of thermal paste diminishes over time and ultimately dries out. Thermal pads, in contrast, have a longer lifespan, eliminating the need to repaste your processor every few years like with regular thermal paste. With the GPE-01, users likely won’t need to replace the pads until their next upgrade, unless they keep their system for over a decade.

We haven’t come across any online reviews for the GPE-01, so we are unable to assess the performance of the graphene thermal pad. Coracer has not disclosed the price or availability of the GPE-01, while the Intel version is priced at approximately $15 on the Chinese e-commerce site Taobao. This price is similar to that of Thermal Grizzly’s KryoSheet graphene thermal pads.

Follow Tom’s Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.

Source link

June 5, 2025 0 comments

Product Reviews

ASRock confirms its BIOS settings are killing Ryzen CPUs, is fully committed to fixing any damaged motherboards

by admin May 29, 2025

ASRock confirmed that it is responsible Ryzen 9000 failures plaguing ASRock motherboard owners. In a discussion between Gamers Nexus and ASRock’s VP of motherboards, Chris Lee, the company confirmed its latest BIOS revision, 3.25, alters certain PBO settings to stop Ryzen 9000 chips from failing.

A few days ago, YouTuber Tech Yes City revealed that ASRock had allegedly solved its Ryzen 9000 problem. The issue surrounded three power and voltage-related functions, Electric Design Current (EDC) and Thermal Design Current (TDC) in Precision Boost Overdrive (PBO), and shadow voltages which were set too high in older ASRock firmware. ASRock has since provided BIOS version 3.25 to its armada of AM5 motherboards with re-tuned EDC, TDC, and shadow voltage settings to rectify the issue.

ASRock Failures Face-to-Face: Motherboards, BIOS, & Burned 9800X3D CPUs – YouTube

Watch On

ASRock confirmed to Gamers Nexus that it is solely responsible, and AMD’s chips aren’t causing problems. ASRock also stated it will cover shipping costs both ways if users RMA their motherboards. However, it notes that it has not seen a single damaged motherboard in this whole ordeal.

You may like

If customers RMA their motherboard with the faulty CPU, ASRock will automatically send the CPU back to the retailer from which the customer bought it. ASRock is not recommending users RMA this way; rather, this was to clarify what happens if this scenario occurs (particularly for less tech-savvy focused customers).

ASRock also recommends that users check if BIOS version 3.25 is installed on brand-new AM5 ASRock motherboards and not to assume it is already preloaded from the factory. ASRock also apologized for not being as transparent with customers about the problem as it could be.

This is the second time ASRock has had to provide BIOS updates to fix its “Ryzen 9000 death” problem. The first time, ASRock pinned the Ryzen 9000 failures on memory compatibility issues and claimed initially that the problem was fixed with a new BIOS update that rectified the memory compatibility issues. Only time will tell if ASRock’s second attempt will be successful.

Evidence has already emerged that ASRock has not fully resolved this issue. At least two Reddit reports online claim that their Ryzen 7 9800X3Ds died with the 3.25 BIOS update.

Source link

May 29, 2025 0 comments